Multi-Level Optimal Transport Aligns Neural Networks with Brain Regions
A novel framework known as Multi-Level Optimal Transport (MOT) has been introduced by researchers to synchronize representations across layers of artificial neural networks and between these networks and biological brains. Unlike traditional techniques that align layers separately, MOT simultaneously determines soft, globally consistent connections between layers and neuron-level transport strategies, enabling source neurons to allocate mass to various target layers. This method provides a unified alignment score for comprehensive network comparisons and effectively addresses depth discrepancies via mass distribution. MOT was tested on vision models, large language models, and recordings from the human visual cortex. It overcomes the shortcomings of conventional representational similarity methods, which often yield asymmetric outcomes, lack global alignment scores, and face challenges with networks of varying depths. The research can be found on arXiv with the identifier 2510.01706.
Key facts
- MOT is a unified framework for representational alignment.
- It infers soft, globally consistent layer-to-layer couplings.
- Neuron-level transport plans are jointly inferred.
- Source neurons can distribute mass across multiple target layers.
- Total transport cost is minimized under marginal constraints.
- MOT produces a single alignment score for entire network comparisons.
- It handles depth mismatches through mass distribution.
- Evaluated on vision models, LLMs, and human visual cortex recordings.
Entities
Institutions
- arXiv