Victor Garcia, Joan Bruna
We propose to study the problem of few-shot learning with the prism of inference on a partially observed graphical model, constructed from a collection of input images whose label can be either observed or not. By assimilating generic message-passing inference algorithms with their neural-network counterparts, we define a graph neural network architecture that generalizes several of the recently proposed few-shot learning models. Besides providing improved numerical performance, our framework is easily extended to variants of few-shot learning, such as semi-supervised or active learning, demonstrating the ability of graph-based models to operate well on 'relational' tasks.
| Task | Dataset | Metric | Value | Model |
|---|---|---|---|---|
| Few-Shot Learning | ChestX | 5 shot | 25.27 | GNN |
| Few-Shot Learning | EuroSAT | 5 shot | 83.64 | GNN |
| Few-Shot Learning | ISIC2018 | 5 shot | 43.94 | GNN |
| Image Classification | Stanford Dogs 5-way (5-shot) | Accuracy | 62.27 | GNN++ |
| Image Classification | Stanford Cars 5-way (5-shot) | Accuracy | 71.25 | GNN++ |
| Image Classification | Stanford Cars 5-way (1-shot) | Accuracy | 55.85 | GNN++ |
| Meta-Learning | ChestX | 5 shot | 25.27 | GNN |
| Meta-Learning | EuroSAT | 5 shot | 83.64 | GNN |
| Meta-Learning | ISIC2018 | 5 shot | 43.94 | GNN |
| Few-Shot Image Classification | Stanford Dogs 5-way (5-shot) | Accuracy | 62.27 | GNN++ |
| Few-Shot Image Classification | Stanford Cars 5-way (5-shot) | Accuracy | 71.25 | GNN++ |
| Few-Shot Image Classification | Stanford Cars 5-way (1-shot) | Accuracy | 55.85 | GNN++ |
| Cross-Domain Few-Shot | ChestX | 5 shot | 25.27 | GNN |
| Cross-Domain Few-Shot | EuroSAT | 5 shot | 83.64 | GNN |
| Cross-Domain Few-Shot | ISIC2018 | 5 shot | 43.94 | GNN |