Text Descriptions are Compressive and Invariant Representations for Visual Learning

TL;DR

This paper introduces SLR-AVD, a novel method that uses multiple visual descriptions generated by language models and sparse logistic regression to create invariant, compressed visual features, improving robustness in few-shot learning.

Contribution

The paper proposes SLR-AVD, a new approach combining language-generated visual descriptions with sparse regression for invariant feature extraction in visual learning.

Findings

Descriptive features are more invariant to domain shifts.

SLR-AVD outperforms state-of-the-art finetuning methods.

Invariant features improve in- and out-of-distribution performance.

Abstract

Modern image classification is based upon directly predicting classes via large discriminative networks, which do not directly contain information about the intuitive visual features that may constitute a classification decision. Recently, work in vision-language models (VLM) such as CLIP has provided ways to specify natural language descriptions of image classes, but typically focuses on providing single descriptions for each class. In this work, we demonstrate that an alternative approach, in line with humans' understanding of multiple visual features per class, can also provide compelling performance in the robust few-shot learning setting. In particular, we introduce a novel method, \textit{SLR-AVD (Sparse Logistic Regression using Augmented Visual Descriptors)}. This method first automatically generates multiple visual descriptions of each class via a large language model (LLM), then uses a VLM to translate these descriptions to a set of visual feature embeddings of each image, and finally uses sparse logistic regression to select a relevant subset of these features to classify each image. Core to our approach is the fact that, information-theoretically, these descriptive features are more invariant to domain shift than traditional image embeddings, even though the VLM training process is not explicitly designed for invariant representation learning. These invariant descriptive features also compose a better input compression scheme. When combined with finetuning, we show that SLR-AVD is able to outperform existing state-of-the-art finetuning approaches on both in-distribution and out-of-distribution performance.