Comment by Aron

I remember writing the RBM in straight CUDA back in ~2008. Back then we had to do our tensor work uphill both ways against the gradient.