Edge computing using biological neurons – IRVIN

AI is used on a large scale on a daily basis to perform various learning and classification tasks. As a result, computing power requirements increase exponentially. It therefore becomes mandatory to explore energy-efficient alternatives. The neuromorphic community tries to get closer to biological efficiency by designing micro-architectures taking into account bio-inspired models. However, a new way would be to use biological systems. Preliminary results published in Nature this year have shown the interest and the possibility of using biology as computation blocks.
In this IRVIN project, we aim to develop a proof of concept of using biological neurons to perform low-power edge computing, with a focus on observation, emulation and interaction. Observation consists of recording and stimulation at the cell level via the nano-electrode system. The emulation of biological dynamics relies on biomimetic algorithms using SNNs. Finally, the interaction is based on real-time neuromodulation at the level of the biological network.
The objectives of the project are: 1) to create energy-efficient biological computational units; 2) to design a bio-hybrid system where biological (in vitro) neurons will be stimulated/trained by a hardware spiking neural network for Edge computing. This project proposes to move forward in the direction of frugal bio-AI by interfacing different systems: living neurons (2D, 3D cultures, spheroids and multi-spheroid systems); recording and stimulation interface (synaptors offering superior spatial selectivity) and a neuromorphic system on FPGA analyzing the output activity of the neural system and controlling the input. The idea is to replicate within this hybrid bio-AI different learning algorithms of STDP-type neural networks or by reservoir computing in order to detect and classify temporal patterns but also speech.
These blocks will create a low-power bio-hybrid AI system whose expected results are: (1) Adaptation: Autonomous evolution over time so that it can adapt to changing conditions. Biological neurons are in 3D and change topology over time. Intelligence (2) is stored locally in the neural network, and only abstract information can exit the system, which requires (3) low computational load and the system will be autonomous and unsupervised.
The IRVIN system will open a new path in neuromorphic engineering and impact society. This biological computing would make it possible to develop systems that are faster, more efficient and more powerful than conventional computing while requiring only minimal energy. Reflections on the ethical issues of bio-hybrid systems and their acceptability in society will be planned throughout the project. Furthermore, the development of biological intelligence may help to elucidate the pathophysiology of developmental and degenerative diseases, potentially helping to identify new therapeutic approaches.