Finding Task-Relevant Features for Few-Shot Learning by Category Traversal

By Hongyang Li, David Eigen, Samuel F. Dodge, Matthew D. Zeiler, Xiaogang Wang. 2019

381 citations recorded by the source corpus.

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 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 miniImageNet and tieredImageNet benchmarks, with overall performance competitive with the most recent state-of-the-art systems.

Subfields: Artificial Intelligence, Cancer Research, Computer Vision and Pattern Recognition, Representation & Self-Supervised Learning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Cancer Research

In Computer Vision and Pattern Recognition

In Representation & Self-Supervised Learning

Citations in the corpus

Rethinking Few-Shot Image Classification: a Good Embedding Is All You Need? (2020, background): “Recently, significant advances [54, 51, 49, 11, 43, 45, 52, 33, 41, 58, 26, 28] have been made to tackle this problem using the ideas of meta-learning or “learning to learn”.”
Learning Adaptive Classifiers Synthesis for Generalized Few-Shot Learning (2019, methodology): “It is notable that some comparison methods such as CTM Li et al. (2019) are evaluated over only 600 unseen class FSL tasks, while we test both Castle and aCastle over 10,000 tasks, leading to more stable results.”
Meta-Transfer Learning Through Hard Tasks (2019, methodology): “In this journal version, we implement two deeper architectures – ResNet-18 and ResNet-25, which have been adopted in newly published related works [42], [57], [72], and we achieve the top performance using ResNet-25.”
An Ensemble of Epoch-Wise Empirical Bayes for Few-Shot Learning (2019, background): “23 – – CTM [40] ResNet-18 64.”
Interventional Few-Shot Learning (2020, background): “8 [35] Hongyang Li, David Eigen, Samuel Dodge, Matthew Zeiler, and Xiaogang Wang.”
Few-Shot Learning via Embedding Adaptation With Set-to-Set Functions (2018, methodology): “Additionally, we list some recent methods (SimpleShot [45], and CTM [23]) using different backbone architectures such as ResNet-18 for reference.”
Tailoring Embedding Function to Heterogeneous Few-Shot Tasks by Global and Local Feature Adaptors (2021, methodology): “This line of works projects instances into a task-specific embedding space by feature adaptation (Oreshkin, López, and Lacoste 2018; Ye et al. 2020; Li et al. 2019; Ravichandran, Bhotika, and Soatto 2019).”
Meta-Learning with Shared Amortized Variational Inference (2020, positive): “CTM (Li et al., 2019) is a recent transductive extension to distance-based models, it identiﬁes task-relevant features using inter-and intra-class relations.”
Few-Shot Open-Set Recognition Using Meta-Learning (2020, background): “Few-Shot Learning: Extensive research on few-shot learning [25, 3, 33, 29, 31, 26, 6, 22, 15] has emerged in recent years.”
Relational Generalized Few-Shot Learning (2019, methodology): “Therefore, efficient on-the-fly model adaptation based on very few examples is at the core of most FSL models [11, 13, 14, 23, 35, 39, 45].”
Compositional Few-Shot Recognition with Primitive Discovery and Enhancing (2020, neutral): “45 CTM [36] ResNet18 64.”
Curriculum-Based Meta-learning (2021, background): “A plenty of ProtoNet variants were proposed for improving model generalization ability[11, 21, 22, 35, 45].”
Empirical Bayes Transductive Meta-Learning with Synthetic Gradients (2020, methodology): “CMT (Li et al., 2019) can be seen as an alternative way to incorporate transduction – it measures the similarity between a query example and the support set while making use of intra- and inter-class relationships.”
Adaptive Subspaces for Few-Shot Learning (2020, neutral): “CTM [52] ResNet-18 62.”
A Universal Representation Transformer Layer for Few-Shot Image Classification (2020, background): “, 2019), or masked based on inter-class commonality and inter-class uniqueness (Li et al., 2019).”