Natalia Berloff – Computing by Selection: Timing-Native Routing and Lookup Without Memory
Prof. Natalia Berloff, University of Cambridge
Leuchs-Russell Auditorium, A.1.500, Staudtstr. 2
Abstract
Most AI hardware is built around dense multiply–accumulate operations and repeated memory access. In this talk I will discuss a different physical primitive: computation by selection, where physics performs discrete routing decisions directly.
First, I will introduce polychronous wave computing, where relative spike timing is encoded as wave phase and evaluated by interference, with a driven–dissipative winner-take-all mechanism performing a physical argmax. This can be implemented naturally in wave-based platforms such as exciton–polariton condensates and oscillator networks, avoiding timestamping and digital comparison altogether.
Second, I will suggest implementing self-indexed holographic lookup tables, where an Ising or XY machine performs robust, margin-controlled address selection, and the selected state emits a reproducible high-dimensional physical signature (e.g. speckle or optical field). A linear optical decoder maps this signature to the desired output value, implementing lookup without explicit memory reads.
I will emphasize the unifying design laws: robustness is controlled by measurable winner–runner-up margins in the selector and by conditioning of the physical signature dictionary. Together, these ideas position polariton and Ising machines not as optimizers or matrix-multipliers, but as physics-native routing elements for spiking networks and sparse mixture-of-experts architectures.
