New AI Tool Teaches Digital Neurons to Behave Like Real Ones

Building realistic digital models of neurons is a monumental task. These 'biophysical models' have thousands of parameters that must be perfectly tuned to replicate the complex electrical and chemical behaviour of real brain cells, a challenge that has historically slowed progress.

A new framework called JAXLEY dramatically speeds up this process. By borrowing powerful techniques from modern artificial intelligence, such as automatic differentiation, and running on GPUs, JAXLEY can efficiently 'train' these models. It automatically adjusts the vast number of parameters, much like how machine learning models are optimised.

The results are impressive. JAXLEY can configure models to precisely match real brain recordings, sometimes orders of magnitude more efficiently than older methods. Researchers have already used it to train a network of detailed neurons with 100,000 parameters to perform a computer vision task, paving the way for a deeper understanding of neural computation.