Evolutionary Computing in Nanomaterial Networks

Motivated by our recent realization of reconfigurable computational functionality in a designless nanoparticle network (Nature Nanotechnology 10, 1048 (2015)), we aim at realizing a hardware implementation of a neural network through artificial evolution, with as ultimate goal efficiently solving real-life computational problems. Artificial Neural Networks (ANNs) are mathematical models inspired by biological neural networks. ANNs are increasingly used throughout society for e.g. pattern recognition, medical diagnosis, data mining and vehicle control. Those tasks are often extremely hard to solve using conventional, program-based computing. ANNs are normally implemented on conventional computers using software, which requires a lot of computational resources. Alternatively, current neuromorphic, i.e. brain-like, architectures are closer to the intrinsic ANN structure, but are still built from conventional components. Here, we propose to use the principle of Evolution-in-Materio, a bottom-up approach where the intrinsic underlying physics of materials is exploited, to realize a hardware implementation of ANNs. More than in other ANN hardware implementations, we expect to take full advantage of the intrinsic computational power of nanoscale matter. Artificial evolution is used to configure the electrical potential landscape of interconnected nanoparticle networks to evolve functionality. The key elements – or artificial neurons – are single-electron transistors formed by the nanoparticles, strongly non-linear switches, giving rise to emerging, overall functional behavior. By massively interconnecting nanoparticle networks into a “super network”, we obtain an exponentially enlarged state space compared to our single network, entering uncharted territory. The enormous state space is expected to allow for more advanced functionality, well beyond our successful proof of concept, such as pattern recognition and vehicle control. The power of our approach is already suggested by numerical simulations based on both physical and neural-network models. We further extend our proof of concept by introducing internal memory into our nanomaterials system. We propose to realize this by including redox-active molecules in our networks, acting as floating gates to the metallic nanoparticles. These molecular floating gates can reversibly change the charge state of neighboring NPs by light irradiation or applied voltages. We deem the inclusion of internal memory a crucial step towards higher complexity. Using molecules and smaller nanoparticles will also lead to higher operating temperatures, significantly facilitating scaling and applicability. To train our networks in a scalable manner, we propose to apply the concepts of reservoir computing and online training. In this project we synergistically combine experiment (Van der Wiel, UT) and theory (Bobbert, TU/e) with the aim of showing that our approach is scalable, defect-tolerant, and efficient in time and energy. This project is fundamental and curiosity-driven. However, realizing an energy- and time-efficient hardware implementation of ANNs makes this project not only scientifically highly relevant, but also technologically very timely. Toyota Research has indicated that it will co-finance this project in cash if granted by NWO. IMEC and Tyndall, the Dutch Foundation for Neural Networks (SNN) and the software company Simbeyond have expressed strong interest in this proposal, and will collaborate.