Improving the Efficient Neural Architecture Search via Rewarding Modifications

Nowadays, a challenge for the scientific community concerning deep learning is to design architectural models to obtain the best performance on specific data sets. Building effective models is not a trivial task and it can be very time-consuming if done manually. Neural Architecture Search (NAS) has achieved remarkable results in deep learning applications in the past few years. It involves training a recurrent neural network (RNN) controller using Reinforcement Learning (RL) to automatically generate architectures. Efficient Neural Architecture Search (ENAS) was created to address the prohibitively expensive computational complexity of NAS using weight sharing. In this paper we propose Improved-ENAS (I-ENAS), a further improvement of ENAS that augments the reinforcement learning training method by modifying the reward of each tested architecture according to the results obtained in previously tested architectures. We have conducted many experiments on different public domain datasets and demonstrated that I-ENAS, in the worst-case reaches the performance of ENAS, but in many other cases it overcomes ENAS in terms of convergence time needed to achieve better accuracies.