Weekly TMLR digest for Feb 23, 2025
TMLR
New certifications
==================
Reproducibility Certification: The Elusive Pursuit of Reproducing PATE-GAN: Benchmarking, Auditing, Debugging
Georgi Ganev, Meenatchi Sundaram Muthu Selva Annamalai, Emiliano De Cristofaro
https://openreview.net/forum?id=wcxrJcJ7vq
---
Survey Certification: Where Do We Stand with Implicit Neural Representations? A Technical and Performance Survey
Amer Essakine, Yanqi Cheng, Chun-Wun Cheng, Lipei Zhang, Zhongying Deng, Lei Zhu, Carola-Bibiane Schönlieb, Angelica I Aviles-Rivero
https://openreview.net/forum?id=QTsJXSvAI2
---
Featured Certification: The 2023 Foundation Model Transparency Index
Rishi Bommasani, Kevin Klyman, Shayne Longpre, Sayash Kapoor, Nestor Maslej, Betty Xiong, Daniel Zhang, Percy Liang
https://openreview.net/forum?id=x6fXnsM9Ez
---
Featured Certification: Loss-to-Loss Prediction: Scaling Laws for All Datasets
David Brandfonbrener, Nikhil Anand, Nikhil Vyas, Eran Malach, Sham M. Kakade
https://openreview.net/forum?id=1Avb4jYjLb
---
Accepted papers
===============
Title: Provable Quantum Algorithm Advantage for Gaussian Process Quadrature
Authors: Cristian A. Galvis-Florez, Ahmad Farooq, Simo Särkkä
Abstract: The aim of this paper is to develop novel quantum algorithms for Gaussian process quadrature methods. Gaussian process quadratures are numerical integration methods where Gaussian processes are used as functional priors for the integrands to capture the uncertainty arising from the sparse function evaluations. Quantum computers have emerged as potential replacements for classical computers, offering exponential reductions in the computational complexity of machine learning tasks. In this paper, we combine Gaussian process quadrature and quantum computing by proposing a quantum low-rank Gaussian process quadrature method based on a Hilbert space approximation of the Gaussian process kernel and enhancing the quadrature using a quantum circuit. The method combines the quantum phase estimation algorithm with the quantum principal component analysis technique to extract information up to a desired rank. Then, Hadamard and SWAP tests are implemented to find the expected value and variance that determines the quadrature. We use numerical simulations of a quantum computer to demonstrate the effectiveness of the method. Furthermore, we provide a theoretical complexity analysis that shows a polynomial advantage over classical Gaussian process quadrature methods. The code is available at https://github.com/cagalvisf/Quantum_HSGPQ.
URL: https://openreview.net/forum?id=K6CvWPtF62
---
Title: Differentially Private Source-Target Clustering
Authors: Shachar Schnapp, Sivan Sabato
Abstract: We consider a new private variant of the Source-Target Clustering (STC) setting, which was introduced by de Mathelin et al. (2022). In STC, there is a target dataset that needs to be clustered by selecting centers, in addition to centers that are already provided in a separate source dataset. The goal is to select centers from the target, such that the target clustering cost given the additional source centers is minimized. We consider private STC, in which the source dataset is private and should only be used under the constraint of differential privacy. This is motivated by scenarios in which the existing centers are private, for instance because they represent individuals in a social network. We derive lower bounds for the private STC objective, illustrating the theoretical limitations on worst-case guarantees for this setting. We then present a differentially private algorithm with asymptotically advantageous results under a data-dependent analysis, in which the guarantee depends on properties of the dataset, as well as more practical variants. We demonstrate in experiments the reduction in clustering cost that is obtained by our practical algorithms compared to baseline approaches.
URL: https://openreview.net/forum?id=ojeCoOKwWp
---
Title: Uncertainty-Based Experience Replay for Task-Agnostic Continual Reinforcement Learning
Authors: Adrian Remonda, Cole Corbitt Terrell, Eduardo E. Veas, Marc Masana
Abstract: Model-based reinforcement learning uses a learned dynamics model to imagine actions and select those with the best expected outcomes. An experience replay buffer collects the outcomes of all actions executed in the environment, which is then used to iteratively train the dynamics model. However, as the complexity and scale of tasks increase, training times and memory requirements can grow drastically without necessarily retaining useful experiences. Continual learning proposes a more realistic scenario where tasks are learned in sequence, and the replay buffer can help mitigate catastrophic forgetting. However, it is not realistic to expect the buffer to infinitely grow as the sequence advances. Furthermore, storing every single experience executed in the environment does not necessarily provide a more accurate model. We argue that the replay buffer needs to have the minimal necessary size to retain relevant experiences that cover both common and rare states. Therefore, we propose using an uncertainty-based replay buffer filtering to enable an effective implementation of continual learning agents using model-based reinforcement learning. We show that the combination of the proposed strategies leads to reduced training times, smaller replay buffer size, and less catastrophic forgetting, all while maintaining performance.
URL: https://openreview.net/forum?id=WxHTSPS2pi
---
Title: Robust Preference Optimization through Reward Model Distillation
Authors: Adam Fisch, Jacob Eisenstein, Vicky Zayats, Alekh Agarwal, Ahmad Beirami, Chirag Nagpal, Peter Shaw, Jonathan Berant
Abstract: Language model (LM) post-training (or alignment) involves maximizing a reward function that is derived from preference annotations. Direct Preference Optimization (DPO) is a popular offline alignment method that trains a policy directly on preference data without the need to train a reward model or apply reinforcement learning. However, the empirical evidence suggests that DPO typically assigns implicit rewards that overfit, and trend towards infinite magnitude. This frequently leads to degenerate policies, sometimes causing even the probabilities of the preferred generations to go to zero. In this work, we analyze this phenomenon and use distillation to get a better proxy for the true preference distribution over generation pairs: we train the LM such that its induced implicit reward, i.e., the scaled log-likelihood ratio of the model to the reference model, matches an explicit reward model trained on the preference data. Moreover, to account for uncertainty in the reward model we are distilling from, we optimize against a family of reward models that, as a whole, is likely to include at least one reasonable proxy for the preference distribution. Our results show that distilling from such a family of reward models leads to improved robustness to distribution shift in preference annotations, while preserving the simple supervised nature of DPO.
URL: https://openreview.net/forum?id=E2zKNuwNDc
---
Title: SimPLR: A Simple and Plain Transformer for Efficient Object Detection and Segmentation
Authors: Duy Kien Nguyen, Martin R. Oswald, Cees G. M. Snoek
Abstract: The ability to detect objects in images at varying scales has played a pivotal role in the design of modern object detectors. Despite considerable progress in removing hand-crafted components and simplifying the architecture with transformers, multi-scale feature maps and pyramid designs remain a key factor for their empirical success. In this paper, we show that shifting the multiscale inductive bias into the attention mechanism can work well, resulting in a plain detector ‘SimPLR’ whose backbone and detection head are both non-hierarchical and operate on single-scale features. We find through our experiments that SimPLR with scale-aware attention is plain and simple architecture, yet competitive with multi-scale vision transformer alternatives. Compared to the multi-scale and single-scale state-of-the-art, our model scales better with bigger capacity (self-supervised) models and more pre-training data, allowing us to report a consistently better accuracy and faster runtime for object detection, instance segmentation, as well as panoptic segmentation. Code is released at \url{https://github.com/kienduynguyen/SimPLR}.
URL: https://openreview.net/forum?id=6LO1y8ZE0F
---
Title: The Elusive Pursuit of Reproducing PATE-GAN: Benchmarking, Auditing, Debugging
Authors: Georgi Ganev, Meenatchi Sundaram Muthu Selva Annamalai, Emiliano De Cristofaro
Abstract: Synthetic data created by differentially private (DP) generative models is increasingly used in real-world settings.
In this context, PATE-GAN has emerged as one of the most popular algorithms, combining Generative Adversarial Networks (GANs) with the private training approach of PATE (Private Aggregation of Teacher Ensembles).
In this paper, we set out to reproduce the utility evaluation from the original PATE-GAN paper, compare available implementations, and conduct a privacy audit.
More precisely, we analyze and benchmark six open-source PATE-GAN implementations, including three by (a subset of) the original authors.
First, we shed light on architecture deviations and empirically demonstrate that none reproduce the utility performance reported in the original paper.
We then present an in-depth privacy evaluation, which includes DP auditing, and show that \textit{all implementations leak more privacy than intended}.
Furthermore, we uncover \textit{19 privacy violations} and 5 other bugs in these six open-source implementations.
Lastly, our codebase is available from: \url{https://github.com/spalabucr/pategan-audit}.
URL: https://openreview.net/forum?id=wcxrJcJ7vq
---
Title: Exploiting Benford's Law for Weight Regularization of Deep Neural Networks
Authors: Julius Ott, Huawei Sun, Enrico Rinaldi, Gianfranco Mauro, Lorenzo Servadei, Robert Wille
Abstract: Stochastic learning of Deep Neural Network (DNN) parameters is highly sensitive to training strategy, hyperparameters, and available training data. Many state-of-the-art solutions use weight regularization to adjust parameter distributions, prevent overfitting, and support generalization of DNNs. None of the existing regularization techniques have ever exploited a typical distribution of numerical datasets with respect to the first non-zero (or significant) digit, called Benford's Law (BL). In this paper, we show that the deviation of the significant digit distribution of the DNN weights from BL is closely related to the generalization of the DNN. In particular, when the DNN is presented with limited training data. To take advantage of this finding, we use BL to target the weight regularization of DNNs. Extensive experiments are performed on image, table, and speech data, considering convolutional (CNN) and Transformer-based neural network architectures with varying numbers of parameters. We show that the performance of DNNs is improved by minimizing the distance between the significant digit distributions of the DNN weights and the BL distribution along with L2 regularization. The improvements depend on the network architecture and how it deals with limited data. However, the proposed penalty term improves consistently and some CNN-based architectures gain up to $15\%$ test accuracy over the default training scheme with L2 regularization on subsets of CIFAR 100.
URL: https://openreview.net/forum?id=TnT59yz7lc
---
Title: Are Large Language Models Really Robust to Word-Level Perturbations?
Authors: Haoyu Wang, Guozheng Ma, Cong Yu, Ning Gui, Linrui Zhang, Zhiqi Huang, Suwei Ma, Yongzhe Chang, Sen Zhang, Li Shen, Xueqian Wang, Peilin Zhao, Dacheng Tao
Abstract: The swift advancement in the scales and capabilities of Large Language Models (LLMs) positions them as promising tools for a variety of downstream tasks. In addition to the pursuit of better performance and the avoidance of violent feedback on a certain prompt, to ensure the responsibility of the LLMs, much attention is drawn to the robustness of LLMs. However, existing evaluation methods mostly rely on traditional question answering datasets with predefined supervised labels, potentially ignoring the superior generation capabilities of contemporary LLMs. To investigate the robustness of LLMs while using their generation ability, we propose a novel rational evaluation pipeline that leverages reward models as diagnostic tools to evaluate the long conversation generated from more challenging open questions by LLMs, which we refer to as the Reward Model for Reasonable Robustness Evaluation (TREvaL). Longer conversations manifest the comprehensive grasp of language models in terms of their proficiency in understanding questions, a capability not entirely encompassed by individual words or letters.Our extensive empirical experiments demonstrate that TREvaL provides an identification for the lack of robustness of nowadays LLMs.Notably, we are surprised to discover that robustness tends to decrease as fine-tuning (SFT and RLHF) is conducted, calling for more attention on the robustness during alignment process.
