A Simplicial Piecewise Linear Approach for Efficient Hardware Realization of Neural Networks

N. RODRIGUEZ, P. JULIAN, E. PAOLINI

Baltimore

Congreso; 2019 53rd Annual Conference on Information Sciences and Systems (CISS); 2019

Johns Hopkins University

Neural Networks (NN) have been a matter of research because of their capability to solve complex problems where other topologies fail. However, current implementations of practical NNs require powerful computers for their deployment, like the ones used in data centers which typically require a high energy budget. This work proposes a simplicial piecewise linear algorithm as an alternative to implement NNs, which in addition can be implemented in low-power microelectronics. We illustrate the feasibility of the approach giving some examples where the simplicial algorithms replace conventional NNs with similar results.