Brain Sparks

Experimental neural network and deep learning code

Calrissian

The main neural network code. Makes heavy use of NumPy.

Implemented and working

• Multi-layer perceptrons (dense layers)
• Backpropagation gradients
• Common activation functions: sigmoid, tanh, ReLU, linear
• Common cost/error functions: MSE, MAE, cross-entropy
• Stochastic gradient descent optimizer
• L1 and L2 regularizers

In the works

• 1-D convolution layer, optional max pooling (gradient nearly done)
• Autoencoders (MLP style)
• GPU acceleration via PyCUDA and/or PyOpenCL (experimenting still)

Future work

• More optimizer options: rprop, momentum, adagrad, etc.
• 2-D convolution, to follow up on 1-D convolution
• Recurrent networks

BrainSparks

Experimental data parallelism hook for Calrissian.

• Attempted some linear algebra with Spark, but not much luck.
• Looking into replicated Calrissian networks for parallelism