Some possible project topics

The following are a few suggestions for project topics. They are all presented in a fiarly open-ended fashion here. Potential projects will need to be decided in detail depending on your interests and the size of the planned project. In each case, the final goals and requirements will have to be decided upon together with Professor Colombetti and us. You are also free to suggest topics of your own.

Please bear in mind that Nic and Fred are both normally in Lugano. They are, of course, contactable by email.

• Implement Real Time Recurrent Learning (Neural Computation, 1, 270-280, 1989). Code up your own implementation of RTRL. Reproduce the results on temporal XOR and sine wave oscillation. Examine the effects of continuous and a discrete periodic inputs on a network trained to oscillate. In what way does the network entrain to an external signal?

• 8-3-8 Encoder Implement a feedforward network with one hidden layer and batch backpropagation, and either momentum or the bold driver method. Set it up as an autoencoder with 8 inputs/outputs and 3 hidden units, and train it on the 8 binary patterns that consist of a single '1' and 7 zeroes. Find values for the free parameters that give you fast, reliable convergence, then compare the speed of learning and ultimate performance for the following cases:
  • linear outputs, sum-squared error
  • logistic outputs, sum-squared error
  • logistic outputs, cross-entropy error
  • softmax outputs, cross-entropy error

• Online Learning Write a program that generates training data for a neural network, such that the function the network must learn to approximate changes periodically. Then implement a neural network that obtains its training patterns from this generator, and performs online learning with local learning rate adaptation on it. Compare the network's performance for various values of the meta-learning rate.

• Tic-Tac-Toe Implement a network that uses Q-learning to learn the game of tic-tac-toe (see figure) from self-play.
  Tic-tac-toe

• Applications You may want to consider applying a neural network as part of a project that relates to other part of your course - for example, in building or simulating a reactive agent.

• Your suggestion here. Have you a favourite dataset you wish to model? Time series data to predict? Pattern recognition problem? Let us know.

[Top] [Back to the first page]