SUPERPIXEL METHODS FOR CONVOLUTIONARY NEURAL NETWORKS

摘要

These are methods, systems, and apparatus for efficiently computing a convolutional neural network layer. Does one method include transforming an input tensor x by y by z into an input tensor x? why y? why z ?,where x? ey? are less than or equal to axey,respectively, ez? is greater than or equal to az; get modified weighting matrices that operate on input tensor x? why y? why z? to generate an output tensor u? why you? why w? which includes an output tensor u by v per w transformed,where u? and v? are less than or equal to auev,respectively, ew? is greater than or equal to aw; and process input tensor x? why y? why z? using modified weighting matrices to generate output tensor u? why you? why w ?.