Getting aligned on representational alignment

TL;DR

The paper identifies representational alignment as a cross-disciplinary core concern in cognitive science, neuroscience, and machine learning, and argues that fragmented terminology hinders progress. It proposes a unifying framework that decomposes studies into data, systems, measurements, embeddings, and an alignment function, with three objectives: measure, bridge, and increase alignment. Through a broad survey, the authors illustrate how this framework maps across domains, including RSA/CKA-based measures, hyperalignment, and model-to-model or human-model mappings. They discuss open problems—from data selection and stimulus design to probing black-box systems and potential risks—aiming to catalyze cross-disciplinary collaboration and accelerate progress in building more aligned information processing systems.

Abstract

Biological and artificial information processing systems form representations of the world that they can use to categorize, reason, plan, navigate, and make decisions. How can we measure the similarity between the representations formed by these diverse systems? Do similarities in representations then translate into similar behavior? If so, then how can a system's representations be modified to better match those of another system? These questions pertaining to the study of representational alignment are at the heart of some of the most promising research areas in contemporary cognitive science, neuroscience, and machine learning. In this Perspective, we survey the exciting recent developments in representational alignment research in the fields of cognitive science, neuroscience, and machine learning. Despite their overlapping interests, there is limited knowledge transfer between these fields, so work in one field ends up duplicated in another, and useful innovations are not shared effectively. To improve communication, we propose a unifying framework that can serve as a common language for research on representational alignment, and map several streams of existing work across fields within our framework. We also lay out open problems in representational alignment where progress can benefit all three of these fields. We hope that this paper will catalyze cross-disciplinary collaboration and accelerate progress for all communities studying and developing information processing systems.

Figures (2)

• Figure 1: Examples of contemporary representational alignment research in cognitive science, neuroscience, and machine learning. We discuss three types of representational alignment research. Measuring representational alignment aims to measure the degree of alignment between two systems as a dependent measure in an experiment (a. jacoby2017integer, b. kriegeskorte2008matching, c. pmlr-v97-kornblith19a). Bridging representational spaces aims to bring representations into a shared space to facilitate some downstream application (d. hebart2020revealing, e. o2018predicting, f. gupta2017aligned). Increasing representational alignment aims to update the internal representations or measurements of one system to increase its alignment with another system (g. muttenthaler2023improving, h. khosla2022high, i. tian2019contrastive). (Reproduced with permission from the cited papers.)
• Figure 2: A general framework for conducting and describing representational alignment research. Most studies of representational alignment involve five components that researchers can control: data is presented via interfaces to the two systems. The systems form internal representations of the data and researchers take measurements of the systems and map them to some embedding space to try to infer the representations. An alignment function is then applied to those inferred representations to compute a single alignment score. These studies typically have one of three objectives: measuring the representational alignment between the two systems (i.e., the alignment score), bridging between two different representational spaces by finding a shared embedding space, or increasing the representational alignment between the two systems either by updating their internal representations (e.g., via learning) or how they are measured.