Weekly TMLR digest for Oct 02, 2022
TMLR
Accepted papers
===============
Title: MixTailor: Mixed Gradient Aggregation for Robust Learning Against Tailored Attacks
Authors: Ali Ramezani-Kebrya, Iman Tabrizian, Fartash Faghri, Petar Popovski
Abstract: Implementations of SGD on distributed and multi-GPU systems creates new vulnerabilities, which can be identified and misused by one or more adversarial agents. Recently, it has been shown that well-known Byzantine-resilient gradient aggregation schemes are indeed vulnerable to informed attackers that can tailor the attacks (Fang et al., 2020; Xie et al., 2020b). We introduce MixTailor, a scheme based on randomization of the aggregation strategies that makes it impossible for the attacker to be fully informed. Deterministic schemes can be integrated into MixTailor on the fly without introducing any additional hyperparameters. Randomization decreases the capability of a powerful adversary to tailor its attacks, while the resulting randomized aggregation scheme is still competitive in terms of performance. For both iid and non-iid settings, we establish almost sure convergence guarantees that are both stronger and more general than those available in the literature. Our empirical studies across various datasets, attacks, and settings, validate our hypothesis and show that MixTailor successfully defends when well-known Byzantine-tolerant schemes fail.
URL: https://openreview.net/forum?id=tqDhrbKJLS
---
Title: LIMIS: Locally Interpretable Modeling using Instance-wise Subsampling
Authors: Jinsung Yoon, Sercan O Arik, Tomas Pfister
Abstract: Understanding black-box machine learning models is crucial for their widespread adoption. Learning globally interpretable models is one approach, but achieving high performance with them is challenging. An alternative approach is to explain individual predictions using locally interpretable models. For locally interpretable modeling, various methods have been proposed and indeed commonly used, but they suffer from low fidelity, i.e. their explanations do not approximate the predictions well. In this paper, our goal is to push the state-of-the-art in high-fidelity locally interpretable modeling. We propose a novel framework, Locally Interpretable Modeling using Instance-wise Subsampling (LIMIS). LIMIS utilizes a policy gradient to select a small number of instances and distills the black-box model into a low-capacity locally interpretable model using those selected instances. Training is guided with a reward obtained directly by measuring the fidelity of the locally interpretable models. We show on multiple tabular datasets that LIMIS near-matches the prediction accuracy of black-box models, significantly outperforming state-of-the-art locally interpretable models in terms of fidelity and prediction accuracy.
URL: https://openreview.net/forum?id=S8eABAy8P3
---
Title: Sparse MoEs meet Efficient Ensembles
Authors: James Urquhart Allingham, Florian Wenzel, Zelda E Mariet, Basil Mustafa, Joan Puigcerver, Neil Houlsby, Ghassen Jerfel, Vincent Fortuin, Balaji Lakshminarayanan, Jasper Snoek, Dustin Tran, Carlos Riquelme Ruiz, Rodolphe Jenatton
Abstract: Machine learning models based on the aggregated outputs of submodels, either at the activation or prediction levels, often exhibit strong performance compared to individual models. We study the interplay of two popular classes of such models: ensembles of neural networks and sparse mixture of experts (sparse MoEs). First, we show that the two approaches have complementary features whose combination is beneficial. This includes a comprehensive evaluation of sparse MoEs in uncertainty related benchmarks. Then, we present efficient ensemble of experts (E$^3$), a scalable and simple ensemble of sparse MoEs that takes the best of both classes of models, while using up to 45% fewer FLOPs than a deep ensemble. Extensive experiments demonstrate the accuracy, log-likelihood, few-shot learning, robustness, and uncertainty improvements of E$^3$ over several challenging vision Transformer-based baselines. E$^3$ not only preserves its efficiency while scaling to models with up to 2.7B parameters, but also provides better predictive performance and uncertainty estimates for larger models.
