Problem 58961. Neural Net: Convolutional NN Back Propagation WP/WH/K LED Digit images; Variant Images Scored
This challenge is to return final WH, WP, and K matrices, given initial XImgset, WH, WP, K, EPY using convolution(Ximg,K), ReLU on the hidden layer and Softmax on the output layer. Training images are LED binary matrices of digits 1:9 and 0. Sets of variants will be tested to show success beyond the training set. (Scale, Offset, Noise, zeroed node(s), intra-shifts).
Back Propagation on WH, WP, and K. BP on K leads to much stronger solution.
Test sets 2 thru 5 show various Pixelations of numbers and the Neural Net success rate. Just fun to look at the data.
Appears more training sets required and possibly more nodes to have high success. One translate not good.
Solution CommentsShow comments
Problem Recent Solvers3