Manifold

Weight separated machine learning framework.

Network types:

• manifold::nn::fc::Manifold Fully connected feedforward network.

Trainer types:

• manifold::optimizers::MiniBatchGradientDescent MBGD trainer with learning rate, decay, early stopping and more.
• manifold::neat::Neat Distributed async NEAT implementation (Neuro Evolution of Augmenting Topologies) using ZMQ workers.

Layer types:

• manifold::layers::Dense Dense (fully connected) layer.

Network types:

• manifold::nn::DNN Adjustable size dense network

Substrate types:

• manifold::Substrate Basic ringbuffer substrate using a Uniform distribution. No curvature.

TODO:

• make hyperparameters trainable via neat as well as network breadth and depth
• add self healing to neat async
• Add multi-machine distributed NEAT.
• Add curvature property to substrate, making it harder to reach edges.
• Add CNN