[Paper Review] Concept Learners for Few-Shot Learning

Kaidi Cao, Maria Brbić+1 more|arXiv (Cornell University)|Jul 14, 2020

Domain Adaptation and Few-Shot Learning51 references4 citations

TL;DR

This paper proposes COMET, a meta-learning framework that improves few-shot generalization by learning concept-specific metric spaces using human-interpretable concepts. By training independent concept learners on structured concept dimensions and ensembling their predictions, COMET achieves 6–15% relative accuracy gains over state-of-the-art methods on diverse few-shot tasks, including image, text, and biological data, while providing inherently interpretable predictions through concept-based reasoning.

ABSTRACT

Developing algorithms that are able to generalize to a novel task given only a few labeled examples represents a fundamental challenge in closing the gap between machine- and human-level performance. The core of human cognition lies in the structured, reusable concepts that help us to rapidly adapt to new tasks and provide reasoning behind our decisions. However, existing meta-learning methods learn complex representations across prior labeled tasks without imposing any structure on the learned representations. Here we propose COMET, a meta-learning method that improves generalization ability by learning to learn along human-interpretable concept dimensions. Instead of learning a joint unstructured metric space, COMET learns mappings of high-level concepts into semi-structured metric spaces, and effectively combines the outputs of independent concept learners. We evaluate our model on few-shot tasks from diverse domains, including fine-grained image classification, document categorization and cell type annotation on a novel dataset from a biological domain developed in our work. COMET significantly outperforms strong meta-learning baselines, achieving 6-15% relative improvement on the most challenging 1-shot learning tasks, while unlike existing methods providing interpretations behind the model's predictions.

Motivation & Objective

• To address the generalization gap in few-shot learning by incorporating structured, human-interpretable concepts into meta-learning.
• To improve model interpretability in low-data regimes by grounding predictions in high-level, reusable concepts.
• To develop a domain-agnostic method that works with both manually defined and unsupervised concept discovery.
• To demonstrate that structured representation learning along concept dimensions enhances generalization without sacrificing accuracy.
• To create a new benchmark for few-shot learning in single-cell biology using meta-learning.

Proposed method

• COMET learns a separate metric space for each high-level concept using concept-specific neural embedding functions.
• Each concept learner defines prototypes in its own metric space that reflect class-level differences in that concept dimension.
• The model aggregates predictions across multiple concept learners using a weighted ensemble mechanism to improve robustness and accuracy.
• Concept importance is determined via local and global scoring mechanisms, enabling selection of relevant concepts from noisy or redundant sets.
• Concepts can be derived from external knowledge bases or discovered in an unsupervised manner from data.
• The architecture is designed to be inherently interpretable, with explanations derived from the contribution of individual concept learners and their prototypes.

Experimental results

Research questions

• RQ1 Can learning along human-interpretable concept dimensions improve few-shot generalization beyond unstructured metric learning?
• RQ2 How does COMET perform compared to state-of-the-art meta-learning baselines across diverse domains, including biology?
• RQ3 To what extent can COMET leverage unsupervised concept discovery without manual annotations?
• RQ4 Can COMET provide faithful, human-understandable explanations for its predictions in low-data regimes?
• RQ5 Does ensembling concept-specific learners lead to better generalization than joint metric space learning?

Key findings

• COMET achieves a 6–15% relative improvement over strong meta-learning baselines on the most challenging 1-shot learning tasks.
• The model significantly outperforms existing methods on fine-grained image classification, document categorization, and a novel single-cell transcriptome benchmark in biology.
• COMET provides inherently interpretable predictions by highlighting which concepts and prototypes contributed to each decision.
• The method generalizes effectively even when given a large, noisy, or redundant set of initial concepts, automatically selecting the most relevant ones.
• COMET demonstrates strong performance with both manually curated and unsupervised concept discovery, proving its domain-agnostic nature.
• Quantitative and qualitative evaluations confirm that the explanations generated by COMET align with human reasoning and are more reliable than post-hoc explanation methods.