What if you could grow an AI chip?
New publication in Nature Communications on electropolymerized dendritic connections for neuromorphic hardware.
GPUs are not very efficient for machine learning, but they dominate the market because custom chips — with their respective AI models — are expensive to design and manufacture. Imagine if we could grow AI devices using data.
I’m thrilled to share that our latest study on the topic has just been published in Nature Communications. In collaboration with IEMN, 3IT, Université de Sherbrooke and LN2, we investigate how electropolymerization of PEDOT-PSS fibers can be used to replicate brain-like growth of dendritic connections.
This new structural plasticity scheme produces sparse artificial neural networks with no oversampling of connections, making them efficient during training and inference on several machine learning tasks.
I want to thank all the co-authors — Kamila Janzakova, Ankush Kumar, Nikhil Garg, Corentin Scholaert, Jean Rouat, Dominique Drouin, Yannick Coffinier, Sébastien Pecqueur and Fabien Alibart — for making this study possible.
Full article: doi.org/10.1038/s41467-023-43887-8