- This is my final year project at Monash University.
- This project is supervised by Dr Mehrtash Harandi and Dr Masoud Faraki.
- ECSE Spark Night 2020 website: https://www.ecsespark.com/draft-b/few-shot-learning
Few-shot learning experiments include:
- Siamese network [paper]
- Graph Neural Network (GNN) [paper]
- Model-Agnostic Meta-Learning (MAML) [paper]
The overall results are summarized below.
| 1 shot 5 ways | 1 shot 20 ways | |
|---|---|---|
| Siamese Net [1] | 95.3% | 85.7% |
| MAML [2] | 99.2% | 96.6% |
| GNN [3] | 99.3% | 97.8% |
| 5 shots 5 ways | 5 shots 20 ways | |
|---|---|---|
| MAML | 99.6% | 98.0% |
| GNN | 99.7% | 98.5% |
| 5 ways | 20 ways |
|---|---|
 |
 |
 |
 |
| 5 ways | 20 ways |
|---|---|
 |
 |
 |
 |
[1] G. Koch, R.Zemel and R.Salakhutdinov, “Siamese Neural Networks for One-shot Image Recognition”, Proceedings of the 32nd International Conference on Machine Learning, vol 37, 2015.
[2] C. Finn, P. Abbeel and S. Levine,“Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks”, Proceedings of the 34th International Conference on Machine Learning (PMLR), 2017.
[3] V. Garcia and J. Bruna, “Few-Shot Learning with Graph Neural Networks”, Sixth International Conference on Learning Representations (ICLR), 2018.