Peiman Mohseni
I'm a PhD candidate in computer science at Texas A&M University, advised by Nick Duffield, where I'm concurrently pursuing an MSc in mathematics. Before that I received a BSc in electrical engineering from the University of Tehran. In summer 2024 I was an AI research intern at Flagship Pioneering in Cambridge, MA, working on discrete diffusion models and GFlowNets (reinforcement learning that samples proportional to reward) for biological sequence design.
I work at the intersection of probabilistic modeling and deep learning, with a current focus on inference in function spaces, particularly neural processes, which use neural networks to parameterize stochastic processes. What makes the framing productive is that a surprising range of problems are, underneath, the same one: condition on part of a signal, predict the rest, and say how uncertain you are. Time-series and spatiotemporal modeling are the obvious case. In-context learning and amortized inference are the same move in disguise, approximate Bayesian inference in a forward pass rather than by retraining. So are inverse problems like image inpainting. Masked autoencoders invert the emphasis: the objective is still prediction from partial observations, but the prediction is a pretext and the representation it forces is the point. Part of the appeal is the interplay of statistics with architectures built around unstructured data and symmetries.
Publications
- Revisiting Neural Processes via Fourier Transform and Volterra Series ICML 2026 · code
- Spectral Convolutional Conditional Neural Processes NeurIPS 2025 · code
- QoS-Efficient Serving of Multiple Mixture-of-Expert LLMs Using Partial Runtime Reconfiguration ICML 2025 · code
- Adaptive Conditional Quantile Neural Processes UAI 2023 · code