Research

Investigating intelligence from its foundations to its frontiers.

Our research spans seven domains — from the mathematics of learning to the use of machine intelligence in the sciences. Each has its own questions and methods; together they describe a single subject studied from several directions.

01 / 07

Artificial Intelligence

We study intelligence as the capacity to act well in situations that were not anticipated in advance: to form plans, revise them in light of evidence, and pursue goals over long horizons. Our interest is in agents that combine learned perception with deliberate reasoning, and in the principles that make such combinations reliable.

Representative subjects

  • Reasoning and planning
  • Intelligent agents
  • Multi-agent systems
  • Knowledge representation
  • Decision-making
  • Reinforcement learning
  • General-purpose intelligent systems
Explore Artificial Intelligence
Fig. 01Schematic. A search tree over possible actions; the highlighted branch marks one plan selected under a learned value estimate.

02 / 07

Machine Learning

Machine learning is the substrate on which most of modern AI is built, yet many of its central phenomena remain poorly explained. We study why learning systems generalize, what their representations encode, how optimization shapes the solutions they find, and how to obtain more capability from less data and compute.

Representative subjects

  • Representation learning
  • Optimization
  • Deep learning
  • Self-supervised learning
  • Generalization
  • Efficient learning
  • Model evaluation
Explore Machine Learning
Fig. 02Schematic. A curved surface stands in for a learned representation space; the trajectory traces an optimizer descending toward a low-loss region.

03 / 07

Generative AI

Generative models learn the structure of a domain well enough to produce new examples from it. That capacity has made them the most general-purpose systems in machine learning. We study how they represent the world, how their outputs can be controlled and verified, and how they can be made efficient and dependable.

Representative subjects

  • Foundation models
  • Large language models
  • Diffusion models
  • Multimodal generation
  • Generative representations
  • Model efficiency
  • Generative system reliability
Explore Generative AI
Fig. 03Schematic. Unstructured samples (left) are progressively transformed into an ordered representation (right), as in iterative generative processes.

04 / 07

Natural Language Processing

Language is both a subject of study and the medium through which people increasingly interact with machines. We investigate how computational systems represent meaning, how they relate words to the world, how they reason in language, and how that ability can extend fairly across the world's languages.

Representative subjects

  • Language understanding
  • Computational semantics
  • Multilingual modeling
  • Grounded language learning
  • Information retrieval
  • Language reasoning
  • Speech–language systems
Explore Natural Language Processing
Fig. 04Schematic. Arcs between tokens illustrate the pairwise dependencies an attention mechanism can represent; line weight indicates relative strength.

05 / 07

Computer Vision

Vision is how many intelligent systems first encounter the world. We study how machines build representations of images and video, recover three-dimensional structure, reason about space and motion, and connect what they see to language and action.

Representative subjects

  • Visual representation learning
  • Image understanding
  • Object recognition
  • Video understanding
  • Spatial reasoning
  • 3D perception
  • Vision–language learning
Explore Computer Vision
Fig. 05Schematic. Nested receptive fields over an image grid feed a hierarchy of increasingly abstract feature maps.

06 / 07

AI Safety

As AI systems become more capable, the gap between what they can do and what we can verify about them becomes a central technical problem. We treat safety as a research discipline: developing methods to understand model internals, measure behavior honestly, quantify uncertainty, and keep systems aligned with the intentions of the people who use them.

Representative subjects

  • Interpretability
  • Alignment
  • Robustness
  • Model evaluation
  • Distribution shift
  • Uncertainty
  • Responsible deployment
Explore AI Safety
Fig. 06Schematic. A model's prediction with an uncertainty band that widens away from observed data — the behavior a well-calibrated system should exhibit.

07 / 07

Scientific Discovery

Machine learning offers new instruments for science: models that approximate expensive simulations, propose candidate structures, find regularities in high-dimensional data, and help formulate hypotheses. We study how to build these instruments so that they respect the constraints of the domains they serve and produce results scientists can check.

Representative subjects

  • Scientific machine learning
  • Computational biology
  • Physics-informed learning
  • Chemistry and materials modeling
  • Simulation and surrogate models
  • Scientific reasoning
  • AI-assisted hypothesis generation
Explore Scientific Discovery
Fig. 07Schematic. Streamlines of a two-dimensional vector field with two sources of rotation — the kind of structure a learned surrogate must reproduce.

Connections

Directions that interact, not departments.

The most interesting problems sit between domains. A question about safety is often a question about representations; a question about scientific discovery is often a question about planning under uncertainty.

01 Artificial Intelligence02 Machine Learning03 Generative AI04 Language05 Vision06 Safety07 Scientific Discovery
  1. 02 ↔ 06: Machine Learning and AI Safety

    Representations and interpretability

    Understanding what models learn is the basis for verifying how they will behave.

    Machine LearningAI Safety

  2. 04 ↔ 03: Natural Language Processing and Generative AI

    Language and generation

    Large language models are simultaneously linguistic objects and general-purpose generators.

    Natural Language ProcessingGenerative AI

  3. 05 ↔ 04: Computer Vision and Natural Language Processing

    Multimodal learning

    Grounding words in perception, and describing perception in words.

    Computer VisionNatural Language Processing

  4. 01 ↔ 07: Artificial Intelligence and Scientific Discovery

    Agents for science

    Experimental design, simulation, and hypothesis search as sequential decision problems.

    Artificial IntelligenceScientific Discovery

  5. 01 ↔ 02: Artificial Intelligence and Machine Learning

    Reasoning and learning

    Combining learned representations with explicit search and planning.

    Artificial IntelligenceMachine Learning

  6. 03 ↔ 06: Generative AI and AI Safety

    Reliable generation

    Evaluating, calibrating, and constraining the outputs of generative systems.

    Generative AIAI Safety

  7. 05 ↔ 07: Computer Vision and Scientific Discovery

    Scientific imaging

    Microscopy, astronomy, and remote sensing as problems in visual inference.

    Computer VisionScientific Discovery

Working on related questions? We would like to hear about it.