Deep Learning Networks on the GPU.

Thanks to Tim Kay (Systems Guru at EECS for getting webots/etc.. working on the robot).

Now I'm going to try to get deep learning working on the robot.

1. Theano is installed 

[http://deeplearning.net/software/theano/]

2. Look at the deep learning tutorials below.. 

[http://deeplearning.net/tutorial/]

Siddharth Sigtia is kindly helping me to apply DBNs to self-organization of sensorimotor representations in the NAO. He's got the above stuff working on the server now. 

Papers on DBNs for Reinforcement Leanring/Function Approximation in Robotics 

1. http://www.frontiersin.org/10.3389/conf.fncom.2011.52.00029/event_abstract

IDEAS

1. Use DBN to learn 

1. Forward models: That is, it should learn to predict the next sensory state given the current sensory state and the motor command.

2. Inverse models: Predict the motor commend required to achieve a next sensory state.

2. Use DBN for tactile object recognition. 

NAO does motor babbling and is allowed to interact with various objects. The higher levels of the DBN might be able to categorise the objects or motor scenarios into various types, and thus act as context nodes.