search

CUFCTL / face-recognition

A GPU-accelerated real-time face recognition system based on classical machine learning algorithms
MIT License
23 stars 11 forks source link

Implement other non-linearity functions in ICA #32

Closed bentsherman closed 7 years ago

bentsherman commented 7 years ago

Not really high priority right now, but the FastICA paper suggests that the pow3 non-linearity is the fastest but also the least accurate non-linearity option, and it is the only one we have implemented. Now that our implementation of ICA is pretty stable, it might be a good project for someone to try and implement the other non-linearity functions to see if they are more accurate, and at what cost in performance.

Since inference time won't be affected by the nonlinearity, I think this is worth looking into eventually.

bentsherman commented 7 years ago

I have added the tanh and gauss functions, and ica_nonl is now a hyperparameter,