Multi-Objective Fine-Grained Neural Architecture Search

The computational complexity of modern deep learning models is extremely high. For onboard robot calculators, their use is impossible without additional manipulations. It is promising to create a neural network for a specific task with appropriate requirements for the complexity of the model. Neural architectural search methods on popular datasets demonstrate promising results in creating small-sized architectures. In practice, not all of them can be used for their own purposes due to their high demands on computing resources at the search stage. In this paper, we propose an algorithm for reducing the computational complexity of the model based on solving a multi-objective optimization problem. The result of the proposed method is a set of trained networks with a different ratio of quality and computational complexity. The obtained results of practical experiments on own data confirm the effectiveness of the proposed method. With a slight loss in quality, it was possible to reduce the computational complexity of the network several times in comparison with the original architecture.