I had some fun putting J Playground through its paces and wrote the following neural network demo:
https://jsoftware.github.io/j-playground/bin/html2#url=https://gist.githubusercontent.com/jpjacobs/6ff68c07bf764ad886e097ffbd19f5f0/raw/52b0563a7daa76df7fa57a3743da45ca0755f1a7/nn.ijs It dowloads the CSV and DSV packages and the Spirals dataset, normalises the data and applies some random jitter. The neural network can be given any number of layers of any size (of course, respecting the input layer to be of size 2 and the output layer of size 3,as there are 2 input dimensions and 3 labels, one-hot encoded), e.g. net=: nn 2 15 8 3 It is trained with stochastic gradient descent and backpropagation with quadratic loss, using e.g. 5000 100 0.1 sgd__net spl =: splitdata data ,. labels When finished, it outputs some stats and some nice pictures of the data and classification borders, using report__net spl. I was **very** impressed to see it run (entire script with 5000 training epochs, and visualisation) without going out of memory in under 35 secs on my mid-range phone… just wow. Some things I noted with respect to the playground: - I had to patch the enclength_jzlib_ verb to allow bigger viewmat pictures (see at the bottom of the script). - It's a pity the text output is blocked until all calculations are done, as this renders it impossible to give status updates. Is there a trick like in JQt (wd 'msgs') to overcome this? - Maybe related: plots and text output do not keep their relative output positions, which I think would be desirable when trying to make a coherent example. - The plot quality is not great yet, while I saw j901 on iPadOS seems to get vector graphics from plot. As the Playground would be the showcase of J, it would be great if it could also include crisp graphics. - Lastly, I noted some strange scrolling glitches after running the script in chrome on Windows (didn't try elsewhere): when scrolling up the terminal output would jump back down from time to time. Any suggestions for improvement are of course welcome, though I kept the features minimal to avoid making it too lengthy. To my taste, the report function is for instance a bit too wordy, but necessary to show some interesting graphs. If one were really ambitious, it could be turned into something like: http://playground.tensorflow.org/ ( open source, by the way). I hope you have fun tinkering with the neural network! Best regards Jan-Pieter ---------------------------------------------------------------------- For information about J forums see http://www.jsoftware.com/forums.htm
