Neural Evolution for Augmenting Topologies in Printed Neuromorphic Circuits

In the context of fast-paced advancements in the Internet of Things and artificial intelligence, the demand for custom solutions in emerging fields like smart packaging and smart bandages has risen significantly. These innovative applications require electronics that are not only ultra-low-cost but also highly flexible and customizable, particularly for edge processing tasks. Traditional silicon-based electronics often fall short in offering cost-effectiveness. In contrast, printed electronics have emerged as a powerful alternative. They use additive manufacturing techniques to create custom electronic circuits at ultra-low cost. These electronics are particularly versatile, with a choice of materials and substrates contributing to notable flexibility and bio-compatibility. To further equip them with computational abilities, there is an increasing interest in printed neuromorphic circuits. These circuits effectively merge the benefits of neuromorphic computing with the capabilities of printed electronics. Nevertheless, the intrinsic constraints of printed electronics, namely large feature sizes and notably reduced integration density, pose challenges for compact application areas. ... mehrTo address these challenges, we propose a method inspired by the Neural Evolution for Augmenting Topologies algorithm. Unlike traditional gradient-based optimization methods, such as pruning, our approach simultaneously optimizes the learnable parameters and the topology structure of printed neuromorphic circuits, thereby facilitating a reduction in circuit area. Through experiments conducted on 11 different datasets, we demonstrate the effectiveness of our approach. Experimental results reveal that, with the proposed approach, 3.1× reduction of the circuit area can be realized while maintaining 100% of classification accuracy of the gradient-based pruning method.