Entropy-based Sampling for Streaming learning with Move-to-Data approach on Video

Abstract

The current paradigm of training deep neural networks relies on large, annotated and representative datasets. They assume a static world where the target domain does not change. However, in the real-world, data changes over time and is often available on the fly. Naive retraining on new data causes catastrophic forgetting and the network is unable to generalize on old data. Streaming learning is a type of incremental learning where networks learn sequentially and as soon as a sample is available from the data stream. Instead of training on every new sample, we propose an uncertainty based selection criteria to improve our previously proposed fast streaming learning method Move-to-Data (MTD), called Entropy-based MTD (EMTD). Besides, streaming learning methods have so far mostly used Convolutional Neural Networks (CNNs) but in recent times Vision Transformers (ViTs) have shown much better performances for many vision tasks. Therefore, we use ViT based Video Transformer to analyse MTD, EMTD and their gradient descent based "retargeting" steps. We have compared the performances of EMTD with MTD (w/wo retargeting) and a popular streaming learning method ExStream for the transformer. EMTD is able to outperform baseline MTD, and EMTD with retargeting achieves close results as ExStream and is ∼ 1.2 times faster.