Sara Achour

Achour joined Stanford as an assistant professor in 2021. She earned undergraduate degrees in computer science and computational biology at UCLA, followed by a computer-science master’s and PhD at MIT. Her research asks how a programmer can use specialized hardware without manually translating every calculation into the device’s signals, connections and physical constraints.[1][3]

Compile equations into a working analog configuration

Achour and Martin Rinard developed Legno for programmable analog hardware. The compiler takes a dynamical-system description and configures computing blocks and connections to reproduce its behavior. It accounts for noise, limited operating ranges, quantization and device variation. Their paper evaluates compiled computations on a physical development board, connecting a high-level mathematical description to an actual circuit’s behavior.[5]

Explore analog designs before committing to a circuit

With Yu-Neng Wang, Glenn Cowan and Ulrich Rührmair, Achour coauthored Ark, a language for describing analog computing approaches. Its validator and compiler let designers introduce physical behaviors progressively and simulate their effects. The work explores a middle ground between a fixed circuit for one purpose and a highly general accelerator, making specialization and imperfect hardware behavior part of the design process.[6]

Her Machine Learning with Dynamics tutorial encodes pen-stroke inputs as nudges to simulated gyroscopes and reads their later positions and velocities. Springs connect the elements so information can reach the outputs. Training adjusts both neural encoder/decoder weights and the dynamical system’s parameters through a differentiable simulation. The example explains dynamics-based computation; its simulation results do not establish the performance of a manufactured accelerator.[4]