Principle of least action in dynamically configured image analysis systems

We propose and justify the development of neural network architectures for training image analysis systems with a dynamically configurable calculation structure. The proposed systems, trained in accordance with the principle of least action, are useful for increasing the processing speed of large amounts of data and can help overcome other shortcomings of deep architectures. The problem of image classification is considered in detail, the solution of which reduces to training a network agent that operates in an environment of classifiers and indirectly perceives images through them. The proposed approach can be used to create algorithms in systems for automatic image analysis, where the critical characteristic is the average processing time of one frame, for example, in systems for image indexing based on their content.