URL: https://openreview.net/forum?id=rWSiBknwQa
---
Title: Where Do We Stand with Implicit Neural Representations? A Technical and Performance Survey
Authors: Amer Essakine, Yanqi Cheng, Chun-Wun Cheng, Lipei Zhang, Zhongying Deng, Lei Zhu, Carola-Bibiane Schönlieb, Angelica I Aviles-Rivero
Abstract: Implicit Neural Representations (INRs) have emerged as a paradigm in knowledge representation, offering exceptional flexibility and performance across a diverse range of applications. INRs leverage multilayer perceptrons (MLPs) to model data as continuous implicit functions, providing critical advantages such as resolution independence, memory efficiency, and generalisation beyond discretised data structures. Their ability to solve complex inverse problems makes them particularly effective for tasks including audio reconstruction, image representation, 3D object reconstruction, and high-dimensional data synthesis. This survey provides a comprehensive review of state-of-the-art INR methods, introducing a clear taxonomy that categorises them into four key areas: activation functions, position encoding, combined strategies, and network structure optimisation. We rigorously analyse their critical properties—such as full differentiability, smoothness, compactness, and adaptability to varying resolutions—while also examining their strengths and limitations in addressing locality biases and capturing fine details. Our experimental comparison offers new insights into the trade-offs between different approaches, showcasing the capabilities and challenges of the latest INR techniques across various tasks. In addition to identifying areas where current methods excel, we highlight key limitations and potential avenues for improvement, such as developing more expressive activation functions, enhancing positional encoding mechanisms, and improving scalability for complex, high-dimensional data. This survey serves as a roadmap for researchers, offering practical guidance for future exploration in the field of INRs. We aim to foster new methodologies by outlining promising research directions for INRs and applications.
URL: https://openreview.net/forum?id=QTsJXSvAI2
---
Title: Single-pass Detection of Jailbreaking Input in Large Language Models
Authors: Leyla Naz Candogan, Yongtao Wu, Elias Abad Rocamora, Grigorios Chrysos, Volkan Cevher
Abstract: Defending aligned Large Language Models (LLMs) against jailbreaking attacks is a challenging problem, with existing approaches requiring multiple requests or even queries to auxiliary LLMs, making them computationally heavy. Instead, we focus on detecting jailbreaking input in a single forward pass. Our method, called SPD, leverages the information carried by the logits to predict whether the output sentence will be harmful. This allows us to defend in just a forward pass. SPD can not only detect attacks effectively on open-source models, but also minimizes the misclassification of harmless inputs. Furthermore, we show that SPD remains effective even without complete logit access in GPT-3.5 and GPT-4. We believe that our proposed method offers a promising approach to efficiently safeguard LLMs against adversarial attacks.
URL: https://openreview.net/forum?id=42v6I5Ut9a
---
Title: On Training-Conditional Conformal Prediction and Binomial Proportion Confidence Intervals
Authors: Rudi Coppola, Manuel Mazo Espinosa
Abstract: Estimating the expectation of a Bernoulli random variable based on $N$ independent trials is a classical problem in statistics, typically addressed using Binomial Proportion Confidence Intervals (BPCI). In the control systems community, many critical tasks—such as certifying the statistical safety of dynamical systems—can be formulated as BPCI problems.
Conformal Prediction (CP), a distribution-free technique for uncertainty quantification, has gained significant attention in recent years and has been applied to various control systems problems, particularly to address uncertainties in learned dynamics or controllers. A variant known as training-conditional CP was recently employed to tackle the problem of safety certification.
In this note, we highlight that the use of training-conditional CP in this context does not provide valid safety guarantees. We demonstrate why CP is unsuitable for BPCI problems and argue that traditional BPCI methods are better suited for statistical safety certification.
URL: https://openreview.net/forum?id=pSk5qyt1ob
---
Title: Generative Risk Minimization for Out-of-Distribution Generalization on Graphs
Authors: Song Wang, Zhen Tan, Yaochen Zhu, Chuxu Zhang, Jundong Li
Abstract: Out-of-distribution (OOD) generalization on graphs aims at dealing with scenarios where the test graph distribution differs from the training graph distributions. Compared to i.i.d. data like images, the OOD generalization problem on graph-structured data remains challenging due to the non-i.i.d. property and complex structural information on graphs. Recently, several works on graph OOD generalization have explored extracting invariant subgraphs that share crucial classification information across different distributions. Nevertheless, such a strategy could be suboptimal for entirely capturing the invariant information, as the extraction of discrete structures could potentially lead to the loss of invariant information or the involvement of spurious information. In this paper, we propose an innovative framework, named Generative Risk Minimization (GRM), designed to generate an invariant subgraph for each input graph to be classified, instead of extraction. To address the challenge of optimization in the absence of optimal invariant subgraphs (i.e., ground truths), we derive a tractable form of the proposed GRM objective by introducing a latent causal variable, and its effectiveness is validated by our theoretical analysis. We further conduct extensive experiments across a variety of real-world graph datasets for both node-level and graph-level OOD generalization, and the results demonstrate the superiority of our framework GRM.
URL: https://openreview.net/forum?id=EcMVskXo1n
---
Title: An Analytical Model for Overparameterized Learning Under Class Imbalance
Authors: Eliav Mor, Yair Carmon
Abstract: We study class-imbalanced linear classification in a high-dimensional Gaussian mixture model. We develop a tight, closed form approximation for the test error of several practical learning methods, including logit adjustment and class dependent temperature. Our approximation allows us to analytically tune and compare these methods, highlighting how and when they overcome the pitfalls of standard cross-entropy minimization. We test our theoretical findings on simulated data and imbalanced CIFAR10, MNIST and FashionMNIST datasets.
URL: https://openreview.net/forum?id=69RntSRF5K
---
Title: Evaluating the Robustness of Analogical Reasoning in Large Language Models
Authors: Martha Lewis, Melanie Mitchell
Abstract: Large language models (LLMs) have performed well on several reasoning benchmarks, including ones that test analogical reasoning abilities. However, there is debate on the extent to which they are performing general abstract reasoning versus employing shortcuts or other non-robust processes, such as ones that overly rely on similarity to what has been seen in their training data. Here we investigate the robustness of analogy-making abilities previously claimed for LLMs on three of four domains studied by Webb et al. (2023): letter-string analogies, digit matrices, and story analogies. For each of these domains we test humans and GPT models on robustness to variants of the original analogy problems—versions that test the same abstract reasoning abilities but that are likely dissimilar from tasks in the pre-training data. The performance of a system that uses robust abstract reasoning should not decline substantially on these variants.
On simple letter-string analogies, we find that while the performance of humans remains high for two types of variants we tested, the GPT models’ performance declines sharply. This pattern is less pronounced as the complexity of these analogy problems is increased, as both humans and GPT models perform poorly on both the original and variant problems requiring more complex analogies. On digit-matrix problems, we find a similar pattern but only on one out of the two types of variants we tested. Lastly, we assess the robustness of humans and GPT models on story-based analogy problems, finding that, unlike humans, the performance of GPT models are susceptible to answer-order effects, and that GPT models also may be more sensitive than humans to paraphrasing.
This work provides evidence that, despite previously reported successes of LLMs on zero-shot analogical reasoning, these models often lack the robustness of zero-shot human analogy- making, exhibiting brittleness on most of the variations we tested. More generally, this work points to the importance of carefully evaluating AI systems not only for accuracy but also robustness when testing their cognitive capabilities.
Code, data, and results for all experiments is available at https://github.com/marthaflinderslewis/robust-analogy.
URL: https://openreview.net/forum?id=t5cy5v9wph
---
Title: The Sparse Matrix-Based Random Projection: A Study of Binary and Ternary Quantization
Authors: Weizhi Lu, Zhongzheng Li, Mingrui Chen, Weiyu Li
Abstract: Random projection is a simple yet effective technique for dimension reduction, widely used in various machine learning tasks. Following the projection step, quantization is often applied to further reduce the complexity of projected data. In general, quantized projections are expected to approximately preserve the pairwise distances between the original data points, to avoid significant performance degradation in subsequent tasks. While this distance preservation property has been investigated for Gaussian matrices, our work further extends the analysis to hardware-friendly $\{0,1\}$-binary matrices, particularly focusing on cases where the projections are quantized into two types of low bit-width codes: $\{0,1\}$-binary codes and $\{0,\pm1\}$-ternary codes. It is found that the distance preservation property tends to be better maintained, when the binary projection matrices exhibit sparse structures. This is validated through classification and clustering experiments, where extremely sparse binary matrices, with only one nonzero entry per column, achieve superior or comparable performance to other denser binary matrices and Gaussian matrices. This presents an opportunity to significantly reduce the computational and storage complexity of the quantized random projection model, without compromising, and potentially even improving its performance.
URL: https://openreview.net/forum?id=dNJmJ8bh1M
---
Title: Comparing the information content of probabilistic representation spaces
Authors: Kieran A. Murphy, Sam Dillavou, Danielle Bassett
Abstract: Probabilistic representation spaces convey information about a dataset and are shaped by factors such as the training data, network architecture, and loss function.
Comparing the information content of such spaces is crucial for understanding the learning process, yet most existing methods assume point-based representations, neglecting the distributional nature of probabilistic spaces.
To address this gap, we propose two information-theoretic measures to compare general probabilistic representation spaces by extending classic methods to compare the information content of hard clustering assignments.
Additionally, we introduce a lightweight method of estimation that is based on fingerprinting a representation space with a sample of the dataset, designed for scenarios where the communicated information is limited to a few bits.
We demonstrate the utility of these measures in three case studies.
First, in the context of unsupervised disentanglement, we identify recurring information fragments within individual latent dimensions of VAE and InfoGAN ensembles.
Second, we compare the full latent spaces of models and reveal consistent information content across datasets and methods, despite variability during training.
Finally, we leverage the differentiability of our measures to perform model fusion, synthesizing the information content of weak learners into a single, coherent representation.
Across these applications, the direct comparison of information content offers a natural basis for characterizing the processing of information.
URL: https://openreview.net/forum?id=adhsMqURI1
---
Title: Adapt then Unlearn: Exploring Parameter Space Semantics for Unlearning in Generative Adversarial Networks
Authors: Piyush Tiwary, Atri Guha, Subhodip Panda, Prathosh AP
Abstract: Owing to the growing concerns about privacy and regulatory compliance, it is desirable to regulate the output of generative models. To that end, the objective of this work is to prevent the generation of outputs containing undesired features from a pre-trained Generative Adversarial Network (GAN) where the underlying training data set is inaccessible. Our approach is inspired by the observation that the parameter space of GANs exhibits meaningful directions that can be leveraged to suppress specific undesired features. However, such directions usually result in the degradation of the quality of generated samples. Our proposed two-stage method, known as 'Adapt-then-Unlearn,' excels at unlearning such undesirable features while also maintaining the quality of generated samples. In the initial stage, we adapt a pre-trained GAN on a set of negative samples (containing undesired features) provided by the user. Subsequently, we train the original pre-trained GAN using positive samples, along with a repulsion regularizer. This regularizer encourages the learned model parameters to move away from the parameters of the adapted model (first stage) while not degrading the generation quality. We provide theoretical insights into the proposed method. To the best of our knowledge, our approach stands as the first method addressing unlearning within the realm of high-fidelity GANs (such as StyleGAN). We validate the effectiveness of our method through comprehensive experiments, encompassing both class-level unlearning on the MNIST and AFHQ dataset and feature-level unlearning tasks on the CelebA-HQ dataset. Our code and implementation is available at: https://github.com/atriguha/Adapt_Unlearn.
URL: https://openreview.net/forum?id=jAHEBivObO
---
Title: On Memorization in Diffusion Models
Authors: Xiangming Gu, Chao Du, Tianyu Pang, Chongxuan Li, Min Lin, Ye Wang
Abstract: Due to their capacity to generate novel and high-quality samples, diffusion models have attracted significant research interest in recent years. Notably, the typical training objective of diffusion models, i.e., denoising score matching, has a closed-form optimal solution that can only generate training-data replicating samples. This indicates that a memorization behavior is theoretically expected, which contradicts the common generalization ability of state-of-the-art diffusion models, and thus calls for a deeper understanding. Looking into this, we first observe that memorization behaviors tend to occur on smaller-sized datasets, which motivates our definition of effective model memorization (EMM), a metric measuring the maximum size of training data at which a model approximates its theoretical optimum. Then, we quantify the impact of the influential factors on these memorization behaviors in terms of EMM, focusing primarily on data distribution, model configuration, and training procedure. Besides comprehensive empirical results identifying the influential factors, we surprisingly find that conditioning training data on uninformative random labels can significantly trigger the memorization in diffusion models. Our study holds practical significance for diffusion model users and offers clues to theoretical research in deep generative models.
