research
ICML 2026
Revisiting the Platonic Representation Hypothesis: An Aristotelian View
What do neural networks actually agree on once you remove the metric's own bias? A walk through the width and depth confounders behind representational similarity, and a refinement of the Platonic story to a smaller, sharper claim about shared local neighborhoods.
- representation learning
- similarity metrics
- multimodal
NeurIPS 2026
The Practical Side of Representation Geometry
Two models often agree on their local representation geometry, and that agreement is cheap to measure and predicts how well they combine. COMPASS uses it to decide whether to join a pair. PORTAL builds the connection in closed form, replacing the projector training stage.
- representation learning
- multimodal alignment
- model merging
- efficiency
NeurIPS 2024 ICASSP 2026 DMLR 2026 ML4H 2023
Representation-Based Data Quality Audits
A learned representation carries traces of the data that trained it, beyond what the labels reveal. We turn those traces into an auditor that flags off-topic samples, near-duplicates, and label errors, with no clean reference data and no per-issue training.
- data-centric ML
- self-supervised learning
- medical imaging
- audio
Under review
Auditing a Scientific Field through its Representations
If we know when to trust representation geometry, the same geometry can audit not just a single dataset but an entire scientific archive. We aggregate 1.1 million dermatology images from 29 public datasets into one shared space and quantify the field's demographic skew, novelty plateaus, and structural voids from that geometry.
- representation learning
- field-scale audit
- medical imaging
- data-centric ML