Event:

A central challenge in neuroscience is that many properties of neural systems cannot be measured exactly. This limits our understanding of these systems and our ability to build simulations that match experimental recordings or predict neural responses to unseen stimuli. Inference allows scientists to identify parameters - the properties that cannot be measured exactly - such that biophysical simulations are consistent with experimental measurements of neural activity. I will present machine learning-based inference methods for biophysical simulations in neuroscience. First, I will discuss simulation-based Bayesian inference using neural networks. Second, I will introduce differentiable simulation as a powerful approach for parameter inference in large-scale biophysical models. This allows us to directly optimize parameters using gradient-based methods, even in morphologically detailed models of single cells or networks - overcoming scalability barriers of previous approaches. Together, these results demonstrate how machine learning opens new possibilities for constructing and fitting biophysical simulations in neuroscience, helping to bridge the gap between experimental data and mechanistic understanding.