Use of programmable switching chips as artificial neural network modules

摘要

A method of performing a binarized neural network (BNN) using a switching chip involves describing an application of an artificial neural network in a binarized form to provide the BNN; configuring a parser of the switching chip to encode an input vector of the BNN into a packet header; configuring a plurality of match-action tables (MATs) of the switching chip to perform one or more of the operations including XNOR, bit count, and sign operations on the input vector encoded in the packet header such that the plurality of MATs are configured by: implementing a bitwise XNOR operation between the input vector and a weighting matrix to generate a plurality of first-level vectors, implementing an algorithm for counting a number of bits set to 1 in the plurality of first-level vectors to generate a plurality of second-stage vectors and to implement a sign operation on the second-stage vectors.