Can't train MNIST digits to Perceptron about neataptic HOT 2 CLOSED

Dope! I'm excited to dig into the code and make some more tests with second order neural nets (like the discrete sequence recall), but that was really fast 👏 !

from neataptic.

Hi. I just fixed the data bug. And while testing the MNIST dataset I realised that weights were initialised way too high for perceptrons. I fixed the JSFiddle (unfortunately, Neataptic has no async training). And I changed yours to output the time (and made it not async):

Open console!

• Neatatpic: https://jsfiddle.net/pn5zerq9/3/
• Synaptic: https://jsfiddle.net/672c4j1g/2/

Results for my PC (chrome + windows 10.0.0):

• Neataptic: time = ~3100ms, error = ~0.008
• Synaptic: time= ~29500ms, error = ~0.008

I just did a few runs. Let me know if your results are different.

I also changed the readme to claim more than 5x as smaller datasets run slower after I removed the weight initialisation. I will soon include weight initialisation methods.

from neataptic.