URL: https://openreview.net/forum?id=D3DBqvSDbj
---
Title: Over-parameterised Shallow Neural Networks with Asymmetrical Node Scaling: Global Convergence Guarantees and Feature Learning
Authors: Francois Caron, Fadhel Ayed, Paul Jung, Hoil Lee, Juho Lee, Hongseok Yang
Abstract: We consider gradient-based optimisation of wide, shallow neural networks, where the output of each hidden node is scaled by a positive parameter. The scaling parameters are non-identical, differing from the classical Neural Tangent Kernel (NTK) parameterisation. We prove that for large such neural networks, with high probability, gradient flow and gradient descent converge to a global minimum and can learn features in some sense, unlike in the NTK parameterisation. We perform experiments illustrating our theoretical results and discuss the benefits of such scaling in terms of prunability and transfer learning.
URL: https://openreview.net/forum?id=Sx1khIIi95
---
Title: Uncertainty Representations in State-Space Layers for Deep Reinforcement Learning under Partial Observability
Authors: Carlos E. Luis, Alessandro Giacomo Bottero, Julia Vinogradska, Felix Berkenkamp, Jan Peters
Abstract: Optimal decision-making under partial observability requires reasoning about the uncertainty of the environment’s hidden state. However, most reinforcement learning architectures handle partial observability with sequence models that have no internal mechanism to incorporate uncertainty in their hidden state representation, such as recurrent neural networks, deterministic state-space models and transformers. Inspired by advances in probabilistic world models for reinforcement learning, we propose a standalone Kalman filter layer that performs closed-form Gaussian inference in linear state-space models and train it end-to-end within a model-free architecture to maximize returns. Similar to efficient linear recurrent layers, the Kalman filter layer processes sequential data using a parallel scan, which scales logarithmically with the sequence length. By design, Kalman filter layers are a drop-in replacement for other recurrent layers in standard model-free architectures, but importantly they include an explicit mechanism for probabilistic filtering of the latent state representation. Experiments in a wide variety of tasks with partial observability show that Kalman filter layers excel in problems where uncertainty reasoning is key for decision-making, outperforming other stateful models.
URL: https://openreview.net/forum?id=rfPns0WJyg
---
Title: Latent Space Energy-based Neural ODEs
Authors: Sheng Cheng, Deqian Kong, Jianwen Xie, Kookjin Lee, Ying Nian Wu, Yezhou Yang
Abstract: This paper introduces novel deep dynamical models designed to represent continuous-time sequences. Our approach employs a neural emission model to generate each data point in the time series through a non-linear transformation of a latent state vector. The evolution of these latent states is implicitly defined by a neural ordinary differential equation (ODE), with the initial state drawn from an informative prior distribution parameterized by an Energy-based model (EBM). This framework is extended to disentangle dynamic states from underlying static factors of variation, represented as time-invariant variables in the latent space. We train the model using maximum likelihood estimation with Markov chain Monte Carlo (MCMC) in an end-to-end manner. Experimental results on oscillating systems, videos and real-world state sequences (MuJoCo) demonstrate that our model with the learnable energy-based prior outperforms existing counterparts, and can generalize to new dynamic parameterization, enabling long-horizon predictions.
URL: https://openreview.net/forum?id=hCxtlfvL22
---
Title: Using representation balancing to learn conditional-average dose responses from clustered data
Authors: Christopher Bockel-Rickermann, Toon Vanderschueren, Jeroen Berrevoets, Tim Verdonck, Wouter Verbeke
Abstract: Estimating the response to an intervention with an associated dose conditional on a unit's covariates, the "conditional-average dose response" (CADR), is a relevant task in a variety of domains, from healthcare to business, economics, and beyond. Estimating such a response is challenging for several reasons: Firstly, it typically needs to be estimated from observational data, which can be confounded and negatively affect the performance of intervention response estimators used for counterfactual inference. Secondly, the continuity of the dose prevents the adoption of approaches used to estimate responses to binary-valued interventions. That is why the machine learning (ML) community has proposed several tailored CADR estimators. Yet, the proposal of most of these methods requires strong assumptions on the distribution of data and the assignment of interventions, which go beyond the standard assumptions in causal inference. Whereas previous works have so far focused on smooth shifts in covariate distributions across doses, in this work, we will study estimating CADR from clustered data and where different doses are assigned to different segments of a population. On a novel benchmarking dataset, we show the impacts of clustered data on model performance. Additionally, we propose an estimator, CBRNet, that enables the application of representation balancing for CADR estimation through clustering the covariate space and a novel loss function. CBRNet learns cluster-agnostic and hence dose-agnostic covariate representations through representation balancing for unbiased CADR inference. We run extensive experiments to illustrate the workings of our method and compare it with the state of the art in ML for CADR estimation.
URL: https://openreview.net/forum?id=U8EMkndyq4
---
Title: ARVideo: Autoregressive Pretraining for Self-Supervised Video Representation Learning
Authors: Sucheng Ren, Hongru Zhu, Chen Wei, Yijiang Li, Alan Yuille, Cihang Xie
Abstract: This paper presents a new self-supervised video representation learning framework \textbf{ARVideo}, which \textit{autoregressively} predict the next video token in a tailored sequence order.
Two key designs are included. First, we organize autoregressive video tokens into
clusters that span both \textit{spatially} and \textit{temporally}, thereby enabling a richer aggregation of contextual information compared to the standard spatial-only or temporal-only clusters. Second, we adopt a randomized spatiotemporal prediction order to facilitate learning from multi-dimensional data, addressing the limitations of a handcrafted spatial-first or temporal-first sequence order. Extensive experiments establish ARVideo as an effective paradigm for self-supervised video representation learning. For example,
when trained with the ViT-B backbone, ARVideo competitively attains 81.2\% on Kinetics-400 and 70.9\% on Something-Something V2, which are on par with the strong benchmark set by VideoMAE. Importantly, ARVideo also demonstrates higher training efficiency, \ie, it trains 14\% faster and requires 58\% less GPU memory compared to VideoMAE.
URL: https://openreview.net/forum?id=TRKwzPnXWQ
---
Title: The 2023 Foundation Model Transparency Index
Authors: Rishi Bommasani, Kevin Klyman, Shayne Longpre, Sayash Kapoor, Nestor Maslej, Betty Xiong, Daniel Zhang, Percy Liang
Abstract: Foundation models have rapidly permeated society, catalyzing a wave of generative AI applications spanning enterprise and consumer-facing contexts. While the societal impact of foundation models is growing, transparency is on the decline, mirroring the opacity that has plagued past digital technologies (e.g. social media). Reversing this trend is essential: transparency is a vital precondition for public accountability, scientific innovation, and effective governance. To assess the transparency of the foundation model ecosystem and help improve transparency over time, we introduce the Foundation Model Transparency Index. The Foundation Model Transparency Index specifies 100 fine-grained indicators that comprehensively codify transparency for foundation models, spanning the upstream resources used to build a foundation model (e.g data, labor, compute), details about the model itself (e.g. size, capabilities, risks), and the downstream use (e.g. distribution channels, usage policies, affected geographies). We score 10 major foundation model developers (e.g. OpenAI, Google, Meta) against the 100 indicators to assess their transparency. To facilitate and standardize assessment, we score developers in relation to their practices for their flagship foundation model (e.g. GPT-4 for OpenAI, PaLM 2 for Google, Llama 2 for Meta). We present 10 top-level findings about the foundation model ecosystem: for example, no developer currently discloses significant information about the downstream impact of its flagship model, such as the number of users, affected market sectors, or how users can seek redress for harm. Overall, the Foundation Model Transparency Index establishes the level of transparency today to drive progress on foundation model governance via industry standards and regulatory intervention.
URL: https://openreview.net/forum?id=x6fXnsM9Ez
---
Title: Dynamics-inspired Structure Hallucination for Protein-protein Interaction Modeling
Authors: Fang Wu, Stan Z. Li
Abstract: Protein-protein interaction (PPI) represents a central challenge within the biology field, and accurately predicting the consequences of mutations in this context is crucial for drug design and protein engineering. Deep learning (DL) has shown promise in forecasting the effects of such mutations but is hindered by two primary constraints. First, the structures of mutant proteins are often elusive to acquire. Secondly, PPI takes place dynamically, which is rarely integrated into the DL architecture design. To address these obstacles, we present a novel framework named Refine-PPI with two key enhancements. First, we introduce a structure refinement module trained by a mask mutation modeling (MMM) task on available wild-type structures, which is then transferred to hallucinate the inaccessible mutant structures. Second, we employ a new kind of geometric network, called the probability density cloud network (PDC-Net), to capture 3D dynamic variations and encode the atomic uncertainty associated with PPI. Comprehensive experiments on SKEMPI.v2 substantiate the superiority of Refine-PPI over all existing tools for predicting free energy change. These findings underscore the effectiveness of our hallucination strategy and the PDC module in addressing the absence of mutant protein structure and modeling geometric uncertainty.
URL: https://openreview.net/forum?id=GGHk5ukO6t
---
Title: Loss-to-Loss Prediction: Scaling Laws for All Datasets
Authors: David Brandfonbrener, Nikhil Anand, Nikhil Vyas, Eran Malach, Sham M. Kakade
Abstract: While scaling laws provide a reliable methodology for predicting train loss across compute scales for a single data distribution, less is known about how these predictions should change as we change the distribution. In this paper, we derive a strategy for predicting one loss from another and apply it to predict across different pre-training datasets and from pre-training data to downstream task data. Our predictions extrapolate well even at 20x the largest FLOP budget used to fit the curves. More precisely, we find that there are simple shifted power law relationships between (1) the train losses of two models trained on two separate datasets when the models are paired by training compute (train-to-train), (2) the train loss and the test loss on any downstream distribution for a single model (train-to-test), and (3) the test losses of two models trained on two separate train datasets (test-to-test). The results hold up for pre-training datasets that differ substantially (some are entirely code and others have no code at all) and across a variety of downstream tasks. Finally, we find that in some settings these shifted power law relationships can yield more accurate predictions than extrapolating single-dataset scaling laws.
URL: https://openreview.net/forum?id=1Avb4jYjLb
---
Title: ALTA: Compiler-Based Analysis of Transformers
Authors: Peter Shaw, James Cohan, Jacob Eisenstein, Kenton Lee, Jonathan Berant, Kristina Toutanova
Abstract: We propose a new programming language called ALTA and a compiler that can map ALTA programs to Transformer weights. ALTA is inspired by RASP, a language proposed by Weiss et al. (2021), and Tracr (Lindner et al., 2023), a compiler from RASP programs to Transformer weights. ALTA complements and extends this prior work, offering the ability to express loops and to compile programs to Universal Transformers, among other advantages. ALTA allows us to constructively show how Transformers can represent length-invariant algorithms for computing parity and addition, as well as a solution to the SCAN benchmark of compositional generalization tasks, without requiring intermediate scratchpad decoding steps. We also propose tools to analyze cases where the expressibility of an algorithm is established, but end-to-end training on a given training set fails to induce behavior consistent with the desired algorithm. To this end, we explore training from ALTA execution traces as a more fine-grained supervision signal. This enables additional experiments and theoretical analyses relating the learnability of various algorithms to data availability and modeling decisions, such as positional encodings. We make the ALTA framework --- language specification, symbolic interpreter, and weight compiler --- available to the community to enable further applications and insights.
