Ben Poole is a research scientist at Google DeepMind, where he works on generative artificial intelligence and machine learning. He is best known as a co-author of several influential papers in deep learning, including DreamFusion, Imagen, and foundational work on score-based generative models. His research focuses on developing methods for generating high-quality images, videos, and 3D content from text descriptions, with applications in creative tools and scientific simulation.
Poole's contributions have helped shape the modern landscape of generative AI, particularly in the areas of text-to-image synthesis and 3D generation. His work bridges theoretical advances in probabilistic modeling with practical engineering solutions, making him a prominent figure in the field. He has collaborated with researchers across academia and industry, and his papers are widely cited in the AI community.
Early Career and Education
Poole completed his PhD in computer science at the University of Toronto, where he was advised by Aaron Courville and worked on deep learning and probabilistic models. During his doctoral studies, he focused on variational inference and generative modeling, publishing papers on topics such as black-box variational inference and neural attention models. His thesis work laid the groundwork for his later contributions to score-based generative models.
After completing his PhD in 2016, Poole joined the research team at OpenAI as a research scientist. At OpenAI, he worked on reinforcement learning and generative models, contributing to projects involving neural networks and loss functions. He remained at OpenAI until 2020, during which time he co-authored several notable papers, including work on implicit generative models and adversarial training.
Key Contributions to Generative Modeling
In 2021, Poole co-authored the paper "Score-Based Generative Modeling with Critically-Damped Langevin Diffusion," which introduced a new framework for score-based generative models. This work, published at the International Conference on Learning Representations (ICLR), demonstrated how to improve sample quality and training stability by using critically-damped Langevin dynamics. The paper is considered a foundational contribution to the field, influencing subsequent research on diffusion models.
In 2022, Poole was a lead author on "DreamFusion: Text-to-3D using 2D Diffusion," a paper presented at ICLR 2023. DreamFusion introduced a method for generating 3D models from text prompts by leveraging a pre-trained 2D diffusion model, such as Imagen, without requiring 3D training data. The approach, which uses a technique called Score Distillation Sampling, became widely adopted and inspired numerous follow-up works in text-to-3D generation. The paper received the Outstanding Paper Award at ICLR 2023.
Also in 2022, Poole contributed to "Imagen: Photorealistic Text-to-Image Diffusion Models," a paper led by Chris Bishop's team at Google Research. Imagen demonstrated state-of-the-art text-to-image synthesis using a large Transformer language model to encode text and a cascade of diffusion models to generate images. The paper, which was released as a preprint and later published at NeurIPS 2022, achieved unprecedented photorealism and helped establish diffusion models as the dominant approach in text-to-image generation.
Research at Google DeepMind
Poole joined Google Research in 2020, which later merged with DeepMind to form Google DeepMind in 2023. At Google DeepMind, he continues to work on generative models, focusing on scaling up diffusion models and improving their efficiency. His recent work includes research on video generation, where he has explored methods for producing temporally coherent videos from text or image inputs. He has also investigated techniques for controllable generation, such as using cross-attention mechanisms to guide the output of diffusion models.
In addition to his research, Poole has contributed to open-source tools and frameworks. He has been involved in developing libraries for machine learning research, including JAX-based implementations of diffusion models. His work has been presented at major conferences such as NeurIPS, ICLR, and ICML, and he has served as a reviewer and area chair for these venues.
Collaborations and Impact
Poole has collaborated with a wide range of researchers, including Jakob Uszkoreit, Llion Jones, and others at Google DeepMind. His work with Anima Anandkumar at Caltech on score-based generative models has been particularly influential, leading to the development of methods that are now standard in the field. The DreamFusion paper, in particular, has spawned a large body of research on text-to-3D generation, with applications in gaming, virtual reality, and robotics.
His papers have accumulated tens of thousands of citations, reflecting their broad impact on both academic research and industry applications. The techniques he helped develop are used in commercial products, such as image generation tools and 3D asset creation pipelines. As of 2024, Poole remains an active researcher at Google DeepMind, continuing to push the boundaries of what is possible with generative AI.
Selected Publications
- "Score-Based Generative Modeling with Critically-Damped Langevin Diffusion" (ICLR 2021)
- "DreamFusion: Text-to-3D using 2D Diffusion" (ICLR 2023, Outstanding Paper Award)
- "Imagen: Photorealistic Text-to-Image Diffusion Models" (NeurIPS 2022)
- "Deep Unsupervised Learning using Nonequilibrium Thermodynamics" (co-author, 2015)
- "Black-Box Variational Inference for Stochastic Differential Equations" (ICML 2016)
These publications highlight his contributions to both theoretical foundations and practical applications of generative modeling. His work continues to influence new generations of researchers in artificial intelligence and related fields.