Dense Classification and Implanting for Few-Shot Learning

Abstract

Training deep neural networks from few examples is a highly challenging and key problem for many computer vision tasks. In this context, we are targeting knowledge transfer from a set with abundant data to other sets with few available examples. We propose two simple and effective solutions: (i) dense classification over feature maps, which for the first time studies local activations in the domain of few-shot learning, and (ii) implanting, that is, attaching new neurons to a previously trained network to learn new, taskspecific features. On miniImageNet, we improve the prior state-of-the-art on few-shot classification, i.e., we achieve 62.5%, 79.8% and 83.8% on 5-way 1-shot, 5-shot and 10shot settings respectively. 1 We use the notation [i] := {1, . . . , i} for i ∈ N.

Cooperative Bi-path Metric for Few-shot Learning

“…During training phase, tasks were sampled, then loss function of the model was calculated based on the tasks, and the network parameters were updated through the back propagation. However, some works [18][19][20][21] did not follow this setting, but trained classification networks in the way of traditional supervised learning.…”

“…To make fair comparisons with the previous works [13,14,16,17,20], we adopt ResNet-12 [3] as the baseline's backbone and use the training set D base to optimize it in a fully-supervised manner. However, it is interesting to find that a network with only a supervised training strategy can also get superior classification results on novel classes (elaborated in Sect.…”

“…2) Temperature in learning: inspired by [39], we also introduce a hyper-parameter called temperature, which was first applied in model distillation to change the smoothness of distribution after softmax normalization and the value of crossentropy. 3) Dense classification: instead of embedding the image features as a vector, we apply dense classification loss [20] to regularize our model, i.e., all the local feature vectors of the feature map before the last fully-connected layer are classified through the fully-connected layer without average pooling. For each training sample (x, y) ∈ D base , the proposed baseline with loss L has the following form:…”

“…Leading by the pioneer work [4], many subsequent researches [5][6][7][8][9][10][11][12][13][14][15][16][17] followed this episodic learning and achieved notable improvements. However, some recent works [18][19][20][21] did not follow this factitious sampling setting, but directly trained model in the way of traditional supervised learning. Thus a natural concern arises, is episodic training necessary for fewshot learning?…”

“…This kind of method classifies query samples by learning a feature extractor on the base classes, extracting features of samples of the novel classes during testing, and measuring the distance or similarity between labeled support samples and unlabeled query samples. However, most of the existing metric learning methods [4,9,11,[14][15][16][17]20] for fewshot learning focused on the correlation between support samples and query samples within novel classes, and did not take full use of the information of base classes. Unlike the above-mentioned approaches, we make use of the base classes to assist in classifying query samples of novel classes during the testing phase.…”