Expediting discovery in Neural Architecture Search by Combining Learning with Planning

摘要

In our previous work, we introduced NASIL as an automated neural architecture search method with imitation learning. Time to discover optimal structures is a key concern in many AML solutions including NASIL. Here, we proposed an extended version called "GNASIL" to speed up the process. Similar to NASIL, GNASIL takes advantage of imitation learning to discover neural architectures for a given device specification. Unlike NASIL that used deep deterministic policy gradient method, GNASIL uses the soft-actor-critic to predict an optimal layer during its search. Furthermore, GNASIL employs a set of probing options and combines learning and planning options to sweep the search space faster. We investigated impact of such deliberative planning on decision making process on a speech recognition task. Reported results demonstrate that probing options in presence of imitation learning enables GNASIL algorithm to automatically learn suitable network structures with very competitive performance both in terms of speed of finding the optimal architectures and their accuracy while keeping computational footprint restrictions into consideration.