Has anyone used SDR to feed to NN.
I’m asking something along the lines of converting BoW or word-embedings or randomly-generated-SDR or word2vec passed trough Spatial pooler and then using them to train NN.
text => bow => SP =SDR=> NN => SDR-labels => text-label
It would have to perform better because the SP->SDP should be related i.e. the input to NN already have correlation.
As I’m on this : Is there a standalone implementation of Spatial Pooler ?
It sounds like a good idea! I’m in! Just to be clear, how are you planing to encode BoW into SDR?
In which language?
HTM.core have one in C++ and Python, Etaler have one in C++.
The MNIST example in htm.core uses SP outputs to train a softmax classifier which is basically a NN.
If I’m not mistaken, that’s pretty much what the SDRClassifier
does - taking a prediction (series of) SDR and running them
through a single-layer, feed-forward network with softmax
where can I find a description of the SP algorithm ?
Here, have a look at BAMI (Biological and Machine Intelligence). Page 43