Multi-Agent Semi-Siamese Training for Long-tail and Shallow Face Learning

With the recent development of deep convolutional neural networks and large-scale datasets, face recognition has made remarkable progress been widely used in various applications. However, unlike existing public many real-world scenarios recognition, depth training dataset is shallow, which means that only two images are available for each ID. non-uniform increase samples, such issue converted to a more general case, known as long-tail learning, suffers from data imbalance intra-class diversity dearth simultaneously. These adverse conditions damage result decline model performance. Based on Semi-Siamese Training, we introduce an advanced solution, named Multi-Agent Training (MASST), address these problems. MASST includes probe network multiple gallery agents—the former aims encode features, latter constitutes stack prototypes (gallery features). For iteration, network, sequentially rotated stack, form pair networks. We give theoretical empirical analysis that, given (or shallow) loss, smooths loss landscape satisfies Lipschitz continuity with help agents updating queue. The proposed method out extra-dependency, thus can be easily integrated functions architectures. It worth noting although employed training, needed inference, without increasing inference cost. Extensive experiments comparisons demonstrate advantages shallow learning.