03 — Capability
Differentiable Optimization
Backprop through time, surrogate gradients, online learning — built into the language of the model.
BrainX supports differentiable optimization techniques, enabling gradient-based learning and parameter tuning directly within biologically realistic neural simulations. The platform combines classical optimization methods with differentiable programming methods:
- Backpropagation through time (BPTT)
- Surrogate gradient methods
- Online gradient algorithms implemented in BrainTrace
From batch to online
Most ML libraries assume offline, batch-style optimization. Brain models trained during simulation need online learning — gradients computed and applied as the run proceeds. BrainTrace gives you that, while reusing the same JAX-native gradient infrastructure that the rest of the ecosystem depends on.