Semi Supervised Learning For Few-shot Audio Classification By Episodic Triplet Mining

Abstract

Few-shot learning aims to generalize unseen classes that appear during testing but are unavailable during training. Prototypical networks incorporate few-shot metric learning, by constructing a class prototype in the form of a mean vector of the embedded support points within a class. The performance of prototypical networks in extreme few-shot scenarios (like one-shot) degrades drastically, mainly due to the desuetude of variations within the clusters while constructing prototypes. In this paper, we propose to replace the typical prototypical loss function with an Episodic Triplet Mining (ETM) technique. The conventional triplet selection leads to overfitting, because of all possible combinations being used during training. We incorporate episodic training for mining the semi hard positive and the semi hard negative triplets to overcome the overfitting. We also propose an adaptation to make use of unlabeled training samples for better modeling. Experimenting on two different audio processing tasks, namely speaker recognition and audio event detection; show improved performances and hence the efficacy of ETM over the prototypical loss function and other meta-learning frameworks. Further, we show improved performances when unlabeled training data are used.

Contrastive Learning of Cough Descriptors for Automatic COVID-19 Preliminary Diagnosis

“…other samples from the support set with the same or different label from that of the sample, respectively) in the latent space learned by the few-shot embedding model (output of f φ (.) [25]).…”

“…The embedding block (referred as f φ (.) in [25]) consists of a stack of 2 dense layers, with 800, 512 hidden units each with ReLU activation. During training, in each episode a triplet loss is computed using the output from the last dense layer as the latent representation.…”

“…We use number of shot as 5 (i.e. each training episode consists of 5 reference samples from each class) and margin = 0.5, where margin defines the maximum difference between the euclidean distances between anchor and positive sample, and between anchor and negative sample (as discussed in [25]).…”