Technical Summary: HERO-SNN

Problem Statement

Spatiotemporal brain data, such as Electroencephalography (EEG) and functional Magnetic Resonance Imaging (fMRI), exhibit complex temporal dynamics and distributed spatial activity that are challenging to model. While Spiking Neural Networks (SNNs) offer a biologically plausible framework for representing these dynamics through spike-based processing, standard unsupervised learning mechanisms like Spike-Timing-Dependent Plasticity (STDP) operate purely on local spike timing. Consequently, the internal representations formed by STDP are task-agnostic; they capture general signal correlations but do not explicitly align with classification objectives or behavioral relevance.

When STDP-trained SNNs are paired with supervised readout layers (e.g., deSNN), the internal connectivity remains static after unsupervised training. This means the network may preserve dominant signal characteristics while failing to capture subtle, discriminative patterns essential for specific cognitive or clinical states. Furthermore, the classification outcome does not influence the internal synaptic structure, leaving a gap between unsupervised spatiotemporal representation learning and task-driven optimization. Existing supervised or reinforcement-based SNN methods often sacrifice biological plausibility (locality) to achieve task performance, or they fail to refine internal representations based on task outcomes.

Methodology: HERO-SNN

To address these limitations, the authors propose HERO-SNN (Homeostatic Eligibility-based Reward-Optimised Spiking Neural Network). This framework extends the existing NeuCube architecture by introducing a third learning phase that refines internal synaptic connectivity based on task performance while preserving local spike-based learning. The methodology consists of three distinct phases:

1. Unsupervised Spatiotemporal Learning

• Data Encoding: EEG signals are converted into spike trains using a threshold-based temporal encoding method. A spike is generated when the change in signal amplitude exceeds a predefined threshold ($\theta$).
• Network Mapping: Input features (EEG channels) are mapped to a 3D SNN reservoir based on anatomical coordinates (e.g., Talairach atlas). The network is initialized with "small-world" connectivity, mimicking biological systems with a mix of excitatory and inhibitory connections.
• STDP Learning: Synaptic weights are updated using unsupervised STDP rules, allowing the network to self-organize according to temporal correlations in the input data.

2. Supervised Learning (deSNN)

• After STDP training, synaptic weights in the reservoir are frozen.
• A Dynamic Evolving Spiking Neural Network (deSNN) is trained to associate the learned spatiotemporal patterns with class labels. Prototype neurons are created for each class, and connections are initialized using Rank Order (RO) learning, where earlier spikes carry higher importance.

3. Task-Aware Refinement (HERO-SNN)

This novel phase introduces a feedback loop where classification confidence modulates internal synaptic updates. The process involves:

• Forward Propagation & Confidence Computation: Training samples are replayed through the trained SNN. A temporary output neuron estimates the class confidence vector ($p_c(s)$) by comparing the current spike pattern against stored prototype neurons.
• Optimization Signal: A scalar optimization signal ($O$) is computed as the difference between the confidence of the true class and the predicted class. This signal indicates whether the internal representation supports or conflicts with the correct task outcome.
• Eligibility Traces: To identify which synapses contributed to the current decision, an eligibility trace ($e_{ij}$) is maintained for each synapse. This trace records activated synapses during the spike propagation of the replayed sample.
• Synaptic Refinement: Synaptic weights are updated using the optimization signal and the eligibility trace: $\Delta w_{ij} = \eta \cdot O(s) \cdot e_{ij}(t)$. This selectively strengthens or weakens synapses involved in task-relevant pathways.
• Homeostatic Regulation: To prevent unstable network dynamics (e.g., excessive firing or silencing) caused by repeated weight updates, a homeostatic mechanism adapts the firing threshold ($V_{th}$) of each neuron. If a neuron fires above a target rate, its threshold increases (reducing excitability); if it fires below the target, the threshold decreases. This ensures sparse, stable spiking activity while allowing adaptation.

Key Contributions

1. Task-Aware Internal Refinement: Unlike hybrid approaches where only the readout layer is supervised, HERO-SNN enables the internal spatiotemporal representations of the SNN to be refined based on task feedback.
2. Integration of Local and Global Signals: The framework successfully bridges the gap between local biological learning rules (STDP, eligibility traces) and global task objectives (classifier confidence) without abandoning spike-based locality.
3. Homeostatic Stability: The inclusion of homeostatic threshold adaptation ensures that the network maintains stable firing dynamics during the refinement process, preventing the instability often associated with reward-modulated plasticity.
4. Interpretability: By refining specific pathways based on task relevance, the model enhances the interpretability of the internal neural representations.

Experimental Results

The framework was validated using a benchmark EEG dataset from a Havening intervention study, distinguishing between a nurturing touch protocol (H+) and a protocol without touch (H−).

• Dataset: 276 spatiotemporal EEG samples (64 channels, 512 time points) from 23 participants, split into training and testing sets at the participant level to ensure generalizability.
• Performance:
  • HERO-SNN achieved a classification accuracy of 91.30% (F1-score: 0.91).
  • The baseline STDP+deSNN (without refinement) achieved 84.62% accuracy.
  • This represents an 11% improvement in accuracy over the standard unsupervised approach.
  • HERO-SNN outperformed state-of-the-art deep learning models, including CNN (83.64%) and LSTM (69.05%), under the same participant-level split conditions.
• Analysis: The results indicated that the refinement process led to localized changes in connectivity, selectively strengthening task-relevant synapses while preserving the overall structure learned during STDP.

Significance and Claims

The paper claims that incorporating task feedback directly into the adaptation of internal spatiotemporal neural representations significantly improves both classification performance and model interpretability. The authors assert that HERO-SNN demonstrates that SNNs can move beyond purely unsupervised structure learning to achieve task-specific optimization while maintaining biological plausibility through local learning rules and homeostatic regulation.

The study suggests that this approach is particularly valuable for high-dimensional, noisy brain signals (like EEG) where relevant information is often weak and distributed. By refining the internal connectivity to better capture discriminative patterns, the model overcomes the limitations of standard STDP, which often fails to align internal representations with specific cognitive or clinical states. The authors position HERO-SNN as a robust framework applicable to other spatiotemporal data types, such as fMRI time series, for mental health and neurological applications.