Sahil Sidheekh: “Building Deep Generative Models That Can Perform Tractable Probabilistic Inference”

My research focuses on developing generative models that combine expressive deep learning representations with tractable probabilistic inference. A key contribution is Probabilistic Flow Circuits (PFCs), introduced in my UAI 2023 paper, which unify probabilistic circuits with normalizing flows. This framework retains the flexibility of deep generative models while enabling exact inference, making it well-suited for density estimation, structured generation, and uncertainty quantification. Additionally, my IJCAI 2024 survey on building expressive and tractable generative models provides a comprehensive outline of the field for understanding the trade-offs between expressivity, tractability, and learning efficiency. It explores how hybrid models—such as deep-learning-enhanced probabilistic circuits—address fundamental challenges in generative modeling, paving the way for more efficient and interpretable AI systems.