A Stochastic Thermodynamics-based Network Architecture (ThN) for Machine Learning

We have developed a new network architecture based on the principles of stochastic thermodynamics for machine learning.  The architecture has been developed based on the principles of thermodynamics in which energy and entropy represented by the attributes.  As these are both extrinsic properties, they are essentially determined by the specific problem for which machine learning is applied to.  The challenges in the formulation have been in the representation of the intrinsic complexity and the presence of disparate length and time scales.  However, given the combinatorial nature of many problems, this architecture has a unique advantage of being able to evolve dynamically in non-equilibrium dynamic mode and also learn in equilibrium conditions.  The dynamics of the network and its stochastic nature provides additional advantages for application of this formalism to both chemical and biological systems.  We will also be presenting the comparisons of the Thermodynamic network (ThN) with more conventional Deep neural Networks.