Skip to content

Layer-by-Layer Guide: Building Your Architecture

Version: 3.15.35 Target: Detailed Component Reference

1. Neuron Types

The sc_neurocore.neurons module provides several neuron models tailored for different layers of the SCPN.

Stochastic LIF (stochastic_lif.py)

The "standard" model for L2-L4 simulations. It combines leaky integration with stochastic noise. * Key Params: tau_mem, v_threshold, noise_std, v_rest. * Use Case: General purpose biological modeling.

Izhikevich (sc_izhikevich.py)

A computationally efficient model that can reproduce a wide variety of firing patterns (regular spiking, bursting, chattering). * Key Params: a, b, c, d. * Use Case: When specific firing temporal dynamics are required.

Homeostatic LIF (homeostatic_lif.py)

Features a dynamic threshold that increases after spiking and decays over time (refractory period). * Use Case: Preventing "runaway" activity in deep recurrent networks.

Fixed-Point LIF (fixed_point_lif.py)

Optimized for hardware deployment (PYNQ-Z2), using integer arithmetic instead of floats.

2. Layer Architectures

Layers in sc_neurocore are containers for ensembles of neurons, providing efficient vectorized processing.

SC Dense Layer (sc_dense_layer.py)

A fully connected layer where every input is connected to every neuron. * Weights: $W \in \mathbb{R}^{N \times M}$ * Connectivity: All-to-all.

SC Conv Layer (sc_conv_layer.py)

Implements spiking convolution, ideal for processing spatial data (L4 Morphology). * Mechanism: Sliding kernels apply spatial filters to the input spike stream.

Memristive Layer (memristive.py)

Simulates non-volatile memory behavior in synapses (L9 Identity). * Feature: Weights persist even if the simulation is reset, modeling long-term memory.

Fusion Layer (fusion.py)

A specialised layer designed to integrate same-width modality vectors with normalised stochastic-computing weights. Positive raw weights are normalised to sum to one; zero or negative total weights fall back to equal weights across the weighted modalities, matching the Rust FusionLayer fallback. Inputs must be one-dimensional arrays with their declared feature length, and unweighted modalities are ignored rather than mixed into the output. * Use Case: Implementing the "Consilium" (L15) optimisation logic. * Verification: tests/test_layers/test_fusion.py covers normalisation, equal-weight fallback, validation failures, output dtype/shape, and Rust weighted-sum parity semantics. Local pytest-cov reports 100% line coverage for src/sc_neurocore/layers/fusion.py. The opt-in non-isolated perf smoke SC_NEUROCORE_PERF=1 ... test_fusion_perf_small passed on 2026-06-27 (1 passed in 1.92s); no formal isolated benchmark artefact exists for this layer yet.

3. Synaptic Components

Located in sc_neurocore.synapses, these define how signals travel between layers.

  • Static Synapse: Fixed weights.
  • STDP Synapse: Weights evolve based on spike timing.
  • Quantum Synapse: Probabilistic transmission based on L1 states.

4. Connectivity Patterns

Feed-Forward

Standard L1 -> L2 -> L3 flow. Used for feature extraction.

Recurrent

Connections within a layer (recurrent.py). Necessary for temporal memory and rhythm generation (L8).

Lateral Inhibition

Neurons within a layer inhibit their neighbors. This "Winner-Take-All" dynamic is used for pattern recognition and focus.

5. Summary Table

Component Module SCPN Mapping Complexity
LIF Neuron neurons.stochastic_lif L2 (Neural) Low
Izhikevich neurons.sc_izhikevich L2 (Dynamics) Medium
Dense Layer layers.sc_dense_layer L3 (Cognition) Medium
Fusion Layer layers.fusion L15 (Consilium) High
STDP learning.stdp L3 (Learning) Medium
Director meta.director L16 (Oversight) Extreme