Abstract
Neuromorphic hardware using memristive devices offers high energy efficiency but suffers from manufacturing defects and operational weight drift. This paper proposes a bio-inspired self-repair mechanism for Spiking Neural Networks (SNNs) based on astrocyte-mediated synaptic regulation. By simplifying complex biological signaling into a local mathematical macro-model, the authors develop Astromorphic Spike-Timing Dependent Plasticity (A-STDP). This local rule allows individual neurons to autonomously recover accuracy from severe faults and drift, outperforming previous global methods in both recovery speed and hardware compatibility.
1. Introduction
Neuromorphic computing promises extreme energy efficiency by mimicking the brain's architecture. However, memristive devices like Phase Change Memory (PCM) used in these systems are often unreliable, suffering from stuck-at-faults (synapses stuck at zero) and temporal weight drift. While traditional AI hardware is precise but brittle, the biological brain maintains stability through complex glial cell interactions. This work explores "astromorphic" repair—leveraging the regulatory role of astrocytes—to create resilient, autonomous neuromorphic hardware.
2. Biological Inspiration and Macro-modeling
The proposed system moves beyond simple neuron-synapse models to include astrocytes, which play a critical role in maintaining neural stability. Astrocytes sense drops in total neuronal activity caused by synaptic failures and respond by boosting the transmission probability of remaining healthy synapses through signaling pathways involving neurotransmitters like 2-AG and glutamate.
The researchers simplified these complex biological interactions into a mathematical macro-model. This model captures how the astrocyte-neuron feedback loop can autonomously adjust synaptic weights to compensate for local failures, providing a blueprint for hardware-level self-healing.
3. Local Astromorphic Spike-Timing Dependent Plasticity (A-STDP)
A key contribution of this work is the translation of the astrocyte macro-model into a hardware-compatible learning rule: local A-STDP. Unlike previous "global" methods that required computationally expensive network-wide statistics (such as calculating the 98th percentile of all weights), the local rule relies only on the synapse's current weight and the total sum of weights for its specific neuron. This is ideal for hardware crossbars, where summing weights along a column is a native operation.
The researchers identified a strong inverse correlation between the severity of a fault ($z$) and the necessary repair ratio ($q$), roughly following $q \approx 1/z$. This "Inverse Proportion Rule" allows the system to effectively redistribute the workload: if a portion of synapses fails, the astrocyte signals the remaining synapses to increase their efficacy proportionally to maintain target activity levels.
4. Experimental Results
The A-STDP (local) rule was evaluated using MNIST and Fashion-MNIST benchmarks. The results demonstrate superior resilience compared to standard STDP and global astromorphic rules:
- Fault Tolerance: On MNIST, the system recovered to ~82.5% accuracy even with 70% of synapses disabled, significantly outperforming the global method (~70.8%).
- Convergence Speed: On the more complex Fashion-MNIST dataset with a 50% fault rate, the local rule achieved peak recovery in just 1,500 steps—a 10x speedup over the 17,500 steps required by global methods.
- Drift Mitigation: The mechanism successfully maintained performance despite the inherent logarithmic conductance drift typical of PCM hardware, ensuring long-term operational stability.
5. Conclusion
Astromorphic self-repair offers a promising path toward autonomous, resilient neuromorphic systems. By implementing local, bio-inspired rules, hardware can "heal" from defects and aging without central oversight or high computational overhead. Future research will focus on scaling these mechanisms to deep supervised networks and exploring the role of inhibitory pathways in network stability.
Reference: Han, Z., Islam, A. N. M. N., & Sengupta, A. (2023). Astromorphic Self-Repair of Neuromorphic Hardware Systems. Proceedings of the AAAI Conference on Artificial Intelligence, 37(6), 7821-7829.
Summary updated