Representational Similarity Analysis (RSA) in rMVPA

Introduction to Representational Similarity Analysis

Representational Similarity Analysis (RSA) compares neural activity patterns with computational models by measuring pattern similarities, matching them against model predictions, and quantifying how well models explain the neural data. The rMVPA package implements this technique for neuroimaging analysis.

Basic Concepts

Dissimilarity Matrices

A dissimilarity matrix represents pairwise differences between conditions or stimuli. In RSA: - Each cell (i,j) represents how different conditions i and j are - Can be derived from neural data or theoretical models - Common measures: correlation distance (1 - correlation), Euclidean distance

RSA Workflow in rMVPA

Create an MVPA dataset
Define your model dissimilarity matrices
Create an RSA design
Build and run the RSA model

Step-by-Step Example

1. Creating Sample Data

Let’s create a simple example dataset with known structure:

# Generate a sample dataset (20x20x8 volume, 80 observations, 4 blocks)
dataset <- rMVPA::gen_sample_dataset(D=c(20,20,8), nobs = 80, blocks=4)

2. Creating Model Dissimilarity Matrices

You can use different types of dissimilarity matrices:

# Method 1: Using dist() on feature vectors
model_features <- matrix(rnorm(80*10), 80, 10)  # 80 trials, 10 features
model_rdm <- dist(model_features)  # Default is Euclidean distance

# Method 2: Direct correlation distance matrix
model_matrix <- 1 - cor(t(model_features))  # Correlation distance

3. Creating an RSA Design

The RSA design specifies how to compare neural and model dissimilarity patterns:

# Basic design with one model RDM
basic_design <- rsa_design(
  formula = ~ model_rdm,
  data = list(model_rdm = model_rdm),
  block_var = factor(dataset$design$block_var)
)

# Design with multiple model RDMs
model_rdm2 <- dist(matrix(rnorm(80*10), 80, 10))
complex_design <- rsa_design(
  formula = ~ model_rdm + model_rdm2,
  data = list(
    model_rdm = model_rdm,
    model_rdm2 = model_rdm2
  ),
  block_var = factor(dataset$design$block_var),
  keep_intra_run = FALSE  # Exclude within-run comparisons
)

4. Creating and Running an RSA Model

The rsa_model() function supports different methods for computing neural dissimilarities and analyzing relationships:

# Create MVPA dataset
dset <- mvpa_dataset(dataset$dataset$train_data, mask=dataset$dataset$mask)

# Create RSA model with different options
rsa_spearman <- rsa_model(
  dataset = dset,
  design = basic_design,
  distmethod = "spearman",  # Method for computing neural dissimilarities
  regtype = "spearman"      # Method for comparing neural and model RDMs
)

# Run searchlight analysis
results <- run_searchlight(
  rsa_spearman,
  radius = 4,
  method = "standard"
)

## function (model_spec, good_results, bad_results = NULL)  
## [1] "combiner is "
## INFO [2025-03-20 21:31:20] Running standard searchlight with radius = 4
## INFO [2025-03-20 21:31:20] creating standard searchlight
## INFO [2025-03-20 21:31:20] running standard searchlight iterator
## INFO [2025-03-20 21:31:20] ⚡ Processing batch 1/11
## INFO [2025-03-20 21:31:21] ⚡ Processing batch 2/11
## INFO [2025-03-20 21:31:21] ⚡ Processing batch 3/11
## INFO [2025-03-20 21:31:22] ⚡ Processing batch 4/11
## INFO [2025-03-20 21:31:23] ⚡ Processing batch 5/11
## INFO [2025-03-20 21:31:23] ⚡ Processing batch 6/11
## INFO [2025-03-20 21:31:24] ⚡ Processing batch 7/11
## INFO [2025-03-20 21:31:25] ⚡ Processing batch 8/11
## INFO [2025-03-20 21:31:25] ⚡ Processing batch 9/11
## INFO [2025-03-20 21:31:26] ⚡ Processing batch 10/11
## INFO [2025-03-20 21:31:27] ⚡ Processing batch 11/11
## INFO [2025-03-20 21:31:27] 
## ✨ MVPA Iteration Complete
## ├─ Total ROIs: 256
## ├─ Processed: 256
## └─ Skipped: 0

Advanced Features

Multiple Comparison Methods

rMVPA supports several methods for comparing neural and model RDMs:

# Pearson correlation
rsa_pearson <- rsa_model(dset, basic_design, 
                        distmethod = "pearson", 
                        regtype = "pearson")

# Linear regression
rsa_lm <- rsa_model(dset, basic_design, 
                    distmethod = "spearman", 
                    regtype = "lm")

# Rank-based regression
rsa_rfit <- rsa_model(dset, basic_design, 
                      distmethod = "spearman", 
                      regtype = "rfit")

Handling Run Structure

RSA can account for the run/block structure of fMRI data. A critical consideration in fMRI analysis is whether to include comparisons between patterns from the same run.