URL: https://openreview.net/forum?id=h751wl9xiR
---
Title: Density of states in neural networks: an in-depth exploration of learning in parameter space
Authors: Margherita Mele, Roberto Menichetti, Alessandro Ingrosso, Raffaello Potestio
Abstract: Learning in neural networks critically hinges on the intricate geometry of the loss landscape associated with a given task. Traditionally, most research has focused on finding specific weight configurations that minimize the loss. In this work, born from the cross-fertilization of machine learning and theoretical soft matter physics, we introduce a novel approach to examine the weight space across all loss values. Employing the Wang-Landau enhanced sampling algorithm, we explore the neural network density of states -- the number of network parameter configurations that produce a given loss value -- and analyze how it depends on specific features of the training set. Using both real-world and synthetic data, we quantitatively elucidate the relation between data structure and network density of states across different sizes and depths of binary-state networks. This work presents and illustrates a novel, informative analysis method that aims at paving the way for a better understanding of the interplay between structured data and the networks that process, learn, and generate them.
URL: https://openreview.net/forum?id=BLDtWlFKhn
---
Title: Rethinking Spectral Augmentation for Contrast-based Graph Self-Supervised Learning
Authors: Xiangru Jian, Xinjian Zhao, Wei Pang, Chaolong Ying, Yimu Wang, Yaoyao Xu, Tianshu Yu
Abstract: The recent surge in contrast-based graph self-supervised learning has prominently featured an intensified exploration of spectral cues. Spectral augmentation, which involves modifying a graph's spectral properties such as eigenvalues or eigenvectors, is widely believed to enhance model performance. However, an intriguing paradox emerges, as methods grounded in seemingly conflicting assumptions regarding the spectral domain demonstrate notable enhancements in learning performance. Through extensive empirical studies, we find that simple edge perturbations - random edge dropping for node-level and random edge adding for graph-level self-supervised learning - consistently yield comparable or superior performance while being significantly more computationally efficient. This suggests that the computational overhead of sophisticated spectral augmentations may not justify their practical benefits. Our theoretical analysis of the InfoNCE loss bounds for shallow GNNs further supports this observation. The proposed insights represent a significant leap forward in the field, potentially refining the understanding and implementation of graph self-supervised learning.
URL: https://openreview.net/forum?id=HjpD5kpfa3
---
New submissions
===============
Title: Accelerating Learned Image Compression Through Modeling Neural Training Dynamics
Abstract: As learned image compression (LIC) methods become increasingly computationally demanding, enhancing their training efficiency is crucial. This paper takes a step forward in accelerating the training of LIC methods by modeling the neural training dynamics.
We first propose a Sensitivity-aware True and Dummy Embedding Training mechanism (STDET) that clusters LIC model parameters into few separate modes where parameters are expressed as affine transformations of reference parameters within the same mode.
By further utilizing the stable intra-mode correlations throughout training and parameter sensitivities, we gradually embed non-reference parameters, reducing the number of trainable parameters. Additionally, we incorporate a Sampling-then-Moving Average (SMA) technique, interpolating sampled weights from stochastic gradient descent (SGD) training to obtain the moving averaged weights, ensuring smooth temporal behavior and minimizing training state variances.
Overall, our method significantly reduces training space dimensions and the number of trainable parameters without sacrificing model performance, thus accelerating model convergence.
We also provide a theoretical analysis on the Noisy quadratic model, showing that the proposed method achieves a lower training variance than standard SGD. Our approach offers valuable insights for further developing efficient training methods for LICs. The code will be publicly available.
URL: https://openreview.net/forum?id=nannw4SGfS
---
Title: Multi-Modal One-Shot Federated Ensemble Learning for Medical Data with Vision Large Language Model
Abstract: Federated learning (FL) has attracted considerable interest in the medical domain due to its capacity to facilitate collaborative model training while maintaining data privacy. However, conventional FL methods typically necessitate multiple communication rounds, leading to significant communication overhead and delays, especially in environments with limited bandwidth. One-shot federated learning addresses these issues by conducting model training and aggregation in a single communication round, thereby reducing communication costs while preserving privacy. Among these, one-shot federated ensemble learning combines independently trained client models using ensemble techniques such as voting, further boosting performance in non-IID data scenarios. On the other hand, existing machine learning methods in healthcare predominantly use unimodal data (e.g., medical images or textual reports), which restricts their diagnostic accuracy and comprehensiveness. Therefore, the integration of multi-modal data is proposed to address these shortcomings. Additionally, vision large language models (vLLMs) have emerged as powerful tools due to their ability to interpret and generate textual descriptions from visual data, making them invaluable for creating textual reports from medical images. In this paper, we introduce FedMME, an innovative one-shot multi-modal federated ensemble learning framework that utilizes multi-modal data for medical image analysis. Specifically, FedMME capitalizes on vision large language models to produce textual reports from medical images, employs a BERT model to extract textual features from these reports, and amalgamates these features with visual features to improve diagnostic accuracy. Experimental results show that our method demonstrated superior performance compared to existing one-shot federated learning methods in healthcare scenarios across four datasets with various data distributions. For instance, it surpasses existing one-shot federated learning approaches by more than 17.5% in accuracy on the RSNA dataset when applying a Dirichlet distribution with ($\alpha$ = 0.3).
URL: https://openreview.net/forum?id=1swYmdt5sM
---
Title: Visual Prompting Reimagined: The Power of Activation Prompts
Abstract: Visual prompting (VP) has emerged as a popular method to repurpose pretrained vision models for adaptation to downstream tasks. Unlike conventional model fine-tuning techniques, VP introduces a universal perturbation directly into the input data to facilitate task-specific fine-tuning rather than modifying model parameters. However, there exists a noticeable performance gap between VP and conventional fine-tuning methods, highlighting an unexplored realm in theory and practice to understand and advance (input-level) VP to reduce its current performance gap. Towards this end, we introduce a generalized concept, termed activation prompt (AP), which extends the scope of (input-level) VP by enabling universal perturbations to be applied to activation maps within the intermediate layers of the model. By using AP to revisit the problem of VP and employing it as an analytical tool, we demonstrate the intrinsic limitations of VP in both performance and efficiency, revealing why input-level prompting may lack effectiveness compared to AP, which exhibits a model-dependent layer preference. We show that AP is closely related to normalization tuning in convolutional neural networks (CNNs) and vision transformers (ViTs), although each model type has distinct layer preferences for prompting. We also theoretically elucidate the rationale behind such preference by analyzing global features across layers. Through extensive experiments across 29 datasets and various model architectures, we provide a comprehensive performance analysis of AP, comparing it with VP and parameter-efficient fine-tuning (PEFT) baselines. Our results demonstrate AP's superiority in both accuracy and efficiency, considering factors such as time, parameters, memory usage, and throughput.
URL: https://openreview.net/forum?id=aRUJc2TU0y
---
Title: Can Kernel Methods Explain How the Data Affects Neural Collapse?
Abstract: A vast amount of literature has recently focused on the ''Neural Collapse'' (NC) phenomenon, which emerges when training neural network (NN) classifiers beyond the zero training error point. The core component of NC is the decrease in the within-class variability of the network's deepest features, dubbed as NC1. The theoretical works that study NC are typically based on simplified unconstrained features models (UFMs) that mask any effect of the data on the extent of collapse. To address this limitation of UFMs, this paper explores the possibility of analyzing NC1 using kernels associated with shallow NNs. We begin by formulating an NC1 metric as a function of the kernel. Then, we specialize it to the NN Gaussian Process kernel (NNGP) and the Neural Tangent Kernel (NTK), associated with wide networks at initialization and during gradient-based training with a small learning rate, respectively. As a key result, we show that the NTK does not represent more collapsed features than the NNGP for Gaussian data of arbitrary dimensions. This showcases the limitations of data independent kernels such as NTK in approximating the NC behavior of NNs. As an alternative to NTK, we then empirically explore a recently proposed data-aware Gaussian Process kernel, which generalizes NNGP to model feature learning. We show that this kernel yields lower NC1 than NNGP but may not follow the trends of the shallow NN. Our study demonstrates that adaptivity to data may allow kernel-based analysis of NC, though, further advancements in this area are still needed. A nice byproduct of our study is showing both theoretically and empirically that the choice of nonlinear activation function affects NC1 (with ERF yielding lower values than ReLU).
URL: https://openreview.net/forum?id=MbF1gYfIlY
---
Title: A Stochastic Polynomial Expansion for Uncertainty Propagation through Networks
Abstract: Network-based machine learning constructs are becoming more prevalent in sensing and decision-making systems. As these systems are implemented in safety-critical environments such as pedestrian detection and power management, it is crucial to evaluate confidence in their decisions. At the heart of this problem is a need to understand and characterize how errors at the input of networks become progressively expanded or contracted as signals move through layers, especially in light of the non-trivial nonlinearities manifest throughout modern machine learning architectures. When sampling methods become expensive due to network size or complexity, approximation is needed and popular methods include Jacobian (first order Taylor) linearization and stochastic linearization. However, despite computational tractability, the accuracy of these methods can break down in situations with moderate to high input uncertainty.
Here, we present a generalized method of propagating variational multivariate Gaussian distributions through neural networks. We propose a modified Taylor expansion function for nonlinear transformation of Gaussian distributions, with an additional approximation in which the polynomial terms act on independent Gaussian random variables (which are identically distributed). With these approximated higher order terms (HOTs), we obtain significantly more accurate estimation of layer-wise distributions. Despite the introduction of the HOTs, this method can propagate a full covariance matrix with a complexity of $\boldsymbol{O}(n^2)$ (and $\boldsymbol{O}(n)$ if only propagating marginal variance), comparable to Jacobian linearization. Thus, our method finds a balance between efficiency and accuracy. We derived the closed form solutions for this approximate Stochastic Taylor expansion for seven commonly used nonlinearities and verified the effectiveness of our method in deep residual neural networks, Bayesian neural networks, and variational autoencoders. This general method can be integrated into use-cases such as Kalman filtering, adversarial training, and variational learning.
URL: https://openreview.net/forum?id=lyDRBhUjhv
---
Title: Reproducibility Study of "Cooperate or Collapse: Emergence of Sustainable Cooperation in a Society of LLM Agents"
Abstract: This study evaluates and extends the findings made by Piatti et al. (2024), who introduced $\texttt{GovSim}$, a simulation framework designed to assess the cooperative decision-making capabilities of large language models (LLMs) in resource-sharing scenarios. By replicating key experiments, we validate claims regarding the performance of large models, such as $\texttt{GPT-4-turbo}$, compared to smaller models. The impact of the universalization principle is also examined, with results showing that large models can achieve sustainable cooperation, with or without the principle, while smaller models fail without it. In addition, we provide multiple extensions to explore the applicability of the framework to new settings. We evaluate additional models, such as $\texttt{DeepSeek-V3}$ and $\texttt{GPT-4o-mini}$, to test whether cooperative behavior generalizes across different architectures and model sizes. Furthermore, we introduce new settings: we create a heterogeneous multi-agent environment, study a scenario using Japanese instructions, and explore an “inverse environment” where agents must cooperate to mitigate harmful resource distributions. Our results confirm that the benchmark can be applied to new models, scenarios, and languages, offering valuable insights into the adaptability of LLMs in complex cooperative tasks. Moreover, the experiment involving heterogeneous multi-agent systems demonstrates that high-performing models can influence lower-performing ones to adopt similar behaviors. This finding has significant implications for other agent-based applications, potentially enabling more efficient use of computational resources and contributing to the development of more effective cooperative AI systems.
URL: https://openreview.net/forum?id=qpu0mtKfh7
---
Title: LINSCAN - A Linearity Based Clustering Algorithm
Abstract: DBSCAN and OPTICS are powerful algorithms for identifying clusters of points in domains where few assumptions can be made about the structure of the data. In this paper, we leverage these strengths and introduce a new algorithm, LINSCAN, designed to seek lineated clusters that are difficult to find and isolate with existing methods. In particular, by embedding points as normal distributions approximating their local neighborhoods and leveraging a distance function derived from the Kullback Leibler Divergence, LINSCAN can detect and distinguish lineated clusters that are spatially close but have orthogonal covariances. We demonstrate how LINSCAN can be applied to seismic data to identify active faults, including intersecting faults, and determine their orientation. Finally, we discuss the properties a generalization of DBSCAN and OPTICS must have in order to retain the stability benefits of these algorithms.
