SDR and sparse networks visualizations

so I’m thinking.

for as much as I like lists if integers and plotting squares with pixels for activated neurons and tangles of random lines for synapses. I wonder if there are other, more human-readable methods to visualize SDR’s data and network weights.

3 Likes

Maybe you could try using a k-nearest-neighbors search to find similar SDR’s?
And then you can show the inputs that yielded the similar SDRs?

Then you would be looking at the input data instead of the raw SDRs.

2 Likes

Felix Andrews @floybix had quite a nice display for his Clojure port of NuPIC 2014-2015 or so. As I recall he had a display of cells against time, inputs and outputs, and you could click on a cell to see its synapses. This Fall 2014 Hackathon video looks to be an early version:

HTM in Clojure [DEMO #6] (2014 Fall NuPIC Hackathon)

HTM higher-order sequence learning - SDRs

http://viewer.gorilla-repl.org/view.html?source=gist&id=95da4401dc7293e02df3&filename=seq-replay.clj

A more comprehensive list of docs, and I guess, still, the source for all his code, should be on his Github:

http://floybix.github.io

I see there’s even a document there on “Visualization driven development”, addressing the importance he gave to the network display:

http://floybix.github.io/2014/07/11/visualization-driven-development-of-the-cortical-learning-algorithm

1 Like

well, those are still kinda unsatisfying visualizations. I mean you can se motion and how SDR’s change but you dont get any more information than that, its like staring at the green screen of random characters from matrix (the movie) and pretending you understand it.

1 Like