This paper is a contribution to the Physical Review Applied collection titled Physics-Inspired Computing.

The human brain functions very differently from artificial neural networks (ANNs) and possesses unique features that are absent in ANNs. An important one among them is “adaptive synaptogenesis,” which modifies synaptic weights when needed to avoid catastrophic forgetting and to promote lifelong learning. In the model described here, a supervised form of adaptive synaptogenesis uses local error signals to modify synaptic weights and reduce errors. In the brain, supervisory signals may come from a variety of brain regions; in the model, supervisory signals are provided as class labels corresponding to each feature vector. In this work, we discuss various algorithmic aspects of adaptive synaptogenesis tailored to edge computing, demonstrate its function using simulations, and design nanomagnetic hardware accelerators for specific functions such as mean-firing-rate estimation. Our approach attempts to combine two disparate fields—neuroscience and spintronics—on a common hardware platform.

Close

PrevNext

View figure in article

• Inhibitory synapses
• Neural encoding
• Neural information
• Spintronics
• Straintronics
• Artificial neural networks
• Brain network
• Magnetic tunnel junctions