The adaptive image classification method using reinforcement learning

The paper proposes a method for image classification that uses in addition to a basic neural network for image classification an additional neural network able to adaptively concentrate on the classified image object. The task of the additional network is the contextual multi-armed bandit problem, which re-duces to predicting such area on the original image, which is, when cut out of the classification process, will increase the confidence of the basic neural network that the object on the image belongs to the correct class. The additional network is trained using reinforcement learning techniques and strategies for compromising between exploration and research when choosing actions to solve the contextual multi-armed bandit problem. Various experiments were carried out on a subset of the ImageNet-1K dataset to choose a neural network architecture, a reinforcement learning algorithm and a learning exploration strategy. We con-sidered reinforcement learning algorithms such as DQN, REINFORCE and A2C and learning exploration strategies such as -greedy, -softmax, -decay-softmax and UCB1 method. Much attention was paid to the description of the experiments performed and the substantiation of the results obtained. The paper proposes application variants of the developed method, which demonstrate an increase in the accuracy of image classification in comparison with the basic ResNet model. It additionally considers the issue of the computational complexity of the developed method.