Hi everyone,
I use NUPIC spatial pooler for image recognition.
Here's my code test on MNIST: got about 92% acc.
https://github.com/tsenggordon/nupic.recogn

I want to extend the code to CIFAR 10 dataset.
I filter the input by max-pooling the pixel value.
Ex: [255 123 24 22] -> [1 0 0 0]
The code is similar to MNIST one
When I am training the net, I got 99% acc.
However, while I test the net using test set, I got 16% acc. (very poor
result)

Any comments are appreciated.

Thanks,

Gordon

Reply via email to