A Universal Representation Transformer Layer for Few-Shot Image Classification

By Lu Liu, William L. Hamilton, Guodong Long, Jing Jiang, Hugo Larochelle. 2020

144 citations recorded by the source corpus.

Few-shot classification aims to recognize unseen classes when presented with only a small number of samples. We consider the problem of multi-domain few-shot image classification, where unseen classes and examples come from diverse data sources. This problem has seen growing interest and has inspired the development of benchmarks such as Meta-Dataset. A key challenge in this multi-domain setting is to effectively integrate the feature representations from the diverse set of training domains. Here, we propose a Universal Representation Transformer (URT) layer, that meta-learns to leverage universal features for few-shot classification by dynamically re-weighting and composing the most appropriate domain-specific representations. In experiments, we show that URT sets a new state-of-the-art result on Meta-Dataset. Specifically, it achieves top-performance on the highest number of data sources compared to competing methods. We analyze variants of URT and present a visualization of the attention score heatmaps that sheds light on how the model performs cross-domain generalization. Our code is available at this https URL.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition

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 Computer Vision and Pattern Recognition

Citations in the corpus

FedProto: Federated Prototype Learning across Heterogeneous Clients (2021, methodology): “Inspired by prototype learning, merging the prototypes over heterogeneous datasets can effectively integrate the feature representations from diverse data distributions (Snell, Swersky, and Zemel 2017; Liu et al. 2020; Dvornik, Schmid, and Mairal 2020).”
Improving Task Adaptation for Cross-domain Few-shot Learning (2021, methodology): “While the methods addressing the first and second challenges are typically applied either in meta-train or meta-test respectively, few methods [11, 24, 42] perform both feature selection in meta-training and also finetuning their feature selection mechanisms in meta-testing.”
Universal Representation Learning from Multiple Domains for Few-shot Classification (2021, methodology): “As in [13, 29], we solve this problem in two steps: i) a meta-training step where a learning algorithm receives a large dataset Db and outputs a general feature extractor f , ii) a meta-test step where the target tasks (S,Q) are sampled from another large dataset Dt by taking the subsets of the data”
Federated Learning from Pre-Trained Models: A Contrastive Learning Approach (2022, methodology): “They are pre-trained on Quick Draw [79], Aircraft [80], and CU-Birds [19] public dataset, respectively.”
A Comprehensive Survey of Few-shot Learning: Evolution, Applications, Challenges, and Opportunities (2022, methodology): “[90] uses attention as a metric strategy to reweight and combine domain-specific representations.”
Learning a Universal Template for Few-shot Dataset Generalization (2021, methodology): “We compare against CNAPs (Requeima et al., 2019), TaskNorm (Bronskill et al., 2020) 3 , SimpleCNAPS (Bateni et al., 2020), SUR(-pf) (Dvornik et al., 2020) and URT(-pf) (Liu et al., 2021).”