Daniel Zoran

Massachusetts Institute of Technology, Google DeepMind (United Kingdom)

Papers

Total Citations

271

H-Index

About

Daniel Zoran is a leading researcher in computer vision and machine learning, with a primary focus on enabling machines to perceive and reason about the physical world in an unsupervised manner. His major contributions lie at the intersection of physical understanding and visual perception, most notably through the introduction of **Visual Interaction Networks** (VINs). In his seminal 2017 paper, which has garnered over 188 citations, Zoran demonstrated that a neural network could learn a physics simulator directly from video, predicting the future state of objects without explicit state information—a task previously requiring domain-specific engineering. This work, along with its follow-up, established a new paradigm for learning intuitive physics from raw visual data. More recently, Zoran has advanced unsupervised object-centric learning with **PARTS** (2021), a method that uses slot attention and independence maximization to decompose visual scenes into coherent objects without labels. His research is notable for pushing the boundaries of what can be learned from passive observation, bridging the gap between human-like physical intuition and artificial systems. Zoran’s work has been highly influential in robotics, graphics, and cognitive AI, inspiring a generation of models that learn structured representations from unstructured video.