Probabilistic photonic computing

The neuroscience principle of free energy minimization (FEM) suggests that living organisms create internal models of their environment in order to minimize surprise and manage uncertainty. This is strikingly different from artificial neural networks (ANNs), which prioritize maximizing accuracy. Although ANNs excel in applications such as natural language processing and weather forecasting, they struggle with real-time, safety-critical tasks like autonomous navigation due to their reliance on deterministic hardware in the von Neumann architecture which is poorly suited for distribution estimation and parameter extraction in probabilistic models. Photonic analog computing enables a paradigm shift for probabilistic processing by exploiting inherent physical stochasticity via direct encoding of information in physical quantities and by permitting ultralow latency and high throughput. Here, I will leverage hybrid photonic integrated circuits to harness physical random number generation (RNG) for probabilistic computing. I will develop chaotic light sources based on Erbium-doped waveguide amplifiers as physical sources of entropy for RNG at telecom wavelengths. Using time-wavelength interleaving of amplitude-bandwidth encoded probabilistic weights and broadband ultrafast waveguide-integrated modulators for vector encoding, I will achieve probabilistic sampling at rates beyond 300 Tera-operations per second. For deterministic convolution processing, I will realize ultra-high throughput programmable photonic crossbar arrays using silicon photonic circuits. By hybrid integration via 2D-3D nanoprinting, I will link different computing platforms into advanced systems. Combining deterministic and probabilistic photonic processors, I will realize disruptive compute architectures for mixed-mode probabilistic-deterministic deep neural networks, achieving Tera-scale probabilistic compute performance, and enabling real-time Bayesian object recognition beyond 100 frames per second.