About the session
As demand for computational power surges - driven by machine learning and optimization - traditional digital electronics are reaching their limits. Meanwhile, advances in component design are enabling novel platforms like quantum and optical computing, challenging conventional systems and calling for a fresh look at computing principles. We believe the moment is right to explore the potential of analog computing.
Leveraging the inherent parallelism and high-bandwidth capabilities of optical systems, we are developing an Analog Optical Computer (AOC) that promises to accelerate critical computations by up to two orders of magnitude over current digital processors. However, fully harnessing the potential of this technology demands a new approach to how applications are encoded and optimized for analog hardware.
In the context of optimization, users of the AOC must express computations using the Quadratic Unconstrained Mixed Optimization (QUMO) formalism. For machine learning, we are exploring several neural network architectures including generative models that can be efficiently realized on analog hardware, offering substantial improvements in both performance and energy efficiency over conventional digital systems.
In this talk, I will present our progress in developing the AOC, including case studies of re-framing key optimization problems into the QUMO model, with applications in financial services (in collaboration with Barclays) and medical imaging (in partnership with Microsoft Health Futures). I will also share insights into designing neural networks optimized for analog hardware, pointing the way to the next generation of machine learning architectures.
Dr Christos Gkantsidis, Principal Researcher, Microsoft Research, UK
