LMLi Meng

Research

Structure is the common thread.

My work asks how learned representations can encode the factors that matter for prediction and decision-making—especially when the observations are high-dimensional, partially observable, or weakly labeled.

01

Representation learning

Learning latent structure without labels

Self-supervised learning becomes most useful when the pretext objective is aligned with the structure of the downstream problem. My research explores contrastive objectives, masked information, and manifold-aware embedding spaces for state representation.

  • Geometry-aware target spaces
  • High-dimensional state representations
  • Evaluation through interpretable latent factors
02

Reinforcement learning

State, memory, and exploration

Agents rarely observe the true state of the world. I study representations for partial observability, mechanisms that preserve behavioral diversity, and transformer-based perception for online reinforcement learning.

  • Partially observable environments
  • Ensemble diversity and deep exploration
  • Visual transformer backbones for control
03

Brain & health AI

From general methods to consequential data

Neural and clinical data make representation quality especially important: labels are limited, recordings vary across people and devices, and useful features often depend on spatial or temporal relationships. My current work translates robust representation learning into medical and brain-data settings.

  • Self-supervision for neural signals
  • Robust spatial and temporal representations
  • Deep-learning applications for medical platforms

Collaboration

Interested in representation learning, reinforcement learning, or AI for neural data?

Start a conversation