Improving Spatiotemporal Self-Supervision by Deep Reinforcement Learning
This work addresses a bottleneck in self-supervised learning for computer vision, offering an incremental improvement in sampling efficiency.
The paper tackled the problem of inefficient permutation sampling in spatiotemporal self-supervised learning by introducing a deep reinforcement learning policy that adapts sampling to the network's state, resulting in competitive performance on image and video classification and retrieval benchmarks.
Self-supervised learning of convolutional neural networks can harness large amounts of cheap unlabeled data to train powerful feature representations. As surrogate task, we jointly address ordering of visual data in the spatial and temporal domain. The permutations of training samples, which are at the core of self-supervision by ordering, have so far been sampled randomly from a fixed preselected set. Based on deep reinforcement learning we propose a sampling policy that adapts to the state of the network, which is being trained. Therefore, new permutations are sampled according to their expected utility for updating the convolutional feature representation. Experimental evaluation on unsupervised and transfer learning tasks demonstrates competitive performance on standard benchmarks for image and video classification and nearest neighbor retrieval.