Machine learning is expanding its scope from mere machine intelligence to full human-like interactions. Similar to humans, machines are now being made with the ability to learn and perform cognitive action or skills based on interactions with the environment. This paper introduces a natural learning theory.

The authors introduce a true learning environment, describing the function of the agent to discard irregularities in the given data. They also mention Euler-Lagrange laws of learning where such equations represent static points and the possibility of non-minimization of cognitive action. They further reduce laws of learning to classic mechanics and find coherence between the definitions of action and potential function.

The authors formally define learning using the Legendre transform in terms of Hamiltonian formulation of the initial conditions of the problem. They further describe BG brackets properties reducing to classic Poisson’s brackets, formulating detailed understanding of energy exchange processes. In an interesting way, the authors describe the connection between energy balancing and cognitive action determining the learning dynamics. The authors emphasize the resemblance of the general form of the cognitive action to the nonlinear form of the recurrent neural networks.

The authors are working with the implementation of their results in semantic learning in particular and in computer vision in general. This paper is an interesting read for those who are working with the development of computational models in the neuroscience and cognitive psychology areas.