Next Week: CV-DL Seminar – ​Jerry Abu Ayoub | Feature-Set Adaptations For Few-Shot Classification

[Naaman Kopty]

Dear all,

You are invited to attend the next talk in the CV-DL Seminar Series.

Jerry Abu Ayoub
Haifa University

Sunday, June 14, 2026
11:15-12:30
Room 508, Amir Building

Title:
Feature-Set Adaptations For Few-Shot Classification

Abstract:

Few-shot classification (FSC) is central to enabling models to rapidly adapt to novel concepts under severe data constraints. While recent large-scale self-supervised vision transformers (ViTs), such as DINOv2, learn powerful representations that substantially boost few-shot performance, standard nearest-class-mean (NCM) classifiers remain highly sensitive to feature bias and geometric misalignment, limiting their robustness—particularly under domain shift.

In this work, we introduce a geometry-aware task adaptation framework that reshapes the representation space according to the intrinsic manifold of each task. By formulating task adaptation as an entropy-regularized optimal transport problem interleaved with a lightweight adapter, we perform principled feature alignment without updating backbone parameters. Beyond inductive FSC, our approach naturally extends to transductive settings, where it effectively bridges the gap between image and text embeddings in a shared multimodal latent space.

Our contributions are threefold: 
(1) We propose a transductive geometry aware pipeline that mitigates feature bias by modeling balanced pairwise affinities within the task; 
(2) we demonstrate a novel transductive alignment between image and text modalities, significantly improving zero-shot and few-shot inference; and 
(3) we provide an extensive evaluation showing that our framework improves cross-domain FSC and clustering tasks across diverse self-supervised and contrastive pre-trained feature extractors.

Relevant papers:

The Balanced-Pairwise-Affinities Feature Transform

Clip-Adapters

Bio:

I am a computer vision and machine learning researcher focusing on few-shot classification and representation learning, under the supervision of Dr. Simon Korman. During my MSc studies, I extended the BPA framework to foundational model applications and multimodal alignments. I also developed the SpurAudio framework to study spurious correlations in few-shot audio classification. In this seminar, I will discuss my work extending a feature-set transform and its application to few-shot classification on foundational models.

We look forward to seeing you there.

CV-DL Club | University of Haifa