[IanCal]

This isn't too related but

> Deep learning took off precisely when the ImageNet paper dropped around 2010. Before nobody believed that backprop can be GPU-accelerated.

Deep learning kicked off with RBMs because you didn't have to do backprop and there was a training algorithm called "contrastive divergence". Each layer could be done in turn, which meant you could stack them way deeper. In ~2008-2009 I implemented Hintons paper on GPUs, which meant I could do the same scale of thing that was taking weeks in matlab in hours on a gpu, and then there were lots of gpus available on the cluster in the uni. Lots of fun cuda hacking (I'm just glad cublas was around by then). The original published learning rates/etc are wrong if I remember right, they didn't match the code.