Layers

Pre-built layer compositions combining neurons, synapses, encoders, and recorders into reusable building blocks.

Class	Architecture	Backend
SCDenseLayer	Fully-connected LIF	NumPy (loop-based)
VectorizedSCLayer	Fully-connected, packed bitwise	NumPy / CuPy GPU
SCConv2DLayer	2D convolution	NumPy
SCRecurrentLayer	Echo state / reservoir	NumPy
SCLearningLayer	Dense + online STDP	NumPy
SCFusionLayer	Multi-modal MUX fusion	NumPy
StochasticAttention	SC attention mechanism	NumPy
MemristiveDenseLayer	Memristive device model	NumPy
JaxSCDenseLayer	Fully-connected LIF	JAX (JIT, GPU/TPU)
HardwareAwareSCLayer	Dense + memristive defects	NumPy
PredictiveCodingSCLayer	XOR error, zero-multiplication	NumPy
RallDendrite	Compartmental dendritic tree	NumPy
LateralInhibition	Gaussian surround suppression	NumPy
WinnerTakeAll	k-WTA competitive layer	NumPy

Dense Layer

SCDenseLayer accepts either a 1-D weight_values vector, which creates one shared stochastic-computing current source for all neurons, or a 2-D (n_neurons, n_inputs) weight matrix, which creates one SC dot-product source per output neuron.

::: sc_neurocore.layers.sc_dense_layer.SCDenseLayer

Vectorized Layer

::: sc_neurocore.layers.vectorized_layer.VectorizedSCLayer

Convolutional Layer

SCConv2DLayer applies deterministic NumPy accumulation over stochastic probability encodings. Unipolar mode accepts finite channel-first tensors in [0, 1]; bipolar mode accepts finite tensors in [-1, 1]. Construction and runtime validation fail closed for invalid channel counts, kernel geometry, stride, padding, bitstream length, input rank, input domain, and empty output geometry before the im2col convolution path runs.

::: sc_neurocore.layers.sc_conv_layer.SCConv2DLayer

Recurrent / Reservoir Layer

::: sc_neurocore.layers.recurrent.SCRecurrentLayer

Learning Layer

::: sc_neurocore.layers.sc_learning_layer.SCLearningLayer

Fusion Layer

::: sc_neurocore.layers.fusion.SCFusionLayer

Attention Layer

StochasticAttention exposes three attention paths:

• forward(...) uses SC-native row-sum normalisation without exponentials;
• forward_softmax(...) uses temperature-scaled softmax with stable max-subtraction;
• forward_bitstream(...) encodes Q, K, and V as unipolar probabilities, computes dot products with bitstream AND gates, and decodes the row-normalised result by popcount.

Bitstream mode accepts Bernoulli streams by default. Passing use_sobol=True uses Sobol low-discrepancy streams for lower deterministic variance at the same bitstream length. All bitstream inputs must be finite probabilities in [0, 1], and length must be positive.

The public contract is tested as follows:

• invalid dim_k, temperature, and unsupported sc_mode fail closed;
• 1-D and 2-D Q, K, and V inputs produce finite outputs with the expected (n_queries, value_dim) shape;
• single-key attention returns the value row for both row-sum and softmax modes;
• softmax remains finite on large scores and sharp temperatures select the best-matching key;
• bitstream mode rejects invalid lengths and out-of-range probabilities;
• Bernoulli bitstream attention approximates row-sum attention within its stochastic tolerance;
• Sobol bitstream attention returns finite bounded probabilities.

::: sc_neurocore.layers.attention.StochasticAttention

Memristive Layer

::: sc_neurocore.layers.memristive.MemristiveDenseLayer

JAX Dense Layer

JaxSCDenseLayer accepts dense inputs shaped (n_inputs,) or (T, n_inputs) and projects them through a validated (n_neurons, n_inputs) weight matrix. It also accepts direct current vectors shaped (n_neurons,) or (T, n_neurons) for low-level LIF experiments. Constructor and runtime inputs validate dimensions, finite values, dense weight shape, known neuron parameter keys, and JAX PRNG seed range before backend execution.

The class is exported through the lazy package facades as sc_neurocore.JaxSCDenseLayer and sc_neurocore.layers.JaxSCDenseLayer. Importing the symbol does not construct backend state; constructing the class still requires the jax optional dependency group.

::: sc_neurocore.layers.jax_dense_layer.JaxSCDenseLayer

Hardware-Aware SC Layer

HardwareAwareSCLayer wraps VectorizedSCLayer with deterministic memristive-defect injection. A seeded stuck_mask selects defective synapses, each defective synapse is forced to a sampled stuck-at value (0.0 or 1.0), and update_weights(...) masks gradients at those locations so training can only adapt the remaining synapses.

The public contract is tested as follows:

• forward(...) preserves the vectorised layer output shape;
• non-zero stuck_rate creates observable stuck synapses;
• stuck weights remain unchanged after gradient updates;
• non-stuck weights update when gradients are applied;
• all weights stay clipped to [0, 1];
• stuck_rate=0.0 produces no stuck synapses.

::: sc_neurocore.layers.hardware_aware.HardwareAwareSCLayer

Predictive Coding SC Layer (Conjecture C9)

PredictiveCodingSCLayer models prediction error in the stochastic-computing domain: predicted and actual Bernoulli bitstreams are compared with XOR, and the popcount of the error stream gives the per-neuron surprise magnitude. Learning moves each prediction weight toward the observed input probability, so repeated exposure to a stable pattern should lower mean prediction error while a switched pattern should raise surprise.

The public contract is tested as follows:

• forward(...) returns prediction_error, per-neuron surprises, and (n_neurons, n_inputs) predictions;
• prediction error and surprise values remain in [0, 1];
• learned weights remain clipped to [0, 1];
• repeated inputs reduce prediction error on average;
• novel inputs after training produce larger surprise than familiar inputs;
• reset() restores the seeded initial weights.

::: sc_neurocore.layers.predictive_coding.PredictiveCodingSCLayer

Rall Branching Dendrite

RallDendrite represents each branch as a fixed-length compartment chain. Inputs are injected at the distal compartment, coupled toward the proximal compartment, and accumulated at the soma through attenuation factors derived from Rall's 3/2 branch-diameter rule.

The public contract is tested as follows:

• the initial and reset states have zero compartment and soma voltages;
• distal input reaches the soma after repeated steps;
• activating more branches increases somatic voltage relative to a single active branch under the same parameters;
• soma voltage decays when input is removed;
• branch_voltages returns a copy shaped (n_branches, branch_length);
• Rall attenuation factors remain normalised;
• distal compartments remain higher than proximal compartments during sustained distal injection.

::: sc_neurocore.layers.rall_dendrite.RallDendrite

Lateral Inhibition

::: sc_neurocore.layers.circuit_primitives.LateralInhibition

Winner-Take-All

::: sc_neurocore.layers.circuit_primitives.WinnerTakeAll