Dynamical System Viewpoint of Machine Learning
Characterization of the ML models' learning behavior with input from data
Once the model is fixed and appropriate learning mechanism is chosen, the typical way of training a model is to consume data from an application and update the model. In order to characterize the behavior of learning a ML model over a series of data samples, one must characterize the impact of each sample on the learning. However, due to the compositional structures prevalent in Neural networks does allow of trivial study of this behavior. Furthermore, to study the existence of solutions for ML model in a variety of learning scenarios such as supervised, unsupervised and reinforcement learning scenarios requires additional tools that are not avaiable in the ML literature.