Expand description
Eight novel neuron models designed for AI workloads, not biological simulation.
Structs§
- Arcane
Neuron - ArcaneNeuron — unified self-referential cognition model.
- Attention
Gated Neuron - Spiking neuron with learned sigmoid attention gate. gate = sigmoid(w_key * I + w_query * v), modulates input before integration.
- Compositional
Binding Neuron - Phase-coding neuron for compositional variable binding. Spike when amplitude * cos(phase) > threshold.
- Continuous
Attractor Neuron - Ring attractor for continuous working memory. Mexican hat connectivity; holds a continuous value in persistent activity.
- Differentiable
Surrogate Neuron - Spiking neuron with learnable surrogate gradient parameters. alpha (decay), beta (steepness), theta (threshold) all trainable.
- Meta
Plastic Neuron - Neuron with self-regulating meta-learning rate. error_trace adapts learning speed: high error → learn faster, low error → stabilize.
- Multi
Timescale Neuron - Three-compartment memory neuron (fast/medium/slow timescales). Slow compartment accumulates context, modulating excitability.
- Predictive
Coding Neuron - Fires only on prediction errors. Silent when input matches prediction.
- Self
Referential Neuron - Introspects on its own spike history; adjusts tau based on firing rate.