Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

EOL: Transductive Few-Shot Open-Set Recognition by Enhancing Outlier Logits

Mateusz Ochal, Massimiliano Patacchiola, Malik Boudiaf, Sen Wang

TL;DR

This work defines a novel transductive inference technique that leverages the InfoMax principle to exploit the unlabelled query set, called the Enhanced Outlier Logit (EOL) method, which refines class prototype representations through model calibration, effectively balancing the inlier-outlier ratio.

Abstract

In Few-Shot Learning (FSL), models are trained to recognise unseen objects from a query set, given a few labelled examples from a support set. In standard FSL, models are evaluated on query instances sampled from the same class distribution of the support set. In this work, we explore the more nuanced and practical challenge of Open-Set Few-Shot Recognition (OSFSL). Unlike standard FSL, OSFSL incorporates unknown classes into the query set, thereby requiring the model not only to classify known classes but also to identify outliers. Building on the groundwork laid by previous studies, we define a novel transductive inference technique that leverages the InfoMax principle to exploit the unlabelled query set. We called our approach the Enhanced Outlier Logit (EOL) method. EOL refines class prototype representations through model calibration, effectively balancing the inlier-outlier ratio. This calibration enhances pseudo-label accuracy for the query set and improves the optimisation objective within the transductive inference process. We provide a comprehensive empirical evaluation demonstrating that EOL consistently surpasses traditional methods, recording performance improvements ranging from approximately $+1.3%$ to $+6.3%$ across a variety of classification and outlier detection metrics and benchmarks, even in the presence of inlier-outlier imbalance.

EOL: Transductive Few-Shot Open-Set Recognition by Enhancing Outlier Logits

TL;DR

This work defines a novel transductive inference technique that leverages the InfoMax principle to exploit the unlabelled query set, called the Enhanced Outlier Logit (EOL) method, which refines class prototype representations through model calibration, effectively balancing the inlier-outlier ratio.

Abstract

In Few-Shot Learning (FSL), models are trained to recognise unseen objects from a query set, given a few labelled examples from a support set. In standard FSL, models are evaluated on query instances sampled from the same class distribution of the support set. In this work, we explore the more nuanced and practical challenge of Open-Set Few-Shot Recognition (OSFSL). Unlike standard FSL, OSFSL incorporates unknown classes into the query set, thereby requiring the model not only to classify known classes but also to identify outliers. Building on the groundwork laid by previous studies, we define a novel transductive inference technique that leverages the InfoMax principle to exploit the unlabelled query set. We called our approach the Enhanced Outlier Logit (EOL) method. EOL refines class prototype representations through model calibration, effectively balancing the inlier-outlier ratio. This calibration enhances pseudo-label accuracy for the query set and improves the optimisation objective within the transductive inference process. We provide a comprehensive empirical evaluation demonstrating that EOL consistently surpasses traditional methods, recording performance improvements ranging from approximately to across a variety of classification and outlier detection metrics and benchmarks, even in the presence of inlier-outlier imbalance.
Paper Structure (26 sections, 14 equations, 3 figures, 3 tables)

This paper contains 26 sections, 14 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Methodology
  4. Task Definitions
  5. Standard Task
  6. Open-Set Task Variation
  7. Preliminaries
  8. Original OSTIM
  9. Limitations of OSTIM
  10. 1) Analysing the Inlier-Outlier Imbalance Problem
  11. 2) Analysing the Role of the Outlier Logit
  12. 3) Understanding Model Calibration
  13. Enhancing the Outlier Logit: Introducing EOL
  14. Decoupling the Outlier Logit
  15. Representing the Inlier-Outlier Scores
  16. ...and 11 more sections

Figures (3)

  • Figure 1: In this study, we investigate the Transductive Open-Set Few-Shot Learning setting (right). In contrast to the nominal Transductive Few-Shot Learning task (left), the algorithms need to give special consideration to outlier samples present in the unlabelled query set to restrict the information flow from unwanted outliers during transductive inference.
  • Figure 2: Diagram comparing OSTIM and our proposed EOL algorithm. Both methods are transductive, utilising the query set as an additional unlabelled training set. Left: OSTIM implicitly assumes a fixed ratio between inlier and outlier classes, which may be easily disrupted in the real world. Right: Key changes introduced in EOL include: (1) decoupling the inlier-outlier representations from the softmax; (2) introducing the balancing hyperparameter $b$ to account for the contribution of inliers and outliers; (3) reformulating and rebalancing the $\textcolor{teal!90}{\mathring{MA}}$ and $\textcolor{blue!75}{\mathring{CO}}$ loss terms; and (4) improving prediction confidence through model calibration.
  • Figure 3: Study of the balancing hyperparameter, b. We evaluate the performance of EOL with different hyperparameters on different imbalance ratios. For comparison, we show the performance of two other transductive baselines, OSLO boudiaf2023osem, and OSTIM boudiaf2022ostim for each of the imbalanced settings.