Design of Hardware Accelerators for Fractal based Machine Learning Applications

Natural processes are non-linear dynamical systems. Samples extracted from such processes are often estimated through non-Euclidean geometry like fractals. Provided such a structure exists in a given data distribution, it becomes imperative to use non-Euclidean space like graphs and manifolds to represent it. Traditional Euclidean machine learning or deep learning techniques may not be sufficient for such spaces. This work provides some insights on how learning or inference for such classes of problems may be accomplished by following some novel hardware software co-design techniques. This work proposes some novel techniques to design hardware accelerators for realizing fractal based machine learning applications.