Reviews: Reinforced Continual Learning

This paper presents a simple but effective way to dynamically grow a neural network in a continual learning setup. The framework uses reinforcement learning to learn a policy that outputs decisions in terms of number of nodes/filters to add, and those additional filters are trained specifically for the new task (while the previous layers are held constant to prevent reduction in performance for old tasks). This is progressively done layer by layer where previous decisions for earlier layers serve as input to an LSTM in order to make decisions for the subsequent layer. Results show similar performance to approaches such as DEM, but have some better results for CIFAR-100 especially and the method results in networks with lower complexity. Strengths - Good overview of two groups of approaches on this topic - The idea of using RL for continual learning is a good one, and one that makes sense given recent advancements in RL for architecture search.