New software toolbox enables brain-like models to learn directly from data (www.biorxiv.org)

Researchers announced Jaxley, a differentiable simulation toolbox that lets detailed biophysical neuron models learn directly from data by combining automatic differentiation with GPU acceleration. The framework turns complex neuron and network simulations into gradient-enabled systems so parameters can be optimized with standard gradient-descent methods. That addresses a long-standing bottleneck in neuroscience: fitting high-dimensional, mechanistic models to physiological measurements at scale. Technically, Jaxley can infer hundreds of biophysical parameters to match intracellular voltage traces or two-photon calcium recordings—sometimes orders of magnitude faster than prior approaches—and scale up to task-driven training. The authors demonstrate training a recurrent network for working memory and a feedforward network of morphologically detailed neurons with ~100,000 parameters to solve a computer-vision task. By enabling efficient, gradient-based fitting of large, mechanistic models to both data and tasks, Jaxley opens the door to building task- and data-constrained biophysical models that link cellular physiology to computation across scales, accelerating hypothesis testing about neural mechanisms.