[Paper Review] Finding Task-Relevant Features for Few-Shot Learning by Category Traversal

Hongyang Li, David Eigen+3 more|arXiv (Cornell University)|May 27, 2019

Domain Adaptation and Few-Shot Learning41 references96 citations

TL;DR

The paper introduces a Category Traversal Module (CTM) that looks across all support classes to identify task-relevant feature dimensions, improving metric-based few-shot learning performance by about 5–10% on miniImageNet and tieredImageNet. CTM is plug-and-play and boosts several baselines.

ABSTRACT

Few-shot learning is an important area of research. Conceptually, humans are readily able to understand new concepts given just a few examples, while in more pragmatic terms, limited-example training situations are common in practice. Recent effective approaches to few-shot learning employ a metric-learning framework to learn a feature similarity comparison between a query (test) example, and the few support (training) examples. However, these approaches treat each support class independently from one another, never looking at the entire task as a whole. Because of this, they are constrained to use a single set of features for all possible test-time tasks, which hinders the ability to distinguish the most relevant dimensions for the task at hand. In this work, we introduce a Category Traversal Module that can be inserted as a plug-and-play module into most metric-learning based few-shot learners. This component traverses across the entire support set at once, identifying task-relevant features based on both intra-class commonality and inter-class uniqueness in the feature space. Incorporating our module improves performance considerably (5%-10% relative) over baseline systems on both mini-ImageNet and tieredImageNet benchmarks, with overall performance competitive with recent state-of-the-art systems.

Motivation & Objective

• Motivate the need for task-aware feature selection in few-shot learning where support classes are treated independently.
• Develop a plug-and-play module that leverages intra-class commonality and inter-class uniqueness to identify task-relevant feature dimensions.
• Integrate CTM with existing metric-based few-shot learners to improve discriminativeness of embeddings.
• Demonstrate CTM’s effectiveness through extensive ablations and comparisons on standard benchmarks.

Proposed method

• Introduce CTM consisting of a concentrator (intra-class commonality) and a projector (inter-class uniqueness).
• Concentrator reduces dimensionality and aggregates within-class features to produce a class-wise embedding ‘o’.
• Projector aggregates across classes to produce a mask ‘p’ that selects task-relevant feature dimensions.
• Apply the mask to support and query embeddings to obtain improved features I(S) and I(Q).
• Embed CTM into existing metric-based few-shot learners (Matching Net, Prototypical Net, Relation Net) by replacing their similarity with M(r(S) ⊙ p, r(Q) ⊙ p).
• Train and evaluate in episodic few-shot settings on miniImageNet and tieredImageNet with standard 5-way/1–5-shot setups.

Experimental results

Research questions

• RQ1 Can a cross-category view of the support set improve the identification of task-relevant features in few-shot learning?
• RQ2 Does CTM improve existing metric-based few-shot learners, and by how much across benchmarks?
• RQ3 Which CTM components (concentrator, projector) are essential for performance gains?
• RQ4 How does CTM affect the learned feature space in terms of discriminability?

Key findings

• CTM yields consistent relative gains of about 5–10% on miniImageNet and tieredImageNet when integrated with existing metric-based methods.
• CTM improves Matching Net, Prototypical Net, and Relation Net by several percentage points across 1-shot and 5-shot settings.
• A deeper backbone (ResNet-18) with CTM substantially increases performance (e.g., 1-shot miniImageNet improves to 62.05% with CTM in Table 4).
• In ablations, removing the concentrator or the projector degrades performance, confirming both components are necessary for best results.
• CTM’s data-augmented version further improves results (e.g., miniImageNet 1-shot 64.12%, tiered 1-shot 68.41% with augmentation).
• t-SNE visualizations show tighter, more discriminative clusters after applying CTM-based masking.