Eli Laird

Appearance

Research

World Models ​

Physics-grounded world models My primary focus is developing physics-grounded world models for robotics and continuous control. Current world models often fail to obey physical laws and to generalize to unseen framerates. This limits their applicability to real-time deployments where noisy observations and varying framerates are common. My work aims to mitigate these limitations by embedding physical principles into network architectures.

Memory Architectures and their impact on learning physical symmetries (ex: Loop Closures) The ability of a world model to represent physical symmetries, like loop closures, in its predicted trajectories is crucial for long-horizon control tasks. Transformer-based models, in particular, struggle with loop closures without large context windows. My research explores how memory architectures can help world models learn and represent these symmetries and provides a thorough analysis of the various memory architectures and their limitations.

Previous Work ​

Graph Neural Networks and Drug Discovery ​

My research explored equivariant graph neural network architectures for modeling quantum chemical properties of molecules. In addition, I explored explainablility techniques for understanding the predictions of these models to aid in discovering new directions for treatment research.

Key Projects:

  • SmartCADD: An open-source virtual drug screening platform combining AI-based filtering with Quantum Mechanical methods.
  • XInsight: A flow-based explanation method for GNNs that identifies key subgraphs responsible for model predictions, offering granular insights into decision-making processes.
  • Message-Passing JEPAs: Exploring Joint Embedding Predictive Architectures (JEPA) combined with message-passing neural networks to learn predictive representations of dynamic systems. (This ultimately ran into issues with overfitting due to graph data being highly imbalanced and overfitting to the graph structure rather than the underlying physics.)
  • Equivariant GNNs: Developing architectures that respect symmetries (e.g., rotation, translation) for applications in quantum chemistry and physical simulation.

Biometrics & Face Recognition ​

My background in data science includes extensive work on fairness, bias, and evaluation in biometric systems. I focused on developing bias metrics and synthetic data pipelines to audit and improve face recognition algorithms.

Key Projects:

  • Synthetic Identity Generation (SIG): A pipeline for generating large-scale, controllable synthetic datasets to evaluate face recognition systems without privacy concerns.
  • Bias Mitigation (GARBE): Proposed the Gini Aggregation Rate for Biometric Equitability (GARBE), a metric now included in ISO standards (ISO/IEC 19795-10:2024) for measuring demographic differentials in biometric performance.
  • Bias Removal: Identifying components of face embeddings produced by open-source face recognition models (like ArcFace) that are responsible for bias and developing methods to remove them.