Corticomorphic CNN–SNN Architecture

• The paper presents a corticomorphic CNN–SNN architecture that integrates cortical motifs such as local receptive fields, lateral inhibition, and STDP for efficient unsupervised feature extraction.
• It employs leaky integrate-and-fire neuron dynamics with shared-weight convolutional kernels to achieve sparse connectivity and competitive performance on object recognition and attention tasks.
• The design enables low-energy, memory-efficient neuromorphic deployment, demonstrating robust generalization and high accuracy across various benchmark applications.

A corticomorphic CNN–SNN architecture is a biologically inspired neural system that integrates convolutional neural networks (CNNs) and spiking neural networks (SNNs), emulating the structural and operational motifs of the cerebral cortex. These architectures deploy local receptive fields, lateral inhibition, and microcolumnar competition, with learning mediated by spike-timing dependent plasticity (STDP), yielding computational models optimized for rapid, unsupervised feature extraction, sparsity, and low-power neuromorphic deployment. The design paradigm leverages shared-weight convolutional kernels, leaky integrate-and-fire (LIF) neuron dynamics, and spike-based event-driven learning, enabling competitive performance on object recognition and attention tasks while reducing memory and energy costs (Panda et al., 2017, Gall et al., 2023).

1. Biological and Theoretical Foundations

Corticomorphic architectures explicitly draw on neurobiological evidence from the organization of sensory cortices, where information is encoded in local receptive fields, processed by columnar microcircuits, and shaped by competitive and plastic synaptic mechanisms. In such models:

• Local receptive fields correspond to small, spatially contiguous clusters of sensory input, mimicking patches of cortex receiving input from neighboring sensory afferents.
• Microcolumnar competition is implemented by lateral inhibition, often via one-to-one interneuron connectivity producing winner-take-all dynamics.
• STDP provides a local, temporally precise synaptic update rule, in contrast to gradient-based optimization found in conventional CNNs.

These mechanisms yield a system capable of learning sparse and robust feature detectors, with neural and synaptic connectivity orders of magnitude lower than fully connected designs (Panda et al., 2017).

2. Architectural Components and Layer Organization

A typical corticomorphic convolutional SNN (“C-SNN,” Editor's term) as described in (Panda et al., 2017) comprises the following layers:

• Input (Retina) Layer: An $N \times N$ array of rate-coded Poisson spiking neurons, each encoding input features (e.g., image pixel intensities) by modulating spike rates (0–100 Hz).
• Convolutional Spiking Layer: $M$ excitatory LIF neurons (e.g., $M=400$ for MNIST), each with a distinct $k \times k$ synaptic kernel $W_j$. Kernels are convolved over the input in discrete strides ($S_h \times S_v$), establishing sparse coverage.
• Inhibitory (Competition) Layer: Each excitatory neuron is paired with an inhibitory interneuron, implementing fast lateral inhibition to enforce sparse, microcolumnar competition.
• Readout (Classifier): The system reads class labels by aggregating spike counts across neuron groups over a defined stimulus presentation window.

This structure is summarized in the table below:

Layer	Biological Motif	Functional Role
Input	Retina	Encodes stimulus in spike trains
Conv SNN	Cortical microcolumn	Spatiotemporal feature extraction
Inhibition	Lateral inhibition	Winner-take-all competition
Readout	Pooling/Decision	Classifies by spike count

3. Neural Dynamics and Plasticity

The excitatory populations operate under leaky integrate-and-fire (LIF) neuron dynamics:

$$C_m \frac{dV_j(t)}{dt} = -\frac{V_j(t) - V_{\text{rest}}}{R_m} + I_j^{\text{syn}}(t)$$

Where postsynaptic currents $I_j^{\text{syn}}$ are generated by summing kernel-weighted, temporally filtered pre-synaptic spikes. Thresholding and refractoriness govern spiking output.

