Is it possible that a neural network on green streams?

Question

Erlang

I pee on Go, I heard about machine learning, I also want it, like everyone else. I found bindings for tensorflow, Python will be richer. But it is impossible without tensors (but which of the artificial intelligences today will multiply two tensors), but directly simulate neurons and connections with light flows and exchange channels? Are there any documented failures? Or even think to forget?

Answer 1

Neurons can be directly simulated, and even the math has already been written. But on modern hardware, it will be possible to launch no more than a million (plus/minus two orders of magnitude) neurons, and in the human brain about 100 billion. Tensorflow operates with matrices, and operations on them are well vectorized and have productive implementations not only for the central processor, but also for the GPU and clusters, as a result, it turns out to count a couple of orders of magnitude more neurons.

Answer 2

It can be done. A question of efficiency.
In one thread, you can process one neuron, but the whole network will be more efficient.
That. you come to the conclusion that you need to process several networks in parallel.
Now, taking into account the fact that you will process the network using TensorFlow, etc. we come to the conclusion that Erlang is needed only in order to be the glue between such parts of the program.