Reinforcement Learning for Architecture Search by Network Transformation

Deep neural networks have shown effectiveness in many challenging tasks and proved their strong capability in automatically learning good feature representation from raw input. Nonetheless, designing their architectures still requires much human effort. Techniques for automatically designing neural network architectures such as reinforcement learning based approaches recently show promising results in benchmarks. However, these methods still train each network from scratch during exploring the architecture space, which results in extremely high computational cost. In this paper, we propose a novel reinforcement learning framework for automatic architecture designing, where the action is to grow the network depth or layer width based on the current network architecture with function preserved. As such, the previously validated networks can be reused for further exploration, thus saves a large amount of computational cost. The experiments on image benchmark datasets have demonstrated the efficiency and effectiveness of our proposed solution compared to existing automatic architecture designing methods.