Learning is mediated by a weight-dependent, pair-based STDP rule, parameterized as:

$$\Delta w = \begin{cases} \eta\,[\,e^{(t_{\text{post}}-t_{\text{pre}})/\tau_{+}} - \mathrm{offset}\,]\, (w_{\max}-w)^{\mu} & t_{\text{post}} > t_{\text{pre}} \ -\eta\,[\,e^{(t_{\text{pre}}-t_{\text{post}})/\tau_{-}} - \mathrm{offset}\,]\, w^{\mu} & t_{\text{pre}} > t_{\text{post}} \end{cases}$$

This update is triggered by the order and timing of pre- and post-synaptic spikes localized within a convolutional receptive field. The weight-dependence and exponential temporal window enable long-term stability without explicit normalization (Panda et al., 2017).

4. Feature Learning and Sparsity

The shared-weight convolutional kernels, adapted via STDP, learn features from the input using significantly fewer parameters than fully connected SNNs. Key operational features include:

• Kernel Tiling: Each $k \times k$ kernel is convolved in $M$0 positions over the $M$1 input space per neuron.
• Weight Sharing: For neuron $M$2, the same $M$3 applies across all patches; across neurons, kernels are distinct, yielding population diversity.
• Lateral Inhibition: Interneuron-driven competition ensures only one excitatory neuron adapts strongly within a patch per presentation, enforcing winner-take-all dynamics.

This approach realizes $M$4–$M$5 sparser connectivity than all-to-all networks, with energy and area benefits critical to neuromorphic hardware.

5. Experimental Protocols and Performance

Simulation protocols utilize rate-coded Poisson input, single-epoch unsupervised training, and spike-count-based readout, as outlined in (Panda et al., 2017). Cross-domain performance has been demonstrated on standard benchmarks:

• MNIST (4-class subset): Convolutional SNN (50 neurons, $M$6 kernels, $M$7 stride positions) achieves $M$8 accuracy with $M$9 fewer synapses than fully connected SNN ($M=400$0).
• MNIST (10 classes, 400 neurons): $M=400$1 test accuracy from $M=400$2 training patterns; fully connected SNNs require $M=400$3 samples for comparable performance.
• Face Detection (binary): $M=400$4 accuracy with $M=400$5 face exemplars, maintaining $M=400$6 synaptic sparsity.
• Caltech Rotated Objects: $M=400$7 accuracy on rotated objects after training on upright samples; fully connected SNNs fail in this transfer regime.
• Generalization: Training on select digit classes (e.g., $M=400$8) enables correct classification of unseen digits ($M=400$9 on $k \times k$0), highlighting feature transfer beyond the training set.

6. Corticomorphic Design for Neuromorphic Edge Applications

Hybrid CNN–SNN designs, particularly those inspired by cortical auditory pathways, have been recently deployed for EEG-based auditory attention detection (AAD) in edge-computing platforms (Gall et al., 2023). Notable attributes include:

• Low-Latency Decoding: Decision windows as low as $k \times k$1 second with eight strategically placed EEG electrodes, achieving $k \times k$2 accuracy.
• Resource Efficiency: $k \times k$3 fewer parameters and $k \times k$4 lower memory footprint compared to conventional CNNs, with reduced bit precision.
• Relevance for Embedded Hardware: These architectures support brain-embedded devices and smart hearing aids, satisfying stringent constraints on power and computation.

7. Implications and Future Directions

Corticomorphic CNN–SNN architectures validate the principle that architectural motifs from the cortex—locality, competition, shared-weights, and event-driven synaptic plasticity—facilitate compact, robust, and generalizable representations, critical for both neuroscience modeling and neuromorphic engineering. These systems combine rapid, unsupervised feature extraction, transfer learning, and robust invariances with memory and energy savings—outcomes foundational for scaling to large, heterogeneous neuromorphic arrays and cortical-scale simulation platforms (Panda et al., 2017, Gall et al., 2023).