neural circuit policies

**Neural Circuit Policies (NCPs)** are **sparse, interpretable recurrent neural network architectures** — derived from Liquid Time-Constant (LTC) networks and wired to resemble biological neural circuits (sensory -> interneuron -> command -> motor). **What Is an NCP?** - **Structure**: A 4-layer architecture inspired by the C. elegans nematode wiring diagram. - **Sparsity**: Extremely sparse connections. A typical NCP might solve a complex driving task with only 19 neurons and 75 synapses. - **Training**: Trained via algorithms like BPTT or evolution, then often mapped to ODE solvers. **Why NCPs Matter** - **Interpretability**: You can look at the weights and say "This neuron activates when the car sees the road edge." - **Efficiency**: Can run on extremely constrained hardware (IoT, microcontrollers). - **Generalization**: The imposed structure prevents overfitting, leading to better out-of-distribution performance. **Neural Circuit Policies** are **glass-box AI** — proving that we don't need millions of neurons to solve control tasks if we wire the few we have correctly.