URL: https://openreview.net/forum?id=ihGMVUjoZx
---
Title: Towards Better Understanding of In-Context Learning Ability from In-Context Uncertainty Quantification
Abstract: Predicting simple function classes has been widely used as a testbed for developing theory and understanding of the trained Transformer's in-context learning (ICL) ability. In this paper, we revisit the training of Transformers on linear regression tasks, and different from all the existing literature, we consider a bi-objective prediction task of predicting both the conditional expectation $\mathbb{E}[Y|X]$ and the conditional variance Var$(Y|X)$. This additional uncertainty quantification objective provides a handle to (i) better design out-of-distribution experiments to distinguish ICL from in-weight learning (IWL) and (ii) make a better separation between the algorithms with and without using the prior information of the training distribution. Theoretically, we show that the trained Transformer reaches near Bayes optimum, suggesting the usage of the information of the training distribution. Our method can be extended to other cases. Specifically, with the Transformer's context window $S$, we prove a generalization bound of $\tilde{\mathcal{O}}(\sqrt{\min\{S, T\}/(n T)})$ on $n$ tasks with sequences of length $T$, providing sharper analysis compared to previous results of $\tilde{\mathcal{O}}(\sqrt{1/n})$. Empirically, we illustrate that while the trained Transformer behaves as the Bayes-optimal solution as a natural consequence of supervised training in distribution, it does not necessarily perform a Bayesian inference when facing task shifts, in contrast to the \textit{equivalence} between these two proposed in many existing literature. We also demonstrate the trained Transformer's ICL ability over covariates shift and prompt-length shift and interpret them as a generalization over a meta distribution.
URL: https://openreview.net/forum?id=0c6iG28rRl
---
Title: Rational Tuning of LLM Cascades via Probabilistic Modeling
Abstract: Understanding the reliability of large language models (LLMs) has recently garnered significant attention. Given LLMs' propensity to hallucinate, as well as their high sensitivity to prompt design, it is already challenging to predict the performance of an individual LLM. However, the problem becomes more complex for compound LLM systems such as cascades, where in addition to each model's standalone performance, we must understand how the error rates of different models interact. In this paper, we present a probabilistic model for the joint performance distribution of a sequence of LLMs, which enables a framework for rationally tuning the confidence thresholds of a LLM cascade using continuous optimization. Compared to selecting confidence thresholds using grid search, our parametric Markov-copula model significantly improves runtime scaling with respect to the length of the cascade and the desired resolution of the cost-error curve, turning them from intractable into low-order polynomial. In addition, the optimal thresholds computed using our continuous optimization-based algorithm increasingly outperform those found via grid search as cascade length grows, improving the area under the cost-error curve by 1.9% on average for cascades with $k\geq3$ models. Overall, our Markov-copula model provides a rational basis for tuning LLM cascade performance and points to the potential of probabilistic methods in analyzing LLM systems.
URL: https://openreview.net/forum?id=YCBVcGSZeR
---
Title: Neural Slot Interpreters: Grounding Object Semantics in Emergent Slot Representations
Abstract: Several accounts of human cognition posit that our intelligence is rooted in our ability to form abstract composable concepts, ground them in our environment, and reason over these grounded entities. This trifecta of human thought has remained elusive in modern intelligent machines. In this work, we investigate whether slot representations extracted from visual scenes serve as appropriate compositional abstractions for grounding and reasoning. We present the Neural Slot Interpreter (NSI), which learns to ground object semantics in slots. At the core of NSI is an XML-like schema that uses simple syntax rules to organize the object semantics of a scene into object-centric schema primitives. Then, the NSI metric learns to ground primitives into slots through a structured contrastive learning objective that reasons over the intermodal alignment. Experiments with a bi-modal object-property and scene retrieval task demonstrate the grounding efficacy and interpretability of correspondences learned by NSI. From a scene representation standpoint, we find that emergent NSI slots that move beyond the image grid by binding to spatial objects facilitate improved visual grounding compared to conventional bounding-box-based approaches. From a data efficiency standpoint, we empirically validate that NSI learns more generalizable representations from a fixed amount of annotation data than the traditional approach. We also show that the grounded slots surpass unsupervised slots in real-world object discovery and scale with scene complexity. Finally, we investigate the reasoning abilities of the grounded slots. Vision Transformers trained on grounding-aware NSI tokenizers using as few as ten tokens outperform patch-based tokens on challenging few-shot classification tasks.
URL: https://openreview.net/forum?id=lyxRBPmmnV
---
Title: Importance Weighting for Aligning Language Models under Deployment Distribution Shift
Abstract: Aligning language models (LMs) with human preferences remains challenging partly because popular approaches, such as reinforcement learning from human feedback and direct preference optimization (DPO), often assume that the training data is sufficiently representative of the environment in which the model will be deployed. However, real-world applications frequently involve distribution shifts, e.g., changes in end-user behavior or preferences during usage or deployment, which pose a significant challenge to LM alignment approaches. In this paper, we propose an importance weighting method tailored for DPO, namely IW-DPO, to address distribution shifts in LM alignment. IW-DPO can be applied to joint distribution shifts in the prompts, responses, and preference labels without explicitly assuming the type of distribution shift. Our experimental results on various distribution shift scenarios demonstrate the usefulness of IW-DPO.
URL: https://openreview.net/forum?id=C7QWN4AXvp
---
Title: Prior Learning in Introspective VAEs
Abstract: Variational Autoencoders (VAEs) are a popular framework for unsupervised learning and data generation. A plethora of methods have been proposed focusing on improving VAEs, with the incorporation of adversarial objectives and the integration of prior learning mechanisms being prominent directions. When it comes to the former, an indicative instance is the recently introduced family of Introspective VAEs aiming at ensuring that a low likelihood is assigned to unrealistic samples. In this study, we focus on the Soft-IntroVAE (S-IntroVAE) and investigate the implication of incorporating a multimodal and learnable prior into this framework. Namely, we formulate the prior as a third player and show that when trained in cooperation with the decoder constitutes an effective way for prior learning, which shares the Nash Equilibrium with the vanilla S-IntroVAE. Furthermore, based on a modified formulation of the optimal ELBO in S-IntroVAE, we develop theoretically motivated regularizations, namely (i) adaptive variance clipping to stabilize training when learning the prior and (ii) responsibility regularization to discourage the formation of inactive prior mode. Finally, we perform a series of targeted experiments on a 2D density estimation benchmark and in an image generation setting comprised of the (F)-MNIST and CIFAR-10 datasets demonstrating the benefit of prior learning in S-IntroVAE in generation and representation learning.
URL: https://openreview.net/forum?id=u4YDVFodYX
---
Title: Responsive Noise-Relaying Diffusion Policy: Responsive and Efficient Visuomotor Control
Abstract: Imitation learning is an efficient method for teaching robots a variety of tasks. Diffusion Policy, which uses a conditional denoising diffusion process to generate actions, has demonstrated superior performance, particularly in learning from multi-modal demonstrates. However, it relies on executing multiple actions to retain performance and prevent mode bouncing, which limits its responsiveness, as actions are not conditioned on the most recent observations. To address this, we introduce Responsive Noise-Relaying Diffusion Policy (RNR-DP), which maintains a noise-relaying buffer with progressively increasing noise levels and employs a sequential denoising mechanism that generates immediate, noise-free actions at the head of the sequence, while appending noisy actions at the tail. This ensures that actions are responsive and conditioned on the latest observations, while maintaining motion consistency through the noise-relaying buffer. This design enables the handling of tasks requiring responsive control, and accelerates action generation by reusing denoising steps. Experiments on response-sensitive tasks demonstrate that, compared to Diffusion Policy, ours achieves 18% improvement in success rate. Further evaluation on regular tasks demonstrates that RNR-DP also exceeds the best acceleration method by 6.9%, highlighting its computational efficiency advantage in scenarios where responsiveness is less critical.
URL: https://openreview.net/forum?id=LLWJkR6gaI
---
Title: Leveraging Uncertainty of Pre-trained Models for Fine-Tuning with Search Engine Retrieval
Abstract: Large pre-trained models can dramatically reduce the amount of task-specific data required to solve a problem, but they often fail to capture domain-specific nuances out of the box. The Web likely contains the information necessary to excel on any specific application, but identifying the right data a priori is challenging without knowing where the model is lacking in knowledge.
This paper shows how to leverage recent advances in multi-modal learning to augment a pre-trained model with search engine retrieval. We propose to retrieve useful data from the Web based on instances the model is uncertain about. These uncertain cases are used without access to their labels to generate search queries with varying granularity of descriptiveness. For the final step of retrieval, we propose a geometry-aware refinement technique to discard images unrelated to the task.
We demonstrate substantial performance improvements, e.g. a remarkable increase of 15 percentage points in accuracy on the StanfordCars and Flowers datasets while requiring two orders of magnitude less data compared to the state-of-the-art. We also present extensive experiments giving insights about what to expect of the proposed approach beforehand while exploring the impact of noisy retrieval and different learning strategies.
URL: https://openreview.net/forum?id=3nA8Hymb6l
---
Title: Adaptive Gradient Normalization and Independent Sampling for (Stochastic) Generalized-Smooth Optimization
Abstract: Recent studies have shown that many nonconvex machine learning problems meet generalized-smooth condition that extends beyond traditional smooth nonconvex optimization. However, the existing algorithms cannot fully adapt to generalized-smooth nonconvex geometry and encounter significant technical limitations on convergence analysis. In this work, we first justify the advantage of using adaptive gradient normalization. We analyze the overall effects of adaptive normalization and function geometry on convergence rate. Our results
provide a comprehensive understanding of the interplay between adaptive gradient normalization and function geometry.
For stochastic generalized-smooth nonconvex optimization, we propose Independent-Adaptively Normalized Stochastic Gradient Descent algorithm, which leverages adaptive gradient normalization, independent sampling, and gradient clipping to achieve an $\mathcal{O}(\epsilon^{-4})$ sample complexity under relaxed assumptions. Experiments on large-scale nonconvex generalized-smooth problems demonstrate the fast convergence of our algorithm.
URL: https://openreview.net/forum?id=KKSQQMlEfw
---
Title: Evaluating Long Range Dependency Handling in Code Generation LLMs
Abstract: As language models support larger and larger context sizes, evaluating their ability to make effective use of that context becomes increasingly important. We analyze the ability of several code generation models to handle long range dependencies using a suite of multi-step key retrieval tasks in context windows up to 8k tokens in length. The tasks progressively increase in difficulty and allow more nuanced evaluation of model capabilities than tests like the popular needle-in-the-haystack test. We find that performance degrades significantly for many models (up to 2x) when a function references another function that is defined later in the prompt. We also observe that models that use sliding window attention mechanisms have difficulty handling references further than the size of a single window. We perform simple prompt modifications using call graph information to improve multi-step retrieval performance up to 3x. Our analysis highlights that long-context performance needs more consideration than just retrieval of single facts within a document.
URL: https://openreview.net/forum?id=yzACI2vFaX
---
Title: Why Solving Multi-agent Path Finding with Large Language Models has not Succeeded Yet
Abstract: With the explosive influence caused by the success of large language models (LLM), there has been an extensive amount of recent work showing that foundation models can be used to solve a large variety of tasks. However, there is very limited work that shares insights on multi-agent planning. Multi-agent planning is different from other domains by combining the difficulty of multi-agent coordination and planning, and making it hard to leverage external tools to facilitate the reasoning needed. In this paper, we focus on the problem of multi-agent path finding (MAPF), which is also known as multi-robot route planning, and study the performance of solving MAPF with LLMs. We first show the motivating success of single-agent planning and multi-agent pathfinding in an empty room map without obstacles, then the failure to plan on the harder room map and maze map of the standard MAPF benchmark. We present our position on why directly solving MAPF with LLMs has not been successful yet, and we use various experiments to support our hypothesis. Based on our results, we discussed how researchers with different backgrounds could help with this problem from different perspectives.
