Machine-learning algorithms
Self-supervised, reinforcement, and geometry-aware learning: objectives, architectures, and training schemes that uncover latent structure when labels are scarce and observations are incomplete.
Li MengPostdoctoral researcher, University of Oslo
Learning representations that preserve structure.
I develop machine-learning algorithms, from self-supervised and reinforcement learning to geometry-aware modeling, and apply them to complex biological signals: brain-computer interfaces, early-embryo development, and genomics.
Currently in the GENESIS convergence environment at the University of Oslo, with a staff page at the Institute of Basic Medical Sciences.
How can a model learn representations that remain useful when observations are incomplete, labels are scarce, and the underlying structure matters?
The question that connects the three directions below.
Three connected directions
Research overviewBrain-computer interfaces
Representation learning for EEG and intracranial neural signals that stays robust across people, sessions, and devices, from speech decoding for implantable BCIs to cognitive-health applications.
Embryo and genomics
Deep learning for early embryogenesis in the GENESIS convergence environment: single-cell transcriptomics, spatial gene expression, epigenetic signatures, and live microscopy.
Selected publications
All publications2024
State Representation Learning Using an Unbalanced Atlas
A manifold-based self-supervised learning paradigm that scales state representations to high-dimensional target spaces.
2024
Maximum Manifold Capacity Representations in State Representation Learning
Extends manifold-aware state representation learning with a nuclear-norm objective and a more efficient route to manifold-capacity regularization.
2023
Unsupervised State Representation Learning in Partially Observable Atari Games
Develops an unsupervised state representation scheme for partially observable environments and evaluates it on Atari-based benchmarks.
2022
Improving the Diversity of Bootstrapped DQN by Replacing Priors With Noise
Replaces fixed randomized priors with Gaussian noise to sustain ensemble diversity and improve deep exploration in Atari games.
Now and recent
Led the speech-decoding experiment at implantable-BCI company Neuralicorn (西湖灵犀) as algorithm lead, from experimental design and code framework to clinical data collection in hospital.
Joined eBRAIN-Health at the University of Oslo, developing deep-learning applications for medical platforms.
Joined the AI-MIND project at OsloMet, working on deep learning for early dementia detection.
Presented “State Representation Learning Using an Unbalanced Atlas” at ICLR 2024.
Writing
All writing- Exploration needs disagreement
Bootstrapped DQN explores well only while its heads disagree. A little zero-mean noise in each head's training target helps keep them apart.
- What should a representation remember when the world is partially observable?
Hide part of each frame from the encoder, ask it to recognise the next full frame anyway, and it learns to infer what it cannot see.
- Letting the atlas be unbalanced
A manifold-shaped representation only started to pay off at scale once we stopped forcing every chart to be used equally.