URL: https://openreview.net/forum?id=i0ZM36d2qU
---
Title: FedShuffle: Recipes for Better Use of Local Work in Federated Learning
Authors: Samuel Horváth, Maziar Sanjabi, Lin Xiao, Peter Richtárik, Michael Rabbat
Abstract: The practice of applying several local updates before aggregation across clients has been empirically shown to be a successful approach to overcoming the communication bottleneck in Federated Learning (FL). Such methods are usually implemented by having clients perform one or more epochs of local training per round while randomly reshuffling their finite dataset in each epoch. Data imbalance, where clients have different numbers of local training samples, is ubiquitous in FL applications, resulting in different clients performing different numbers of local updates in each round. In this work, we propose a general recipe, FedShuffle, that better utilizes the local updates in FL, especially in this regime encompassing random reshuffling and heterogeneity. FedShuffle is the first local update method with theoretical convergence guarantees that incorporates random reshuffling, data imbalance, and client sampling — features that are essential in large-scale cross-device FL. We present a comprehensive theoretical analysis of FedShuffle and show, both theoretically and empirically, that it does not suffer from the objective function mismatch that is present in FL methods that assume homogeneous updates in heterogeneous FL setups, such as FedAvg (McMahan et al., 2017). In addition, by combining the ingredients above, FedShuffle improves upon FedNova (Wang et al., 2020), which was previously proposed to solve this mismatch. Similar to Mime (Karimireddy et al., 2020), we show that FedShuffle with momentum variance reduction (Cutkosky & Orabona, 2019) improves upon non-local methods under a Hessian similarity assumption.
URL: https://openreview.net/forum?id=Lgs5pQ1v30
---
Title: Representation Alignment in Neural Networks
Authors: Ehsan Imani, Wei Hu, Martha White
Abstract: It is now a standard for neural network representations to be trained on large, publicly available datasets, and used for new problems. The reasons for why neural network representations have been so successful for transfer, however, are still not fully understood. In this paper we show that, after training, neural network representations align their top singular vectors to the targets. We investigate this representation alignment phenomenon in a variety of neural network architectures and find that (a) alignment emerges across a variety of different architectures and optimizers, with more alignment arising from depth (b) alignment increases for layers closer to the output and (c) existing high-performance deep CNNs exhibit high levels of alignment. We then highlight why alignment between the top singular vectors and the targets can speed up learning and show in a classic synthetic transfer problem that representation alignment correlates with positive and negative transfer to similar and dissimilar tasks.
URL: https://openreview.net/forum?id=fLIWMnZ9ij
---
New submissions
===============
Title: InfoNCE is a variational autoencoder
Abstract: Unsupervised learning typically involves learning a full generative model of the inputs, and goes back at least to the Boltzmann (Ackley et al. 1985) and Helmholtz machines (Dayan et al. 1995). More recently, it was noted that we can get good representations of unlabelled datapoints without the need to learn a full generative model, and this gave birth to the modern field of self-supervised learning (SSL). We reconcile these critically important families of machine learning method, by showing that modern SSL methods including InfoNCE which maximize mutual information are equivalent to a particular unsupervised learning method, the variational autoencoder. Additionally, our approach resolves mysteries purely in SSL. In particular, recent work (Tschannen et al., 2019) has argued that mutual information objectives can give arbitrarily entangled representations. Instead, they argue that the excellent performance of InfoNCE arises from the use of a simplified, linear mutual information estimator. How can we understand the success of InfoNCE if better mutual information estimators lead to worse representations? Remarkably, under one choice of prior, the VAE objective (i.e. the ELBO) is exactly equal to the mutual information (up to constants). Under an alternative choice of prior, the SSVAE objective is exactly equal to the simplified parametric mutual information estimator used in InfoNCE (up to constants). As such, the SSVAE framework naturally provides a principled justification for using simplified mututal information estimators, because they are equivalent to structured priors.
URL: https://openreview.net/forum?id=SGNIcTOtvG
---
Title: A Ranking Game for Imitation Learning
Abstract: We propose a new framework for imitation learning---treating imitation as a two-player ranking-based game between a policy and a reward. In this game, the reward agent learns to satisfy pairwise performance rankings between behaviors, while the policy agent learns to maximize this reward. In imitation learning, near-optimal expert data can be difficult to obtain, and even in the limit of infinite data cannot imply a total ordering over trajectories as preferences can. On the other hand, learning from preferences alone is challenging as a large number of preferences are required to infer a high-dimensional reward function, though preference data is typically much easier to collect than expert demonstrations. The classical inverse reinforcement learning (IRL) formulation learns from expert demonstrations but provides no mechanism to incorporate learning from offline preferences and vice versa. We instantiate the proposed ranking-game framework with a novel ranking loss giving an algorithm that can simultaneously learn from expert demonstrations and preferences, gaining the advantages of both modalities. Our experiments show that the proposed method achieves state-of-the-art sample efficiency and can solve previously unsolvable tasks in the Learning from Observation (LfO) setting.
