2026 workshop program
Accelerated Computation with Auto-Differentiation: AD/PyTorch/JAX
The work
Abstract
This talk introduces automatic differentiation as a foundation for modern accelerated computation, showing how derivatives can be computed exactly and efficiently through computational graphs rather than by hand or finite differences. I will first motivate autodiff through examples in optimization, inference, and scientific modeling, then show how PyTorch uses autograd to support model building, training, and GPU-accelerated machine learning. Finally, I will discuss JAX as a more function-transform–oriented framework, emphasizing grad, jit, and vmap or writing fast, differentiable, and composable numerical code for scientific computing. This tutorial will describe the overall methods used when working with PyTorch and JAX, while also providing simple examples that can serve as building blocks for developing larger models.
Requirements: GitHub Repo