URL: https://openreview.net/forum?id=8hAxEFRVQT
---
Title: Shared Imagination: LLMs Hallucinate Alike
Abstract: Despite the recent proliferation of large language models (LLMs), their training recipes -- model architecture, pre-training data and optimization algorithm -- are often very similar. This naturally raises the question of the similarity among the resulting models. In this paper, we propose a novel setting, imaginary question answering (IQA), to better understand model similarity. In IQA, we ask one model to generate purely imaginary questions (e.g., on completely made-up concepts in physics) and prompt another model to answer. Surprisingly, despite the total fictionality of these questions, all models can answer each other's questions with remarkable consistency, suggesting a "shared imagination space" in which these models operate during such hallucinations. We conduct a series of investigations into this phenomenon and discuss the implications of such model homogeneity on hallucination detection and computational creativity. We will release and maintain code and data on a public website.
URL: https://openreview.net/forum?id=NUXpBMtDYs
---
Title: Local Differential Privacy-Preserving Spectral Clustering for General Graphs
Abstract: Spectral clustering is a widely used algorithm to find clusters in networks. Several researchers have studied the stability of spectral clustering under local differential privacy with the additional assumption that the underlying networks are generated from the stochastic block model (SBM). However, we argue that this assumption is too restrictive since social networks do not originate from the SBM.
Thus, we delve into an analysis for general graphs in this work. Our primary focus is the edge flipping method -- a common technique for protecting local differential privacy. We show that, when the edges of an $n$-vertex graph satisfying some reasonable well-clustering assumptions are flipped with a probability of $O(\log n/n)$, the clustering outcomes are largely consistent. Empirical tests further corroborate these theoretical findings. Conversely, although clustering outcomes have been stable for dense and well-clustered graphs produced from the SBM, we show that in general, spectral clustering may yield highly erratic results on certain dense and well-clustered graphs when the flipping probability is $\omega(\log n/n)$. This indicates that the best privacy budget obtainable for general graphs is $\Theta(\log n)$.
URL: https://openreview.net/forum?id=zo5b60AuAH
---
Title: Task-Specific Exploration in Meta-Reinforcement Learning via Task Reconstruction
Abstract: Reinforcement learning trains policies specialized for a single task. Meta-reinforcement learning (meta-RL) improves upon this by leveraging prior experience to train policies for few-shot adaptation to new tasks. However, existing meta-RL approaches often struggle to explore and learn tasks effectively. We introduce a novel meta-RL algorithm for learning to learn task-specific, sample-efficient exploration policies. We achieve this through task reconstruction, an original method for learning to identify and collect small but informative datasets from tasks. To leverage these datasets, we propose a meta-learned hyper-reward that encourages policies to learn to adapt. Empirical evaluations demonstrate that our algorithm adapts to a larger variety of tasks and achieves higher returns than existing meta-RL methods. Additionally, we show that even with full task information, adaptation is more challenging than previously assumed. However, policies trained with our hyper-reward adapt to new tasks successfully.
URL: https://openreview.net/forum?id=1tWfApVmRH
---
Title: What Matters for Model Merging at Scale?
Abstract: Model merging aims to combine multiple expert models into a more capable single model, offering benefits such as reduced storage and serving costs, improved generalization, and support for decentralized model development. Despite its promise, previous studies have primarily focused on merging a few small models. This leaves many unanswered questions about the effect of scaling model size and how it interplays with other key factors—like the base model quality and number of expert models—, to affect the merged model’s performance. This work systematically evaluates the utility of model merging at scale, examining the impact of these different factors. We experiment with merging fully fine-tuned models using four popular merging methods—Averaging, Task Arithmetic, Dare-TIES, and TIESMerging across model sizes ranging from 1B to 64B parameters and merging up to 8 different expert models. We evaluate the merged models on both held-in tasks, i.e., the expert’s training tasks, and zero-shot generalization to unseen held-out tasks. Our wide range of experiments provide several new insights about model merging at scale and the interplay between different factors. First, we find that merging is more effective when experts are created from strong base models, i.e., models with good zero-shot performance, compared to pre-trained ones. Second, larger models facilitate easier merging. Third merging consistently improves generalization capabilities. Notably, when merging eight large expert models, the merged models often generalize better compared to the multitask trained models. Fourth, we can better merge more expert models when working with larger models. Fifth, different merging methods behave very similarly at larger scales. Overall, our findings shed light on some interesting properties of model merging while also highlighting some limitations. We hope that this study will serve as a reference point on large-scale merging.
URL: https://openreview.net/forum?id=HW26XyHp3P
---
Title: Towards improving saliency map interpretability using feature map smoothing
Abstract: Input-gradient-based feature attribution methods, such as Vanilla Gradient, Integrated Gradients, and SmoothGrad, are widely used to explain image classifiers by generating saliency maps. However, these methods struggle to provide explanations that are both visually clear and quantitatively robust. Key challenges include ensuring that explanations are sparse, stable, and faithfully reflect the model’s decision-making. Adversarial training, known for enhancing model robustness, have been shown to produce sparser explanations with these methods; however, this sparsity often comes at the cost of stability. In this work, we investigate the trade-off between stability and sparsity in saliency maps and propose the use of a smoothing layer during adversarial training. Through extensive experiments and evaluation, we demonstrate this smoothing technique improves the stability and faithfulness of saliency maps without sacrificing sparsity. Furthermore, a qualitative user study reveals that human evaluators tend to distrust explanations that are overly noisy or excessively sparse—issues commonly associated with explanations in naturally and adversarially trained models, respectively and prefer explanations produced by our proposed approach. Our findings offer a promising direction for generating reliable explanations with robust models, striking a balance between clarity and usability.
URL: https://openreview.net/forum?id=POaVh2Gu14
---
Title: Epanechnikov Variational Autoencoder
Abstract: In this paper, we bridge Variational Autoencoders (VAEs) and kernel density estimations (KDEs) by approximating the posterior by KDEs and deriving a new lower bound of empirical log likelihood. The flexibility of KDEs makes the optimization of posteriors in VAEs possible, which not only addresses the limitations of Gaussian latent space in vanilla VAE but also provides a new perspective of estimating the KL-divergence term in original evidence lower bound (ELBO). We then propose the Epanechnikov kernel based VAE which enjoys some functional optimality under appropriate conditions. Compared with Gaussian kernel, Epanechnikov kernel has compact support which should make the generated sample less noisy and blurry. The implementation of Epanechnikov kernel in VAE is straightforward as it lies in the "location-scale" family of distributions where the reparametrization tricks can be directly employed. A series of experiments on benchmark datasets such as MNIST, Fashion-MNIST, CIFAR-10 and CelebA further demonstrate the superiority of Epanechnikov Variational Autoenocoder (EVAE) over vanilla VAE and other baseline models in the quality of reconstructed images, as measured by the FID score and Sharpness.
URL: https://openreview.net/forum?id=Qi02eBNCxQ
---
Title: Riemann-Lebesgue Forest for Regression
Abstract: We propose a novel ensemble method called Riemann-Lebesgue Forest (RLF) for regression. The core idea in RLF is to mimic the way how a measurable function can be approximated by partitioning its range into a few intervals. With this idea in mind, we develop a new tree learner named Riemann-Lebesgue Tree (RLT) which has a chance to perform Lebesgue type cutting,i.e splitting the node from response $Y$ at certain non-terminal nodes. In other words, we introduce the "splitting type randomness" in our ensemble method. We show that the optimal Lebesgue type cutting results in larger variance reduction in response $Y$ than ordinary CART cutting (an analogue of Riemann partition). Such property is beneficial to the ensemble part of RLF, which is verified by extensive experiments. We also generalize the asymptotic normality of RLF under different parameter settings. Two one-dimensional examples are provided to illustrate the flexibility of RLF. The competitive performance of RLF against original random forest (RF) and boosting methods such as XGboost is demonstrated by experiments in simulation data and real world datasets. Additional experiments further illustrate that RLF could achieve decent performance comparable to that of RF with less running time, especially in large datasets.
URL: https://openreview.net/forum?id=Gx8ujJTnG9
---
Title: Cluster and Predict Latents Patches for Improved Masked Image Modeling
Abstract: Masked Image Modeling (MIM) offers a promising approach to self-supervised representation learning, however existing MIM models still lag behind the state-of-the-art.
In this paper, we systematically analyze target representations, loss functions, and architectures, to introduce CAPI -- a novel pure-MIM framework that relies on the prediction of latent clusterings.
Our approach leverages a clustering-based loss, which is stable to train, and exhibits promising scaling properties.
Our ViT-L backbone, CAPI, achieves 83.8\% accuracy on ImageNet and 32.1\% mIoU on ADE20K with simple linear probes, substantially outperforming previous MIM methods and approaching the performance of the current state-of-the-art, DINOv2.
URL: https://openreview.net/forum?id=Ycmz7qJxUQ
---
Title: Diffusion Model Predictive Control
Abstract: We propose Diffusion Model Predictive Control (D-MPC), a novel MPC approach that learns a multi-step action proposal and a multi-step dynamics model, both using diffusion models, and combines them for use in online MPC. On the popular D4RL benchmark, we show performance that is significantly better than existing model-based offline planning methods using MPC and competitive with state-of-the-art (SOTA) model-based and model-free reinforcement learning methods. We additionally illustrate D-MPC's ability to optimize novel reward functions at run time and adapt to novel dynamics, and highlight its advantages compared to existing diffusion-based planning baselines.
URL: https://openreview.net/forum?id=pvtgffHtJm
---
Title: Model-free reinforcement learning with noisy actions for automated experimental control in optics
Abstract: Setting up and controlling optical systems is often a challenging and tedious task. The high number of degrees of freedom to control mirrors, lenses, or phases of light makes automatic control challenging, especially when the complexity of the system cannot be adequately modeled due to noise or non-linearities. Here, we show that reinforcement learning (RL) can overcome these challenges when coupling laser light into an optical fiber, using a model-free RL approach that trains directly on the experiment without pre-training. By utilizing the sample-efficient algorithms Soft Actor-Critic (SAC) or Truncated Quantile Critics (TQC), our agent learns to couple with 90% efficiency, comparable to the human expert. We demonstrate that direct training on an experiment can replace extensive system modeling. Our result exemplifies RL's potential to tackle problems in optics, paving the way for more complex applications where full noise modeling is not feasible.
URL: https://openreview.net/forum?id=DAYsM4zDNg
---
Title: Double Horizon Model-Based Policy Optimization
Abstract: Model-based reinforcement learning (MBRL) reduces the cost of
real-environment sampling by generating synthetic trajectories (called
rollouts) from a
learned dynamics model. However, choosing the length of the rollouts
poses two dilemmas:
(1) Longer rollouts better preserve on-policy training
but amplify model bias, indicating the need for an intermediate
horizon to mitigate distribution shift (i.e., the gap between
on-policy and past off-policy samples). (2) Moreover, a longer
model rollout may reduce value estimation bias but raise the variance
of policy gradients due to backpropagation through multiple steps,
implying another intermediate horizon for stable gradient estimates.
However, these two optimal horizons may differ. To resolve this
conflict, we propose Double Horizon Model-Based Policy Optimization
(DHMBPO), which divides the rollout procedure into a long
``distribution rollout'' (DR) and a short ``training rollout'' (TR).
The DR generates on-policy state samples for mitigating distribution
shift. In contrast, the short TR leverages differentiable
transitions to offer accurate value gradient estimation with stable
gradient updates, thereby requiring fewer updates and reducing overall
runtime. We demonstrate that the double-horizon approach effectively
balances distribution shift, model bias, and gradient instability, and
surpasses existing MBRL methods on continuous-control benchmarks in
terms of both sample efficiency and runtime.
