Evaluation
Representational Alignment
2023ActivePublished: 29 September 2026Updated: 29 September 2026Published
Key
innovation
The degree to which the internal representations of two systems (two models, or a model and a brain) encode information similarly — measured by representation-similarity metrics.
Category
Evaluation
Abstraction level
Paradigm
Operation level
Evaluation (runtime)Model
Use cases
Comparing models to each otherNeuroAI: model–brain comparisonsAssessing human-likeness of representationsDiagnosing transfer and feature similarity
How it works
One collects both systems representations for the same stimulus set and compares them with a metric: RSA compares (dis)similarity matrices across stimuli; CKA measures space similarity with invariance to rotation and scale; linear (encoding) regression tests how well one representation predicts the other (neural predictivity). The output is an alignment score, usually in [0,1].
Problem solved
One needs to compare how different models (or a model and a brain) represent information — without this it is hard to speak of shared mechanisms or human-like processing.
Components
RSAMetric
Comparing representational (dis)similarity matrices across stimuli.
CKAMetric
Centered kernel alignment — space similarity robust to rotation/scale.
Regression/encoding (predictivity)Metric
Predicting one representation from another.
Evolution
2014
RSA and model–brain comparisons in computational neuroscience
2019
CKA popularised as a representation-similarity metric in ML
2023
Consolidation of representational alignment as a shared NeuroAI language
Inflection point