Sparsely ensembled convolutional neural network classifiers via reinforcement learning
This work provides a method for more resource-efficient deep learning inference, which is beneficial for applications with computational constraints.
This paper addresses the problem of efficiently ensembling convolutional neural networks by dynamically selecting classifiers to minimize computational resources while maximizing accuracy. The proposed reinforcement learning agent implicitly learns a classifier selection function, outperforming conventional ensemble learning methods.
We consider convolutional neural network (CNN) ensemble learning with the objective function inspired by least action principle; it includes resource consumption component. We teach an agent to perceive images through the set of pre-trained classifiers and want the resulting dynamically configured system to unfold the computational graph with the trajectory that refers to the minimal number of operations and maximal expected accuracy. The proposed agent's architecture implicitly approximates the required classifier selection function with the help of reinforcement learning. Our experimental results prove, that if the agent exploits the dynamic (and context-dependent) structure of computations, it outperforms conventional ensemble learning.