Robust 5-way 5-shot Classification on MiniImageNet (Outlier Noise)
68.51Accuracy (0% Noise)Similarity weighted prototypes (Euclidean)
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Similarity weighted prototypes (Euclidean)Aggregation method=Euclidean distance weighting, Backbone=4-layer convolutional2022.04 | 68.51 | 64.57 | 58.01 | 47.25 | |
| Spatial median prototypes (Median)Aggregation method=Median, Backbone=4-layer convolutional2022.04 | 68.37 | 64.46 | 57.85 | 47.19 | |
| Similarity weighted prototypes (Cosine)Aggregation method=Cosine similarity weighting, Backbone=4-layer convolutional2022.04 | 68.2 | 64.78 | 58.36 | 47.34 | |
| OracleProtocol=5-way 5-shot2022.04 | 68.18 | 66.08 | 62.6 | 56.89 | |
| Vanilla ProtoNetBackbone=4-layer convolutional2022.04 | 68.18 | 63.92 | 51.11 | 40.11 | |
| RNNPBackbone=4-layer convolutional2022.04 | 68.17 | 63.8 | 56.97 | 46.92 | |
| Similarity weighted prototypes (Absolute)Aggregation method=Absolute distance weighting, Backbone=4-layer convolutional2022.04 | 68.13 | 64.69 | 58.3 | 47.39 | |
| TraNFS-3Number of Layers=3, Backbone=4-layer convolutional2022.04 | 68.11 | 64.96 | 59.03 | 47.69 | |
| Baseline++Backbone=4-layer convolutional2022.04 | 67.85 | 63.49 | 57.07 | 46.99 | |
| TraNFS-2Number of Layers=2, Backbone=4-layer convolutional2022.04 | 67.76 | 64.47 | 58.29 | 47.37 | |
| Linear ClassifierBackbone=4-layer convolutional2022.04 | 66.7 | 61.13 | 53.86 | 44.05 | |
| MAMLBackbone=4-layer convolutional2022.04 | 63.21 | 57.35 | 50 | 40.9 | |
| Matching NetworksBackbone=4-layer convolutional2022.04 | 62.05 | 57.69 | 51.32 | 42.39 | |
| Nearest k = 5k=5, Backbone=4-layer convolutional2022.04 | 56.34 | 52.32 | 46.49 | 38.44 | |
| Nearest k = 1k=1, Backbone=4-layer convolutional2022.04 | 55.87 | 50.9 | 45.28 | 38.75 | |
| Nearest k = 3k=3, Backbone=4-layer convolutional2022.04 | 55.28 | 50.53 | 44.4 | 37.03 |