Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Aspect-Based Few-Shot Learning

Tim van Engeland, Lu Yin, Vlado Menkovski

TL;DR

This paper reframes few-shot learning by introducing an aspect-based matching concept, where the correspondence between query and support data is driven by an aspect formed from the support set rather than a fixed class label. It proposes the Deep Set Traversal Module (DSTM), combining a permutation-equivariant neighborhood encoder with a permutation-invariant set aggregator and a reshaper to infer and apply the aspect for context-aware matching. Experimental results on synthetic Geometric Shapes and Sprites datasets show that aspect-based embeddings yield larger positive-to-negative distance ratios than traditional FSL baselines, validating the approach's feasibility in controlled settings. The work highlights the potential of aspect-guided matching for flexible, label-efficient recognition and points to future work with natural images and human-in-the-loop validation.

Abstract

We generalize the formulation of few-shot learning by introducing the concept of an aspect. In the traditional formulation of few-shot learning, there is an underlying assumption that a single "true" label defines the content of each data point. This label serves as a basis for the comparison between the query object and the objects in the support set. However, when a human expert is asked to execute the same task without a predefined set of labels, they typically consider the rest of the data points in the support set as context. This context specifies the level of abstraction and the aspect from which the comparison can be made. In this work, we introduce a novel architecture and training procedure that develops a context given the query and support set and implements aspect-based few-shot learning that is not limited to a predetermined set of classes. We demonstrate that our method is capable of forming and using an aspect for few-shot learning on the Geometric Shapes and Sprites dataset. The results validate the feasibility of our approach compared to traditional few-shot learning.

Aspect-Based Few-Shot Learning

TL;DR

This paper reframes few-shot learning by introducing an aspect-based matching concept, where the correspondence between query and support data is driven by an aspect formed from the support set rather than a fixed class label. It proposes the Deep Set Traversal Module (DSTM), combining a permutation-equivariant neighborhood encoder with a permutation-invariant set aggregator and a reshaper to infer and apply the aspect for context-aware matching. Experimental results on synthetic Geometric Shapes and Sprites datasets show that aspect-based embeddings yield larger positive-to-negative distance ratios than traditional FSL baselines, validating the approach's feasibility in controlled settings. The work highlights the potential of aspect-guided matching for flexible, label-efficient recognition and points to future work with natural images and human-in-the-loop validation.

Abstract

We generalize the formulation of few-shot learning by introducing the concept of an aspect. In the traditional formulation of few-shot learning, there is an underlying assumption that a single "true" label defines the content of each data point. This label serves as a basis for the comparison between the query object and the objects in the support set. However, when a human expert is asked to execute the same task without a predefined set of labels, they typically consider the rest of the data points in the support set as context. This context specifies the level of abstraction and the aspect from which the comparison can be made. In this work, we introduce a novel architecture and training procedure that develops a context given the query and support set and implements aspect-based few-shot learning that is not limited to a predetermined set of classes. We demonstrate that our method is capable of forming and using an aspect for few-shot learning on the Geometric Shapes and Sprites dataset. The results validate the feasibility of our approach compared to traditional few-shot learning.

Paper Structure

This paper contains 20 sections, 6 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Method
  4. Aspect definition
  5. Deep Set Traversal Module
  6. Permutation-equivariant Deep Set Model
  7. Permutation-invariant Deep Set Model
  8. Reshaper
  9. Design choice in DSTM
  10. Evaluation
  11. Data and setup
  12. Data
  13. Support Sets
  14. Data split
  15. Loss
  16. ...and 5 more sections

Figures (3)

  • Figure 1: Examples of different FSL cases that illustrate how taking different aspects results with different matching between the query and support set data points
  • Figure 2: Detailed depiction of the DSTM components
  • Figure 3: Example of the distance between the query and support set instances for the geometric shape data. Differentiating aspects leads to different distances between query and the same image.