I am working on an open source GPU supported language for real neural network processing.
Called JOONN (June) it will be the first solution focused on neural modeling and large neural cubes.
It is a response to the statistical convolution network appraches which are think layer and matrix based. Not neural modeling based.
One issue is GPU support. I am so utterly frustrated with the crap of JOCL and JCUDA that finding the right library to bridge calls efficiently is difficult. ROCM is an interesting development that bridges all cpu types.
Existing software does not easily allow setting up complex neurons.
The modeling consists of creating a neuron class with
a connection profile
a propagation function
an activation function
a decay function
a chaos / edelman function for connectivity and spontaneous link generation
When you reach the wall with nothing existing thats worth a shit after a years investigation, you realize the need to create your own.
If you are a developer who wants to help work on this project shoot me a message or post a comment below. Esp if you are a GPU coder. thanks.
No comments:
Post a Comment