Research map / Learning theory and algorithms / Optimal transport
Representational Alignment Across Model Layers and Brain Regions with Hierarchical Optimal Transport
Shaan Shah, Meenakshi Khosla
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use global soft layer coupling (MOT) instead of independent pairwise best-layer matching | MOT · LLM alignment (Llama-3.2 1B vs 3B) | held-out reconstruction correlation: 0.505 -> 0.558 | Improved |
| use MOT instead of pairwise best OT | MOT · NSD cortex alignment (Subject A vs B) | held-out reconstruction correlation: 0.245 -> 0.244 | No clear effect |
| use vanilla MOT instead of pairwise best OT | MOT · Vision models (ViT-MAE Base vs Huge) | held-out reconstruction correlation: 0.417 -> 0.149 | Hurt |
| use vanilla MOT instead of pairwise best OT | MOT · Vision models (ViT-MAE Base vs DINOv2 Large) | held-out reconstruction correlation: 0.624 -> 0.577 | Improved |
| add rotation (MOT+R) to MOT (ablation) | MOT · Vision models (DINOv2 Small vs Large) | held-out reconstruction correlation: 0.353 -> 0.778 | Improved |
| use MOT+R instead of rotation-aware pairwise best | MOT+R · Vision models (DINOv2 Small vs Giant) | held-out reconstruction correlation: 0.418 -> 0.790 | Improved |
| keep soft layer coupling instead of top-1 hard assignment (ablation) | MOT · LLM alignment (Llama-3.2 1B vs 3B) | held-out reconstruction correlation: 0.502 -> 0.558 | Improved |
| optimize layer coupling instead of random permutation of P (ablation) | MOT · LLM alignment (Llama-3.2 1B vs 3B) | held-out reconstruction correlation: 0.510 -> 0.558 | Improved |
| optimize layer coupling instead of random permutation of P (ablation) | MOT · NSD cortex alignment (Subject A vs B) | held-out reconstruction correlation: 0.135 -> 0.244 | Improved |
| use layer-wise MOT instead of flattening all neurons into one global OT (ablation) | MOT · NSD cortex alignment (Subject A vs B) | held-out reconstruction correlation: 0.184 -> 0.244 | Improved |
| use MOT instead of global linear predictivity | MOT · NSD cortex alignment (Subject A vs B) | held-out reconstruction correlation: 0.305 -> 0.244 | Hurt |
| use MOT instead of pairwise linear predictivity | MOT · LLM alignment (Llama-3.2 1B vs 3B) | held-out reconstruction correlation: 0.737 -> 0.558 | Hurt |
Closest papers by meaning
- Bridging Critical Gaps in Convergent Learning: How Representational Alignment Evolves Across Layers, Training, and Distribution Shifts (NeurIPS 2025 · similarity 0.89)
- Sparse components distinguish visual pathways & their alignment to neural networks (ICLR 2025 · similarity 0.88)
- Tracing Representation Progression: Analyzing and Enhancing Layer-Wise Similarity (ICLR 2025 · similarity 0.87)
- Scaling Laws for Task-Optimized Models of the Primate Visual Ventral Stream (ICML 2025 · similarity 0.87)
- Hierarchical Multi-Marginal Optimal Transport for Network Alignment (AAAI 2024 · similarity 0.87)
- How Not to Stitch Representations to Measure Similarity: Task Loss Matching Versus Direct Matching (AAAI 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
