Evolving interpretable plasticity for spiking networks

• DOI: 10.7554/elife.66273
• Publication Date: 2021-10-28T10:00:13Z
• Authors: Jakob Jordan; Maximilian Schmidt; Walter Senn; Mihai A Petrovici
• Keywords: Neurons; 0301 basic medicine; 0303 health sciences; synaptic plasticity; Neuronal Plasticity; spiking neuronal networks; QH301-705.5; Science; Q; Models, Neurological; R; 610 Medicine & health; 03 medical and health sciences; learning to learn; metalearning; Medicine; Animals; Humans; genetic programming; Biology (General); Nerve Net; Computational and Systems Biology

Continuous adaptation allows survival in an ever-changing world. Adjustments the synaptic coupling strength between neurons are essential for this capability, setting us apart from simpler, hard-wired organisms. How these changes can be mathematically described at phenomenological level, as so-called 'plasticity rules', is both understanding biological information processing and developing cognitively performant artificial systems. We suggest automated approach discovering biophysically plausible plasticity rules based on definition of task families, associated performance measures biophysical constraints. By evolving compact symbolic expressions, we ensure discovered amenable to intuitive understanding, fundamental successful communication human-guided generalization. successfully apply our typical learning scenarios discover previously unknown mechanisms efficiently rewards, recover efficient gradient-descent methods target signals, uncover various functionally equivalent STDP-like with tuned homeostatic mechanisms.Our brains incredibly adaptive. Every day form memories, acquire new knowledge or refine existing skills. This stands contrast current computers, which typically only perform pre-programmed actions. Our own ability adapt result a process called plasticity, connections change. To better understand brain function build adaptive machines, researchers neuroscience intelligence (AI) modeling underlying mechanisms. So far, most work towards goal was guided by human intuition – that is, strategies scientists think likely succeed. Despite tremendous progress, has two drawbacks. First, time limited expensive. And second, have natural reasonable tendency incrementally improve upon models, rather than starting scratch. Jordan, Schmidt et al. now developed ‘evolutionary algorithms’. These computer programs search solutions problems mimicking evolution, such concept fittest. The exploits increasing availability cheap but powerful computers. Compared its predecessors (or indeed brains), it also uses less biased previous models. evolutionary algorithms were presented three scenarios. In first, had spot repeating pattern continuous stream input without receiving feedback how well doing. second scenario, received virtual rewards whenever behaved desired manner example reinforcement learning. Finally, third ‘supervised learning’ told exactly much behavior deviated behavior. For each scenarios, able solve successfully. Using study computers ‘learn’ will provide insights into health disease. It could pave way intelligent machines needs their users.