NEAT Algorithm implementation

[crisbodnar]

Would it be possible to write an algorithm like NEAT which evolves the topology of a neural network in Fold? From my understanding, Fold does not support dynamic computation graphs in this sense, but these dynamic graphs are rather input dependent.

[delesley]

The original NEAT algorithm evolved small neural networks at the level of individual neurons. That obviously won't work with tensorflow; calling a python function to schedule a tensorflow operation to invoke a cuda kernel to multiply two scalars would have ridiculously high overhead. However, if you want to implement NEAT at the level of NN layers, rather than individual neurons, then Tensorflow Fold should work quite well. You'll want to use the loom library, not the blocks library. First, create a separate LoomOp class for every NN operation you want to support. Second, implement NEAT on a population of programs. Each program in the population is a DAG of NN operations. Write a recursive python function which traverses the DAG for every program in the population, and invokes the appropriate LoomOp for each node. There is a calculator example in loom that shows how to do this for arithmetic expressions. Loom will handle the dynamic batching for you, and evaluate the LoomOps for all programs using TensorFlow. It might be easier to build a prototype using TensorFlow Eager, and then switch to loom. Loom should give you a nice ~30x speedup over eager due to dynamic batching. -DeLesley

…

On Mon, Oct 29, 2018 at 10:22 AM Cristian Bodnar ***@***.***> wrote: Would it be possible to write an algorithm like NEAT which evolves the topology of a neural network in Fold? From my understanding, Fold does not support dynamic computation graphs in this sense, but these dynamic graphs are rather input dependent. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#100>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AGGbTeryJ8zn8Q4ol6FgWrioSsqJavvrks5upzlwgaJpZM4X_rAF> .

-- DeLesley Hutchins | Software Engineer | delesley@google.com | 505-206-0315