Understanding keep_intra_run

The keep_intra_run = FALSE parameter tells RSA to exclude comparisons between patterns within the same run/block. This is important because:

Temporal Autocorrelation: BOLD responses within the same run are temporally autocorrelated
Scanner Drift: Within-run patterns may share scanner drift effects
Physiological Noise: Within-run patterns may share structured noise from breathing, heart rate, etc.

Here’s a visualization of what keep_intra_run = FALSE does:

# Create a small example with 2 runs, 4 trials each
mini_data <- matrix(1:8, ncol=1)  # Trial numbers 1-8
run_labels <- c(1,1,1,1, 2,2,2,2)  # Two runs with 4 trials each

# Create distance matrix
d <- dist(mini_data)
d_mat <- as.matrix(d)

# Show which comparisons are kept (TRUE) or excluded (FALSE)
comparison_matrix <- outer(run_labels, run_labels, "!=")
# Only show lower triangle to match distance matrix structure
comparison_matrix[upper.tri(comparison_matrix)] <- NA

# Display the matrices
cat("Trial numbers:\n")

## Trial numbers:

print(matrix(1:8, nrow=8, ncol=8)[lower.tri(matrix(1:8, 8, 8))])

##  [1] 2 3 4 5 6 7 8 3 4 5 6 7 8 4 5 6 7 8 5 6 7 8 6 7 8 7 8 8

cat("\nRun comparisons (TRUE = across-run, FALSE = within-run):\n")

## 
## Run comparisons (TRUE = across-run, FALSE = within-run):

print(comparison_matrix[lower.tri(comparison_matrix)])

##  [1] FALSE FALSE FALSE  TRUE  TRUE  TRUE  TRUE FALSE FALSE  TRUE  TRUE  TRUE
## [13]  TRUE FALSE  TRUE  TRUE  TRUE  TRUE  TRUE  TRUE  TRUE  TRUE FALSE FALSE
## [25] FALSE FALSE FALSE FALSE

When we create an RSA design with keep_intra_run = FALSE:

# Create design excluding within-run comparisons
blocked_design <- rsa_design(
  formula = ~ model_rdm,
  data = list(model_rdm = model_rdm),
  block_var = factor(dataset$design$block_var),
  keep_intra_run = FALSE  # Exclude within-run comparisons
)

This creates an RSA design where: - Comparisons between patterns from different runs are included - Comparisons between patterns within the same run are excluded - The analysis focuses on more reliable between-run pattern similarities

When to Use keep_intra_run = FALSE

You should consider setting keep_intra_run = FALSE when: - Your experiment has multiple runs/blocks - You want to control for temporal autocorrelation - You want to minimize the impact of run-specific noise - You’re following conservative analysis practices

Setting keep_intra_run = TRUE (default) might be appropriate when: - You have very few runs and need more samples - Your runs are very short - You’ve carefully controlled for temporal confounds - You’re doing exploratory analyses

Visualizing Results

You can examine and visualize the RSA results:

# Get range of correlation values
range(results$results$model_rdm$data)

## [1] -0.03065412  0.02820032

# Basic summary of results
print(results)

## 
##  █▀▀ Searchlight Analysis Results ▀▀█ 
## 
## ├─ Coverage 
## │  ├─ Total Voxels:  3,200 
## │  └─ Active Voxels:  256 
## └─ Performance Metrics 
##    ├─  model_rdm 
##    │  ├─ Mean:  -0.0070 
##    │  ├─ SD:  0.0070 
##    │  ├─ Min:  -0.0307 
##    │  └─ Max:  0.0282

# Save results (commented out)
# write_vol(results$model_rdm, "RSA_results.nii.gz")

Summary

The rMVPA package provides a comprehensive RSA implementation with flexible model specification, multiple dissimilarity computation methods, and support for complex experimental designs with run/block structures. It integrates seamlessly with searchlight analysis and offers various statistical approaches including correlation, regression, and rank-based methods.

When using RSA in rMVPA, carefully consider your experimental design when setting block variables and intra-run parameters, choose distance methods that match your theoretical framework, and select statistical approaches appropriate for your analysis goals.

References

For more information on RSA: - Kriegeskorte et al. (2008). Representational similarity analysis - connecting the branches of systems neuroscience. Front Syst Neurosci. - Nili et al. (2014). A toolbox for representational similarity analysis. PLoS Comput Biol.

Bradley Buchsbaum

2025-03-20