Enhancing Few-Shot Image Classification with Unlabelled Examples

17 Jun 2020  ยท  Peyman Bateni, Jarred Barber, Jan-Willem van de Meent, Frank Wood ยท

We develop a transductive meta-learning method that uses unlabelled instances to improve few-shot image classification performance. Our approach combines a regularized Mahalanobis-distance-based soft k-means clustering procedure with a modified state of the art neural adaptive feature extractor to achieve improved test-time classification accuracy using unlabelled data. We evaluate our method on transductive few-shot learning tasks, in which the goal is to jointly predict labels for query (test) examples given a set of support (training) examples. We achieve state of the art performance on the Meta-Dataset, mini-ImageNet and tiered-ImageNet benchmarks. All trained models and code have been made publicly available at github.com/plai-group/simple-cnaps.

PDF Abstract

Results from the Paper


Task Dataset Model Metric Name Metric Value Global Rank Uses Extra
Training Data
Result Benchmark
Few-Shot Image Classification Meta-Dataset Transductive CNAPS Accuracy 70.32 # 9
Few-Shot Image Classification Meta-Dataset Rank Transductive CNAPS Mean Rank 3.05 # 2
Few-Shot Image Classification Mini-Imagenet 10-way (1-shot) Transductive CNAPS Accuracy 42.8 # 4
Few-Shot Image Classification Mini-Imagenet 10-way (1-shot) Transductive CNAPS + FETI Accuracy 68.5 # 1
Few-Shot Image Classification Mini-Imagenet 10-way (5-shot) Transductive CNAPS + FETI Accuracy 85.9 # 1
Few-Shot Image Classification Mini-Imagenet 10-way (5-shot) Transductive CNAPS Accuracy 59.6 # 4
Few-Shot Image Classification Mini-Imagenet 5-way (1-shot) Transductive CNAPS + FETI Accuracy 79.9 # 12
Few-Shot Image Classification Mini-Imagenet 5-way (1-shot) Transductive CNAPS Accuracy 55.6 # 78
Few-Shot Image Classification Mini-Imagenet 5-way (5-shot) Transductive CNAPS Accuracy 73.1 # 67
Few-Shot Image Classification Mini-Imagenet 5-way (5-shot) Transductive CNAPS + FETI Accuracy 91.5 # 6
Few-Shot Image Classification Tiered ImageNet 10-way (1-shot) Transductive CNAPS + FETI Accuracy 65.1 # 1
Few-Shot Image Classification Tiered ImageNet 10-way (1-shot) Transductive CNAPS Accuracy 54.6 # 3
Few-Shot Image Classification Tiered ImageNet 10-way (5-shot) Transductive CNAPS Accuracy 72.5 # 3
Few-Shot Image Classification Tiered ImageNet 10-way (5-shot) Transductive CNAPS + FETI Accuracy 80.6 # 1
Few-Shot Image Classification Tiered ImageNet 5-way (1-shot) Transductive CNAPS + FETI Accuracy 73.8 # 22
Few-Shot Image Classification Tiered ImageNet 5-way (1-shot) Transductive CNAPS Accuracy 65.9 # 40
Few-Shot Image Classification Tiered ImageNet 5-way (5-shot) Transductive CNAPS + FETI Accuracy 87.7 # 18
Few-Shot Image Classification Tiered ImageNet 5-way (5-shot) Transductive CNAPS Accuracy 81.8 # 39

Methods