search

AntonioTepsich / Convolutional-KANs

This project extends the idea of the innovative architecture of Kolmogorov-Arnold Networks (KAN) to the Convolutional Layers, changing the classic linear transformation of the convolution to learnable non linear activations in each pixel.
MIT License
779 stars 76 forks source link

How to set the bias #5

Closed entuo closed 5 months ago

entuo commented 6 months ago

Congratulations, this is groundbreaking work. I was wondering if it could be used as a replacement for nn.conv2d and I realized that there is no bias attribute in it, how should I set it up, thanks for your answer!

AlexBodner commented 5 months ago

Hi! In a KAN that uses B Splines you dont need a bias! It is different than a normal network, as you learn an activation function that includes in its representation a potential bias that it might need.