RESEARCH / FOUR PERSPECTIVES

Ideas meet at the edges.

Four real fields. Four ways of understanding the world. Connect disciplines through original papers, open projects and questions worth pursuing.

01 / SPATIAL INTELLIGENCE

Spatial intelligence & simulation

How can machines understand a world in three dimensions?

From images to 3D scenes, from geometry to physical laws. Explore the connections between neural rendering, differentiable simulation and embodied perception.

NeRF3D Gaussian SplattingSimulation

NeRF represents a scene as a neural radiance field; 3D Gaussian Splatting uses explicit 3D Gaussians for high-quality real-time novel-view synthesis. Rendering quality, reconstruction accuracy and physical consistency are distinct evaluation questions.

3D Gaussian Splatting · SIGGRAPH 2023↗NeRF · ECCV 2020↗

02 / BIOMEDICAL IMAGING

Biomedical imaging

See more detail with fewer measurements.

Explore accelerated MRI, image reconstruction and inverse problems, where data-driven methods meet imaging physics, uncertainty and the limits of clinical applicability.

MRIInverse ProblemsReconstruction

fastMRI is an open research initiative from NYU Langone Health and Facebook AI Research to advance accelerated MRI. Dataset access is subject to its terms; benchmark performance is not evidence of clinical effectiveness.

fastMRI · NYU Langone Health↗fastMRI · Open-source code↗

03 / COMPUTER & SOFTWARE SCIENCE

Computer & software science

Reliable software begins with verifiable reasoning.

From program semantics to interactive theorem proving, from system design to reproducible experiments: investigate correctness, expressiveness and engineering practice.

LeanFormal MethodsSystems

Lean is a functional programming language and interactive theorem prover. Its community library mathlib supports formalized mathematics. Begin with the relationship between propositions, types and proofs.

Lean · Language & theorem prover↗Mathematics in Lean · Open textbook↗

04 / LEARNING & DECISION MAKING

Game science & reinforcement learning

Learn through interaction. Decide under uncertainty.

Follow the thread of sequential decision-making through agents, environments, rewards and policies, linking control, games and multi-agent cooperation.

Reinforcement LearningAgentsControl

Gymnasium maintains a standard API and reference environments for reinforcement learning. Record environment versions, random seeds, training budgets and evaluation protocols rather than treating one high score as a robust result.

Gymnasium · Farama Foundation↗Spinning Up · OpenAI↗