URL: https://openreview.net/forum?id=HRvHCd03HM
---
Title: Offset Unlearning for Large Language Models
Abstract: Despite the strong capabilities of Large Language Models (LLMs) to acquire knowledge from their training corpora, the memorization of sensitive information in the corpora such as copyrighted, biased, and private content has led to ethical and legal concerns. In response to these challenges, unlearning has emerged as a potential remedy for LLMs affected by problematic training data. However, previous unlearning techniques are either not applicable to black-box LLMs due to required access to model internal weights, or violate data protection principles by retaining sensitive data for inference-time correction. We propose $\delta$-unlearning, an offset unlearning framework for black-box LLMs. Instead of tuning the black-box LLM itself, $\delta$-unlearning learns the logit offset needed for unlearning by contrasting the logits from a pair of smaller models. Experiments demonstrate that $\delta$-unlearning can effectively unlearn target data while maintaining similar or even stronger performance on general out-of-forget-scope tasks. $\delta$-unlearning also effectively incorporates different unlearning algorithms, making our approach a versatile solution to adapting various existing unlearning algorithms to black-box LLMs.
URL: https://openreview.net/forum?id=A4RLpHPXCu
---
Title: On the Utility of Existing Fine-Tuned Models on Data-Scarce Domains
Abstract: Large Language Models (LLMs) have been observed to perform well on a wide range of downstream tasks when fine-tuned on domain-specific data. However, such data may not be readily available in many applications, motivating zero-shot or few-shot approaches using existing domain or task adjacent (fine-tuned) models, which we call DAFT. While several fine-tuned models for various tasks are available, finding one appropriate DAFT model for a given task is often not straight forward. In this paper, we explore different utilization techniques of these existing DAFT models for data-scarce problems, i.e., tasks for which data is not available or limited. We observe that for zero-shot problems, ensembling of DAFT models provides an accuracy performance close to that of the single best model. With few-shot problems (few data from target domain available), this performance can be improved further by picking or putting more weights to the DAFT models that are expected to perform better on the target task.
URL: https://openreview.net/forum?id=kY2fKLOGkI
---
Title: Investigating Continual Pretraining in Large Language Models: Insights and Implications
Abstract: Continual learning (CL) in large language models (LLMs) is an evolving domain that focuses on developing efficient and sustainable training strategies to adapt models to emerging knowledge and achieve robustness in dynamic environments. Our primary emphasis is on continual domain-adaptive pretraining, a process designed to equip LLMs with the ability to integrate new information from various domains while retaining previously learned knowledge. Since existing works concentrate mostly on continual fine-tuning for a limited selection of downstream tasks or training domains, we introduce a new benchmark designed to measure the adaptability of LLMs to changing pretraining data landscapes. We further examine the impact of model size on learning efficacy and forgetting, as well as how the progression and similarity of emerging domains affect the knowledge transfer within these models.
Our findings uncover several key insights: (i) continual pretraining consistently improves <1.5B models studied in this work and is also superior to domain adaptation, (ii) larger models always achieve better perplexity than smaller ones when continually pretrained on the same corpus, (iii) smaller models are particularly sensitive to continual pretraining, showing the most significant rates of both learning and forgetting, (iv) continual pretraining boosts downstream task performance of GPT-2 family, (v) continual pretraining enables LLMs to specialize better when the sequence of domains shows semantic similarity while randomizing training domains leads to better transfer and final performance otherwise. We posit that our research establishes a new benchmark for CL in LLMs, providing a more realistic evaluation of knowledge retention and transfer across diverse domains.
URL: https://openreview.net/forum?id=aKjJoEVKgO
---
Title: Bad Values but Good Behavior: Learning Highly Misspecified Bandits and MDPs
Abstract: Parametric, feature-based reward models are employed by a variety of algorithms in decision-making settings such as bandits and Markov decision processes (MDPs). The typical assumption under which the algorithms are analysed is realizability, i.e., that the true values of actions are perfectly explained by some parametric model in the class. We are, however, interested in the situation where the true values are (significantly) misspecified with respect to the model class. For parameterized bandits, contextual bandits and MDPs, we identify structural conditions, depending on the problem instance and model class, under which basic algorithms such as $\epsilon$-greedy, LinUCB and fitted Q-learning provably learn optimal policies under even highly misspecified models. This is in contrast to existing worst-case results for, say misspecified bandits, which show regret bounds that incur a linear scaling with time horizon, and shows that there can be a nontrivially large set of bandit instances that are robust to misspecification.
URL: https://openreview.net/forum?id=iB4S16CRKz
---
Title: Ensemble Kalman Diffusion Guidance: A Derivative-free Method for Inverse Problems
Abstract: When solving inverse problems, one increasingly popular approach is to use pre-trained diffusion models as plug-and-play priors. This framework can accommodate different forward models without re-training while preserving the generative capability of diffusion models. Despite their success in many imaging inverse problems, most existing methods rely on privileged information such as derivative, pseudo-inverse, or full knowledge about the forward model. This reliance poses a substantial limitation that restricts their use in a wide range of problems where such information is unavailable, such as in many scientific applications. We propose Ensemble Kalman Diffusion Guidance (EnKG), a derivative-free approach that can solve inverse problems by only accessing forward model evaluations and a pre-trained diffusion model prior. We study the empirical effectiveness of EnKG across various inverse problems, including scientific settings such as inferring fluid flows and astronomical objects, which are highly non-linear inverse problems that often only permit black-box access to the forward model.
URL: https://openreview.net/forum?id=XPEEsKneKs
---
Title: Efficient pooling of predictions via kernel embeddings
Abstract: Probabilistic predictions are probability distributions over the set of possible outcomes. Such predictions quantify the uncertainty in the outcome, making them essential for effective decision making. By combining multiple predictions, the information sources used to generate the predictions are pooled, often resulting in a more informative forecast. Probabilistic predictions are typically combined by linearly pooling the individual predictive distributions; this encompasses several ensemble learning techniques, for example. The weights assigned to each prediction can be estimated based on their past performance, allowing more accurate predictions to receive a higher weight. This can be achieved by finding the weights that optimise a proper scoring rule over some training data. By embedding predictions into a Reproducing Kernel Hilbert Space (RKHS), we illustrate that estimating the linear pool weights that optimise kernel-based scoring rules is a convex quadratic optimisation problem. This permits an efficient implementation of the linear pool when optimally combining predictions on arbitrary outcome domains. This result also holds for other combination strategies, and we additionally study a flexible generalisation of the linear pool that overcomes some of its theoretical limitations, whilst allowing an efficient implementation within the RKHS framework. These approaches are compared in an application to operational wind speed forecasts, where this generalisation is found to offer substantial improvements upon the traditional linear pool.
URL: https://openreview.net/forum?id=dji9MfONNP
---
Title: RefinedFields: Radiance Fields Refinement for Planar Scene Representations
Abstract: Planar scene representations have recently witnessed increased interests for modeling scenes from images, as their lightweight planar structure enables compatibility with image-based models. Notably, K-Planes have gained particular attention as they extend planar scene representations to support in-the-wild scenes, in addition to object-level scenes. However, their visual quality has recently lagged behind that of state-of-the-art techniques. To reduce this gap, we propose RefinedFields, a method that leverages pre-trained networks to refine K-Planes scene representations via optimization guidance using an alternating training procedure. We carry out extensive experiments and verify the merit of our method on synthetic data and real tourism photo collections. RefinedFields enhances rendered scenes with richer details and improves upon its base representation on the task of novel view synthesis. Our code is publicly available as open-source.
URL: https://openreview.net/forum?id=S6JpSsYBDZ
---
Title: Low-Rank Graph Contrastive Learning for Noisy Node Classification
Abstract: Graph Neural Networks (GNNs) have been widely used to learn node representations and with outstanding performance on various tasks such as node classification. However, noise, which inevitably exists in real-world graph data, would considerably degrade the performance of GNNs revealed by recent studies. In this work, we propose a novel and robust GNN encoder, Low-Rank Graph Contrastive Learning (LR-GCL). Our method performs transductive node classification in two steps. First, a low-rank GCL encoder named LR-GCL is trained by prototypical contrastive learning with low-rank regularization. Next, using the features produced by LR-GCL, a linear transductive classification algorithm is used to classify the unlabeled nodes in the graph. Our LR-GCL is inspired by the low frequency property of the graph data and its labels, and it is also theoretically motivated by our sharp generalization bound for transductive learning. To the best of our knowledge, our theoretical result is among the first to theoretically demonstrate the advantage of low-rank learning in graph contrastive learning supported by strong empirical performance. Extensive experiments on public benchmarks demonstrate the superior performance of LR-GCL and the robustness of the learned node representations. The code of LR-GCL is available at \url{https://anonymous.4open.science/r/LRGCL/}.
URL: https://openreview.net/forum?id=6bR6SRgLxx
---
Title: Generalized Flow Matching for Transition Dynamics Modeling
Abstract: Simulating transition dynamics between metastable states is a fundamental challenge in dynamical systems and stochastic processes with wide real-world applications in understanding protein folding, chemical reactions and neural activities. However, the computational challenge often lies on sampling exponentially many paths in which only a small fraction ends in the target metastable state due to existence of high energy barriers. To amortize the cost, we propose a data-driven approach to warm-up the simulation by learning nonlinear interpolations from local dynamics. Specifically, we infer the kinetic energy or ``potential energy'' of the system from local dynamics data. To find plausible paths between two metastable states, we formulate a generalized flow matching framework that learns a vector field to sample probable paths between the two marginal densities under the learned energy function. Furthermore, we iteratively refine the model by assigning importance weights to the sampled paths and buffering more likely paths for training. We validate the effectiveness of the proposed method to sample probable paths on both synthetic and real-world molecular systems.
URL: https://openreview.net/forum?id=L46qxJbOx8
---
Title: Mode-Aware GAN: Continual Adaptation for Conditional Image Generation
Abstract: Continuously learning new modes in generative models while preserving previously learned ones is a significant challenge, particularly with limited training samples. Here, we propose a Mode Affinity Score tailored for continual learning within conditional generative adversarial networks. This score, derived from the discriminators, measures the similarity between generative tasks. By leveraging this score, new modes can be seamlessly integrated into the model through an interpolation process among the closest learned modes, guided by the computed affinity scores. This approach enhances generation performance and mitigates the risk of catastrophic forgetting. Extensive experiments demonstrate the efficacy of our method compared to existing techniques, even when using significantly fewer training samples.
URL: https://openreview.net/forum?id=DxDqc9rNkd
---
Title: Dynamical-VAE-based Hindsight to Learn the Causal Dynamics of Factored-POMDPs
Abstract: Learning the underlying Markovian dynamics of an environment, from partial observations, is a key first step towards model-based reinforcement learning. Considering the environment as a Partially Observable Markov Decision Process (POMDP), state representations are typically inferred from the history of past observations and actions. Instead, we design a Dynamical Variational Auto-Encoder (DVAE) to learn causal Markovian dynamics from offline trajectories in a factored-POMDP setting. In doing so, we derive that incorporating future information is essential to accurately capture causal dynamics and the underlying Markovian states. Our method employs an extended hindsight framework that integrates past, current, and multi-step future information, to infer hidden factors in a principled way, while simultaneously learning transition dynamics as a structural causal model. Our framework is derived from maximizing the log-likelihood of complete trajectories factorized in time and state. Empirical results in a 1-hidden factored-POMDP setting, reveal that this approach uncovers the hidden factor up to a simple transformation, as well as the transition model and causal graph, more effectively than history based, typical 1-step hindsight based, and full trajectory bidirectional-RNN-based models.
URL: https://openreview.net/forum?id=OCD9mxjRyd
---
Title: Synthetic Data is Sufficient for Zero-Shot Visual Generalization from Offline Data
Abstract: Offline reinforcement learning (RL) offers a promising framework for training agents using pre-collected datasets without the need for further environment interaction. However, policies trained on offline data often struggle to generalise due to limited exposure to diverse states.The complexity of visual data introduces additional challenges such as noise, distractions, and spurious correlations, which can misguide the policy and increase the risk of overfitting if the training data is not sufficiently diverse. Indeed, this makes it challenging
to leverage vision-based offline data in training robust agents that can generalize to unseen environments. To solve this problem, we propose a simple approach—generating additional synthetic training data. We propose a two-step process, first augmenting the originally collected offline data to improve zero-shot generalization by introducing diversity, then using a diffusion model to generate additional data in latent space. We test our method across both continuous action spaces (Visual D4RL) and discrete action spaces (Procgen), demonstrating that it significantly improves generalization without requiring any algorithmic changes to existing model-free offline RL methods. We show that our method not only increases the diversity of the training data but also significantly reduces the generalization gap at test time while maintaining computational efficiency. We believe this approach could fuel additional progress in generating synthetic data to train more general agents in the future.
