2 ms·
I have zero doubt that given sufficient resources ANY function can be learned by a neural network. Why not just initialize part of the layers in a network to d
by mikewarot 6y ago
I have zero doubt that given sufficient resources ANY function can be learned by a neural network.
Why not just initialize part of the layers in a network to do the FFT, or DFT when you're setting up a system to learn, that it doesn't have to waste megawatt hours of power? Consider it a form of discretized / explicit transfer learning.
Why not use the fact that you could just set the coefficients in a neural network to do DFTs or FFTs to utilize neural network chips as ersatz DSP chips, without learning?
Just like when GPUs turned out to be useful for neural networks, the TPUs will turn out to be useful for DSP and other things. These neural network chips showing up can do a lot of cool things unrelated to AI, if we use them properly.