Cross-homeostatic plasticity enables analog neuromorphic circuits to exhibit robust computational primitives

Many neural computations emerge from self-sustained patterns of activity in recurrent neural circuits, which rely on balanced excitation and inhibition. Neuromorphic electronic circuits that use the physics of silicon to emulate neuronal dynamics represent a promising approach for implementing the brain's computational primitives, including self-sustained neural activity. However, achieving the same robustness of biological networks in neuromorphic computing systems remains a challenge, due to the high degree of heterogeneity and variability of their analog components. Inspired by the strategies used by real cortical networks, we apply a biologically-plausible cross-homeostatic learning rule to balance excitation and inhibition in neuromorphic implementations of spiking neural networks. We demonstrate how this learning rule allows the neuromorphic system to overcome device mismatch and to autonomously tune the spiking network to produce robust, self-sustained attractor dynamics in an inhibition-stabilized regime. We also show that this rule can implement a stable working memory, and that the electronic circuits can reproduce biologically relevant emergent neural dynamics, including the so-called "paradoxical effect". In addition to validating neuroscience models on a substrate that shares many similar properties and limitations with biological systems, this work enables the construction of ultra-low power, mixed-signal neuromorphic technologies that can be automatically configured to compute reliably, despite the large on-chip and chip-to-chip variability of their analog components.