Home

BUGS / TODO

• there is now support for neural network reinforcement learning (RL) to try to learn in roguelike environments (calls python/minihack package). Improve RL algorithms so that it can learn to play in minihack environments.

• add more support for drop-out code (in RL and L-BFGS and other places)

• test if superresolution code works with a simple regularization (0.5*||w||^2 term) and don't diverge to large numbers.

• RBM code seem to be broken but runs/computes for now.

Deep learning: for a simple test problem (test_data_residual.sh) neural network can learn the problem with 40 layers in 10 minutes (dense residual neural network with leaky rectifier unit non-linearity), 20 layers residual neural network gives perfect results.