Representational Similarity Analysis
Representational Similarity Analysis (RSA) · Also known as: RSA, representational geometry, similarity structure analysis
Representational Similarity Analysis (RSA) is a framework for comparing representational geometry across brain regions, computational models, and behavioral measures. Introduced by Kriegeskorte and colleagues in 2008, RSA measures how similarly a brain region represents different stimuli or concepts by examining pairwise similarity structure rather than absolute activity patterns.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
RSA is ideal for comparing representational geometry across brains and models, when interest is in abstract structure rather than absolute activity, and when many conditions are available. Use RSA to test whether brain and model share organizational principles. Avoid RSA if localizing activation is the primary goal (use univariate fMRI) or sample sizes are very small (RDM estimates become noisy).
Strengths & limitations
- Reveals abstract representational structure independent of activity magnitude or anatomical alignment
- Bridges brain, behavior, and computation—directly compares representations across modalities
- Model-agnostic; can compare to any source (another brain region, neural network, behavioral judgments)
- Robust to some preprocessing artifacts (e.g., global scaling) that affect univariate and MVPA approaches
- Requires many conditions (typically 8–100) for reliable RDM estimates; low-condition studies produce unstable estimates
- Rank correlation (Spearman rho) used for inference; limited statistical power, especially with few conditions
- Interpretation of representational dissimilarity structure can be abstract; unclear what specific representational features drive effects
- Symmetric RDMs collapse information about directionality; cannot distinguish A→B vs. B→A relationships
Frequently asked
What is a representational dissimilarity matrix (RDM)?
An RDM is a symmetric square matrix where each element is the dissimilarity (1 - correlation) between two conditions. Diagonal is zero (perfect similarity with self). High values indicate low pairwise correlation; low values indicate high correlation. RDM summarizes the overall geometry—how conditions cluster in representational space.
How many conditions do I need for RSA?
Minimum 8–10 conditions to get stable estimates. More is better; 20–100 conditions yield robust RDMs. With fewer conditions, sampling noise dominates. Conduct power simulations or use bootstrap confidence intervals to assess stability. Report number of conditions used and RDM reliability estimates.
How do I compare RSA results across regions or subjects?
Use rank correlation (Spearman rho) between RDMs from different sources (e.g., two brain regions, brain vs. model). Higher correlation indicates more similar representational structure. Construct statistical null distributions via permutation or bootstrap to establish significance; compare correlations with paired tests across subjects for group inference.
Can RSA replace MVPA?
No; they address different questions. MVPA asks 'can conditions be classified?' RSA asks 'does representational structure match a model?' MVPA is sensitive to overall separability; RSA is sensitive to relational structure. Use both when possible; they provide complementary information.
Sources
- Kriegeskorte, N., Mur, M., & Bandettini, P. A. (2008). Representational similarity analysis—connecting the branches of systems neuroscience. Frontiers in Systems Neuroscience, 2, 4. DOI: 10.3389/neuro.06.004.2008 ↗
- Nili, H., Wingfield, C., Walther, A., et al. (2014). Inferring population attitude towards candidates from social media and electoral history. PLOS ONE, 9(5), e95809. link ↗
How to cite this page
ScholarGate. (2026, June 3). Representational Similarity Analysis (RSA). ScholarGate. https://scholargate.app/en/neuroimaging/representational-similarity-analysis
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Dynamic Causal ModelingNeuroimaging↔ compare
- Graph Brain Network AnalysisNeuroimaging↔ compare
- Multivariate Pattern AnalysisNeuroimaging↔ compare