URL: https://openreview.net/forum?id=d3rHk4VAf0
---
Title: Fairness and robustness in anti-causal prediction
Abstract: Robustness to distribution shift and fairness have independently emerged as two important desiderata required of modern machine learning models. While these two desiderata seem related, the connection between them is often unclear in practice. Here, we discuss these connections through a causal lens, focusing on anti-causal prediction tasks, where the input to a classifier (e.g., an image) is assumed to be generated as a function of the target label and the protected attribute. By taking this perspective, we draw explicit connections between a common fairness criterion - separation - and a common notion of robustness - risk invariance. These connections provide new motivation for applying the separation criterion in anticausal settings, and inform old discussions regarding fairness-performance tradeoffs. In addition, our findings suggest that robustness-motivated approaches can be used to enforce separation, and that they often work better in practice than methods designed to directly enforce separation. Using a medical dataset, we empirically validate our findings on the task of detecting pneumonia from X-rays, in a setting where differences in prevalence across sex groups motivates a fairness mitigation. Our findings highlight the importance of considering causal structure when choosing and enforcing fairness criteria.
URL: https://openreview.net/forum?id=mrTXGDZns2
---
Title: A Unified Domain Adaptation Framework with Distinctive Divergence Analysis
Abstract: Unsupervised domain adaptation enables knowledge transfer from a labeled source domain to an unlabeled target domain by aligning the learnt features of both domains. The idea is theoretically supported by the generalization bound analysis in Ben-David et al. (2007), which specifies the applicable task (binary classification) and designates a specific distribution divergence measure. Although most domain adaptation models seek theoretical grounds from this particular bound analysis, they do not actually fit into the stringent conditions. In this paper, we bridge the long-standing theoretical gap in literature by providing a unified generalization bound. Our analysis can well accommodate the classification/regression tasks and most commonly-used divergence measures, and more importantly, it can theoretically recover a large amount of previous models. In addition, we identify the key difference in the distribution divergence measures underlying the diverse models and commit a comprehensive in-depth comparison of the commonly-used divergence measures. Based on the unified generalization bound, we propose new domain adaptation models and conduct experiments on real-world datasets that corroborate our theoretical findings. We believe these insights are helpful in guiding the future design of domain adaptation algorithms.
URL: https://openreview.net/forum?id=yeT9cBq8Cn
---
Title: Separable Self-attention for Mobile Vision Transformers
Abstract: Mobile vision transformers (MobileViT) can achieve state-of-the-art performance across several mobile vision tasks, including classification and detection. Though these models have fewer parameters, they have high latency as compared to convolutional neural network-based models. The main efficiency bottleneck in MobileViT is the multi-headed self-attention (MHA) in transformers, which requires $O(k^2)$ time complexity with respect to the number of tokens (or patches) $k$. Moreover, MHA requires costly operations (e.g., batch-wise matrix multiplication) for computing self-attention, impacting latency on resource-constrained devices. This paper introduces a separable self-attention method with linear complexity, i.e. $O(k)$. A simple yet effective characteristic of the proposed method is that it uses element-wise operations for computing self-attention, making it a good choice for resource-constrained devices. The improved model, MobileViTv2, is state-of-the-art on several mobile vision tasks, including ImageNet object classification and MS-COCO object detection. With about three million parameters, MobileViTv2 achieves a top-1 accuracy of 75.6% on the ImageNet dataset, outperforming MobileViT by about 1% while running $3.2\times$ faster on a mobile device.
URL: https://openreview.net/forum?id=tBl4yBEjKi
---
Title: Causal Inference from Small High-dimensional Datasets
Abstract: Many methods have been proposed to estimate treatment effects with observational data. Often, the choice of the method considers the application's characteristics, such as type of treatment and outcome, confounding effect, and the complexity of the data. These methods implicitly assume that the sample size is large enough to train such models, especially the neural network-based estimators. What if this is not the case? In this work, we propose Causal-Batle, a methodology to estimate treatment effects in small high-dimensional datasets in the presence of another high-dimensional dataset in the same feature space. We adopt an approach that brings transfer learning techniques into causal inference. Our experiments show that such an approach helps to bring stability to neural network-based methods and improve the treatment effect estimates in small high-dimensional datasets. The code for our method and all our experiments is available at \url{github.com/HiddenForAnonymization}.
URL: https://openreview.net/forum?id=LcizPy6xhW
---