[1502.02551] Deep Learning with Limited Numerical Precision

http://arxiv.org/abs/1502.02551

Our results show that deep networks can be trained using only 16-bit wide fixed-point number representation when using stochastic rounding, and incur little to no degradation in the classification accuracy.

16ビットの浮動小数点数（半精度浮動小数点数）でも確率的に丸め処理をすれば使い物になるらしい。

脳に「インスパイア」されたチップがコンピューティングの未来を変える | ReadWrite Japan

http://readwrite.jp/archives/2389

脳と同じように機能するコンピューター « WIRED.jp

http://wired.jp/2013/09/02/truenorth/

IBMが人間の脳と同じ構造を持つプロセッサーの開発に成功 - GIGAZINE

http://gigazine.net/news/20140808-ibm-brain-similar-processor-chip/

IBM's TrueNorth processor mimics the human brain - CNET

http://www.cnet.com/news/ibms-truenorth-processor-mimics-the-human-brain/

Yann LeCun - My comments on the IBM TrueNorth neural net...

https://www.facebook.com/yann.lecun/posts/10152184295832143

Neuron states are binary (spikes) and synaptic weights are 9-bit signed integers with a 4-bit time delay. The overall peak performance is 4096x256x256x1000 = 266 GSops (billion synaptic operations per second).