robertleeplummerjr
commented
6 years ago The short answer: yes The longer answer: This is something we've put a great deal of thinking into, and ended up adopting gpu.js and assisting getting it to version 1. I just merged in an experimental branch this is where the above answer "yes" is satisfied. That gives us our short term gain for feed forward neural networks.
The long term "yes" is that we are (re-?)structuring the recurrent architecture (which is still experiment) to create a single unified way of composing layers. This branch: https://github.com/BrainJS/brain.js/tree/nn-gpu-layers embodies where we are going next. Not only will it be fast (gpu), but easy, and straightforward. This is something we feel is greatly missing from the machine learning community. The api is simple:
const net = new FeedForward({
inputLayer: () => { /* return a layer */ }
hiddenLayers: [() => { /* return a layer */ },() => { /* return a layer */ },() => { /* return a layer */ }],
outputLayer: () => { /* return a layer */ }
});What is most important about this is that it will run on GPU first, with a CPU fallback, and that the Recurrent neural network will share the same basic architectural frontend, so no matter if you are building a simple perceptron, or a Vincent Van Gogh imagine manipulator, or a Johnny5 robot, or a cancer detector, you'll have the same easy means of composition.
I could go on and on and on, this has been far over a year in the making representing a lot of very smart people asking if the way current neural networks have been implemented right, and the answer we keep coming back to is "no", but I digress, and I thank you for asking this very fundamental question.
nsantini
arguiot
Hello 👋, I would like to know if it would be possible to improve performances of the
trainfunction, using the GPU, or using multiple cores/threads in the node version.Thank you for building this library