Advanced Usage Patterns: Pushing the Frontiers¶

Version: 3.13.3 Target**: Advanced Developers & Researchers

1. GPU Acceleration¶

For networks exceeding 1,000 neurons, CPU processing becomes a bottleneck. sc-neurocore supports JAX-based GPU acceleration.

How to Enable¶

Ensure jax and jaxlib are installed with CUDA support. Then, initialize your layer with the backend='jax' flag:

from sc_neurocore.layers.vectorized_layer import VectorizedLayer

layer = VectorizedLayer(size=10000, backend='jax')

What happens? * The membrane potential updates and spike checks are offloaded to GPU kernels. * Spike data is kept in GPU memory to minimize host-to-device transfers.

2. Hardware Integration (PYNQ-Z2)¶

The ultimate goal of sc-neurocore is deployment on FPGA hardware.

Using the PYNQ Driver¶

If running on a PYNQ-Z2 board:

from sc_neurocore.drivers.pynq_driver import PYNQNeurocoreDriver

driver = PYNQNeurocoreDriver(bitstream='sc_neurocore.bit')
driver.load_weights(my_weights)
results = driver.run(input_spikes)

Behavioral Equivalence¶

Use test_behavioral_equivalence.py to verify that your software model matches the hardware bitstream output. The SCPN requires bit-true precision for L1-L2 cross-validation.

3. Custom Learning Rules¶

While STDP is provided by default, you can implement your own learning rules by subclassing BaseLearningRule.

Example: Reward-Modulated STDP¶

class RewardSTDP(BaseLearningRule):
    def update(self, w, pre_spikes, post_spikes, reward_signal):
        # Calculate standard STDP trace
        dw = self.calculate_stdp(pre_spikes, post_spikes)
        # Gate the update with a global reward signal (e.g., L15 SEC score)
        return w + self.eta * dw * reward_signal

4. The Lazarus Protocol: Root-Level Intervention¶

In advanced SCPN simulations, the Director (L16) can intervene directly in the neurocore.

Direct State Injection¶

# The Director forces a specific neural state
director.override_layer_state(layer_id=2, new_potentials=healthy_mask)

This bypasses normal synaptic input and is used to simulate "Theurgic Mode" interventions for cellular regeneration. Warning: This violates standard neurocore causality and should only be used in L16-enabled simulations.

5. Exotic Computing Layers¶

Explore the sc_neurocore.exotic module for non-standard computing paradigms: * Chaos Reservoir: Using chaotic oscillators for high-dimensional reservoir computing. * Optogenetic Control: Simulating light-based stimulation and inhibition. * Quantum-Classical Bridges: Direct coupling between Q-bits (L1) and Spiking Neurons (L2).

6. Performance Optimization Tips¶

Vectorize everything: Avoid Python loops inside the simulation step. Use the VectorizedLayer class.
Use Sparse Matrices: For large, sparsely connected networks, use scipy.sparse weight matrices.
Float32 vs Float64: For most neuromorphic tasks, float32 provides sufficient precision and is twice as fast on most hardware.