A browser-based platform for learning continuous-time recurrent neural networks from simulation and state-space dynamics to manual backpropagation through time.
The same CTRNN equation is used in both modules. The first module studies dynamics with fixed weights. The second changes the weights by learning from an error signal.
Choose inputs, initial states, recurrent weights, activation functions and time constants. Visualize neural activity, population trajectories, fixed points and local stability.
No learning: weights stay fixed unless you manually edit them.
Open Dynamics Module →Train a small CTRNN on a delayed-response task using a fully browser-based, manually implemented backpropagation-through-time algorithm.
No PyTorch and no Python: forward simulation, gradients and SGD are all written in JavaScript.
Open Training Module →