URL: https://openreview.net/forum?id=gFmSFa408D
---
Title: Enhancing deep neural networks through complex-valued representations and Kuramoto synchronization dynamics
Abstract: Neural synchrony is hypothesized to play a crucial role in how the brain organizes visual scenes into structured representations, enabling the robust encoding of multiple objects within a scene. However, current deep learning models often struggle with object binding, limiting their ability to represent multiple objects effectively. Inspired by neuroscience, we investigate whether synchrony-based mechanisms can enhance object encoding in artificial models trained for visual categorization. Specifically, we combine complex-valued representations with Kuramoto dynamics to promote phase alignment, facilitating the grouping of features belonging to the same object. We evaluate two architectures employing synchrony: a feedforward model and a recurrent model with feedback connections to refine phase synchronization using top-down information. Both models outperform their real-valued counterparts and complex-valued models without Kuramoto synchronization on tasks involving multi-object images, such as overlapping handwritten digits, noisy inputs, and out-of-distribution transformations. Our findings highlight the potential of synchrony-driven mechanisms to enhance deep learning models, improving their performance, robustness, and generalization in complex visual categorization tasks.
URL: https://openreview.net/forum?id=zx6QGmBL43
---
Title: Tackling Feature and Sample Heterogeneity in Decentralized Multi-Task Learning: A Sheaf-Theoretic Approach
Abstract: Federated multi-task learning (FMTL) aims to simultaneously learn multiple related tasks across clients without sharing sensitive raw data. However, in the decentralized setting, existing FMTL frameworks are limited in their ability to capture complex task relationships
and handle feature and sample heterogeneity across clients. To address these challenges, we introduce a novel sheaf-theoretic-based approach for FMTL. By representing client relationships using cellular sheaves, our framework can flexibly model interactions between
heterogeneous client models. We formulate the sheaf-based FMTL optimization problem using sheaf Laplacian regularization and propose the Sheaf-FMTL algorithm to solve it. We show that the proposed framework provides a unified view encompassing many existing
federated learning (FL) and FMTL approaches. Furthermore, we prove that our proposed algorithm, Sheaf-FMTL, achieves a sublinear convergence rate in line with state-of-the-art decentralized FMTL algorithms. Extensive experiments demonstrate that Sheaf-FMTL exhibits
communication savings by sending significantly fewer bits compared to decentralized FMTL baselines.
URL: https://openreview.net/forum?id=JlPq0LmApB
---
Title: Seeking Flat Minima with Mean Teacher on Semi- and Weakly-Supervised Domain Generalization for Object Detection
Abstract: Object detectors do not work well when domains largely differ between training and testing data. To overcome this domain gap in object detection without requiring expensive annotations, we consider two problem settings: semi-supervised domain generalizable object detection (SS-DGOD) and weakly-supervised DGOD (WS-DGOD). In contrast to the conventional domain generalization for object detection that requires labeled data from multiple domains, SS-DGOD and WS-DGOD require labeled data only from one domain and unlabeled or weakly-labeled data from multiple domains for training. In this paper, we show that object detectors can be effectively trained on the two settings with the same Mean Teacher learning framework, where a student network is trained with pseudo-labels output from a teacher on the unlabeled or weakly-labeled data. We provide novel interpretations of why the Mean Teacher learning framework works well on the two settings in terms of the relationships between the generalization gap and flat minima in parameter space. On the basis of the interpretations, we also show that incorporating a simple regularization method into the Mean Teacher learning framework leads to flatter minima. The experimental results demonstrate that the regularization leads to flatter minima and boosts the performance of the detectors trained with the Mean Teacher learning framework on the two settings.
URL: https://openreview.net/forum?id=0FQmlp0pgJ
---
Title: Conflict-Aware Adversarial Training
Abstract: Adversarial training is the most effective method to obtain adversarial robustness for deep neural networks by directly involving adversarial samples in the training procedure. To obtain an accurate and robust model, the weighted-average method is applied to optimize standard loss and adversarial loss simultaneously. In this paper, we argue that the weighted-average method does not provide the best tradeoff for standard performance and adversarial robustness. We argue that the failure of the weighted-average method is due to the conflict between gradients derived from standard and adversarial loss, and further demonstrate such a conflict increases with attack budget theoretically and practically. To alleviate this problem, we propose a new trade-off paradigm for adversarial training with a conflict-aware factor for the convex combination of standard and adversarial loss, named Conflict-Aware Adversarial Training (CA-AT). Comprehensive experimental results show that CA-AT consistently offers a superior trade-off between standard performance and adversarial robustness under the settings of adversarial training from scratch and parameter-efficient finetuning.
URL: https://openreview.net/forum?id=PfvgILgAgB
---
Title: DynFed: Dynamic Test-Time Adaptation for Federated Learning with Adaptive Rate Networks
Abstract: Test-Time Personalized Federated Learning (TTPFL) has emerged as a promising approach for adapting models to distribution shifts in federated Learning (FL) environments without relying on labeled data during testing. However, existing methods often struggle with heterogeneous shifts across clients and lack the flexibility to handle diverse distribution changes effectively. In this paper, we introduce \our, a novel algorithm that dynamically optimizes test-time adaptation (TTA) in FL scenarios with heterogeneous distribution shifts.
Our method leverages Adaptive Rate Networks (ARNs) to generate client-specific adaptation rates, enabling more effective handling of diverse shift types, including label skew and feature shifts. \our employs an innovative iterative adaptation process, where ARNs continuously refine adaptation rates based on the current adaptation state, without direct access to raw client data. Crucially, we uncover a fundamental dichotomy: optimal adaptation strategies for one-type and multi-type distribution shifts are diametrically opposed. \our navigates this challenge by automatically adjusting its approach based on the nature of the encountered shifts.
Extensive experiments demonstrate that \our significantly outperforms existing TTPFL and TTA methods across various shift scenarios. Our method shows particularly robust performance in complex multi-type shift environments, where previous approaches often struggle. This work opens new avenues for adaptive and resilient FL in real-world applications where distribution shifts are diverse and unpredictable.
URL: https://openreview.net/forum?id=6Av7iP4AXc
---
Title: Mixture of Balanced Information Bottlenecks for Long-Tailed Visual Recognition
Abstract: Deep neural networks (DNNs) have achieved significant success in various applications with large-scale and balanced data. However, data in real-world visual recognition are usually long-tailed, bringing challenges to efficient training and deployment of DNNs. Information bottleneck (IB) is an elegant approach for representation learning. In this paper, we propose a balanced information bottleneck (BIB) approach, in which loss function re-balancing and self-distillation techniques are integrated into the original IB network. BIB is thus capable of learning a sufficient representation with essential label-related information fully preserved for long-tailed visual recognition. To further enhance the representation learning capability, we also propose a novel structure of mixture of multiple balanced information bottlenecks (MBIB), where different BIBs are responsible for combining knowledge from different network layers. MBIB facilitates an end-to-end learning strategy that trains representation and classification simultaneously from an information theory perspective. We conduct experiments on commonly used long-tailed datasets, including CIFAR100-LT, ImageNet-LT, and iNaturalist 2018. Both BIB and MBIB reach state-of-the-art performance for long-tailed visual recognition.
URL: https://openreview.net/forum?id=9eiALSuZGA
---
Title: Feedback-Guided Black-box Attack in Federated Learning: A Cautious Attacker Perspective
Abstract: Federated Learning (FL) is a robust approach to collaborative machine learning that upholds the integrity of data privacy by ensuring that data remains with the owners. However, FL systems are vulnerable to sophisticated adversarial attacks from malicious clients, especially those leveraging black-box settings. Unlike centralized data poisoning, attacking FL presents unique challenges (i) server-side defense mechanisms can detect and discard suspicious client updates, requiring attacks to maintain minimal visibility across multiple training rounds, and (ii) malicious clients must repeatedly generate poisoned data using only their local black-box model for each round of training, as previous poisoning attempts may be nullified during global aggregation. This forces adversaries to craft stealthy poisoned data locally in a black-box context for each round, maintaining low visibility while ensuring impact. Existing FL attack methods often show high visibility while maintaining impact due to their attack nature, the scale of the introduced perturbations, and the lack of detection strategies. Also, these methods often rely on maximizing cross-entropy loss on the true class, resulting in delayed attack convergence and highly noticeable perturbations. Hence, it is crucial to develop a stealthy data poisoning attack with low visibility for black-box settings in order to comprehend the motives of a cautious attacker in designing an FL attack. To address these challenges, we propose a Feedback-guided Causative Image Black-box Attack (F-CimBA), which is specifically designed for FL by adding random perturbation noise to the data. F-CimBA minimizes the loss of the most confused class (i.e., the incorrect class that the model confuses with the highest probability) instead of the true class, allowing it to exploit local model vulnerabilities for early attack convergence. This approach ensures that poisoned updates maintain low visibility, reducing the likelihood of server-side rejection. Furthermore, F-CimBA adapts effectively under non-IID data distributions and varying attack scenarios, consistently degrading the global model's performance. Additionally, we analyze its impact on system hardware metrics, highlighting the stealth and efficiency of F-CimBA, considering the computational overhead of repeated poisoning attempts in the FL context. Our evaluation demonstrates F-CimBA's consistent ability to poison the global model with minimal visibility under varying attack scenarios and non-IID data distributions, even in the presence of robust server-side defenses.
URL: https://openreview.net/forum?id=C3xxQVfve0
---
Title: Text to Stealthy Adversarial Face Masks
Abstract: Recent studies have demonstrated that modern facial recognition systems, which are based on deep neural networks, are vulnerable to adversarial attacks, including the use of accessories, makeup patterns, or precision lighting. However, developing attacks that are both robust (resilient to changes in viewing angles and environmental conditions) and stealthy (do not attract suspicion by, for example, incorporating obvious facial features) remains a significant challenge. In this context, we introduce a novel diffusion-based method (DAFR) capable of generating robust and stealthy face masks for dodging recognition systems (where the system fails to identify the attacker). Specifically our approach is capable of producing high-fidelity printable textures using the guidance of textual prompts to determine the style. This method can also be adapted for impersonation purposes, where the system misidentifies the attacker as a specific other individual. Finally, we address a gap in the existing literature by presenting a comprehensive benchmark (FAAB) for evaluating adversarial accessories in three dimensions, assessing their robustness and stealthiness.
URL: https://openreview.net/forum?id=XYqCx026AI
---
Title: Sparsity regularization via tree-structured environments for disentangled representations
Abstract: Many causal systems such as biological processes in cells can only be observed indirectly via measurements, such as gene expression. Causal representation learning---the task of correctly mapping low-level observations to latent causal variables---could advance scientific understanding by enabling inference of latent variables such as pathway activation. In this paper, we develop methods for inferring latent variables from multiple related datasets (environments) and tasks. As a running example, we consider the task of predicting a phenotype from gene expression, where we often collect data from multiple cell types or organisms that are related in known ways. The key insight is that the mapping from latent variables driven by gene expression to the phenotype of interest changes sparsely across closely related environments. To model sparse changes, we introduce Tree-Based Regularization (TBR), an objective that minimizes both prediction error and regularizes closely related environments to learn similar predictors. We prove that under assumptions about the degree of sparse changes, TBR identifies the true latent variables up to some simple transformations. We evaluate the theory empirically with both simulations and ground-truth gene expression data. We find that TBR recovers the latent causal variables better than related methods across these settings, even under settings that violate some assumptions of the theory.
URL: https://openreview.net/forum?id=ZzUz0jo200
---