4 ms·

I wonder if one could store only the binary representation at training and sample a floating point representation (both weights and gradient) during backprop.

• adastra22 1y ago
  Back propagation on random data that is then thrown away would be pretty useless.