Abstract
While neuromorphic computing architectures based on Spiking Neural Networks (SNNs) are increasingly gaining interest as a pathway toward bio-plausible machine learning, attention is still focused on computational units like the neuron and synapse. Shifting from this neuro-synaptic perspective, this paper attempts to explore the self-repair role of glial cells, in particular, astrocytes. The work investigates stronger correlations with astrocyte computational neuroscience models to develop macro-models with a higher degree of bio-fidelity that accurately captures the dynamic behavior of the self-repair process. Hardware-software co-design analysis reveals that bio-morphic astrocytic regulation has the potential to self-repair hardware realistic faults in neuromorphic hardware systems with significantly better accuracy and repair convergence for unsupervised learning tasks on the MNIST and F-MNIST datasets.
Introduction
Neuromorphic hardware, particularly systems utilizing memristive devices like Phase Change Memory (PCM), offers a promising route toward extreme energy efficiency in AI. However, these physical substrates are inherently "noisy" and prone to hardware-level non-idealities. Specifically, they suffer from high fault rates, such as stuck-at-zero synapses, and temporal instabilities like conductance drift.
Traditional mitigation strategies often rely on complex fault-detection schemes or computationally expensive global retraining, which are difficult to deploy on resource-constrained edge devices. While previous bio-inspired attempts at self-repair have been made, they often relied on global network statistics that lack biological plausibility and hardware efficiency. This research proposes a local, autonomous self-repair mechanism integrated directly into the learning rule, inspired by the regulatory role of astrocytes.
Biological Inspiration and Macro-modeling
To develop a hardware-efficient repair mechanism, the authors first bridged the gap between complex neuroscience and practical engineering. They began with a high-fidelity computational model of astrocyte-neuron interaction, focusing on signaling pathways involving 2-AG, Calcium waves, and Glutamate exocytosis.
Through simulation of these interactions, a key mathematical relationship was identified: the "self-repair ratio" ($q$) is inversely proportional to the "fault severity" ($z$). In biological terms, as a neuron's functional output is compromised by damage, the astrocyte detects the drop in activity and signals the remaining healthy synapses to increase their strength, effectively load-balancing the neural activity.
The A-STDP (Local) Learning Rule
This macro-model was translated into a modified Spike-Timing-Dependent Plasticity (STDP) rule, termed A-STDP (local). Unlike previous global astromorphic rules that required network-wide weight percentiles, this local version only requires information available at the individual neuron level: the current synaptic weight and the total sum of weights for that specific neuron.
This locality is critical for hardware implementation. In memristive crossbar arrays, a neuron's total weight sum can be calculated simply by measuring the total current, making the A-STDP rule perfectly compatible with existing neuromorphic architectures.
Experimental Evaluation
The proposed self-repairing SNNs were evaluated using the MNIST (handwritten digits) and Fashion-MNIST (clothing items) benchmarks. The local A-STDP rule was compared against standard STDP and a previously proposed global astromorphic rule.
Key results include:
- Superior Accuracy: On MNIST with a 70% fault rate, local self-repair achieved ~82.5% accuracy, significantly outperforming standard STDP (~70.1%) and the global repair rule (~70.8%).
- Faster Convergence: On the more challenging Fashion-MNIST dataset, the local rule reached peak accuracy up to 10x faster than the global version.
- Drift Mitigation: The system successfully compensated for PCM-specific weight drift, which typically causes performance degradation over time in memristive hardware.
Conclusion
By moving from global parameters to local, bio-morphic regulation, this work demonstrates that bio-fidelity can lead directly to hardware efficiency. The A-STDP (local) rule provides a robust, autonomous pathway for self-repairing neuromorphic systems. Future work will look toward scaling these mechanisms to deeper, multi-layer architectures and exploring the role of astrocytic modulation in inhibitory synapses.
Summary updated