Simulating the flip-flop task using Echo State Networks and FORCE learning
The project has two parts. In the first part, a standard echo state network (ESN) model is implemented with feedback of the output. The read-out weights are optimized using regularized least-square method. The simulations are performed on flip-flop task with 2, 3, and 4 input/output bits, and the results are analyzed by varying hyperparameters and other settings.
In the second part, the online training algorithm known as FORCE learning is implemented. The read-out weights are optimized using the same method as part 1. The simulations are performed on the same flip-flop tasks as in part 1, and conclusions are drawn on the differences between the two approaches.
- Ceni et al. Interpreting Recurrent Neural Networks Behaviour via Excitable Network Attractors. Cognitive Computation, 2019
- Section 3.3 of Sussillo and Barak. Opening the Black Box: Low-Dimensional Dynamics in High-Dimensional Recurrent Neural Networks. Neural Computation, 2013
- Soheil Changizi ( @cocolico14 )