A bit puzzled about the results of using this FFNN about swift-ai HOT 4 CLOSED

Forget about it, just noob initialization of the weights.

from swift-ai.

Glad you got it figured out. Initializing the weights as nil will automatically assign random weights, although it currently only works well for Sigmoid activation.

It sounds like you're using the MNIST training set. I've had good results training FFNN with that set, and I'm preparing an example for it that will be finished soon. Check out the iOS example project - I update it daily and it includes some neat features. The handwriting recognition part is almost done.

from swift-ai.

Ah, that's kind of good to know (that means we're working on the same thing at the same time, so not so good :D ), interested to see what will come out! Will check the iOS example project in a few days

from swift-ai.

I'm only writing it for iOS right now. If you're writing an OS X version, that will be great to have. I'll move the MNIST data to the Examples directory so it's accessible for both apps (so we don't package two copies).

I'm adding Softmax activation to FFNN right now. It will be useful for handwriting recognition, since it helps with classification problems.

from swift-ai.