Skip to contents

Naive Cross-Decoding

rMVPA authors

2026-05-08

Source: vignettes/Naive_Cross_Decoding.Rmd
Naive_Cross_Decoding.Rmd

ReNA / REMAP terminology

This section of the documentation covers four related models for cross-domain representational analysis. They share a vocabulary that’s easier to keep straight if defined once:

Reach for naive cross-decoding first: it’s the cheapest baseline. The other four answer different questions when naive transfer is insufficient.

Overview

You use cross-decoding when the data come from two related domains, such as perception and memory, and you want to know whether a pattern learned in the source domain transfers to the target domain. Naive cross-decoding is the direct-transfer baseline: it classifies target-domain patterns by correlating them with source-domain prototypes, with no learned correction between the two domains.

This makes the analysis deliberately conservative. If it works, the representations are already aligned well enough to transfer directly. If it does not, the result is still useful because it tells you how much work a domain-adaptive model such as remap_rrr_model() or an encoding-retrieval partition analysis has to explain.

Algorithm:

  1. Compute a prototype (mean pattern) for each category in the training domain.
  2. Correlate every test pattern with each prototype.
  3. Assign the category whose prototype has the highest correlation.

Correlations are passed through softmax to produce pseudo-probabilities.

Quick example 

set.seed(42)
data_info <- gen_sample_dataset(
  D = c(10, 10, 10), nobs = 80, nlevels = 4,
  blocks = 4, external_test = TRUE
)

# Three ROIs
roi_mask <- NeuroVol(
  sample(1:3, length(data_info$dataset$mask), replace = TRUE),
  space(data_info$dataset$mask)
)

model <- naive_xdec_model(
  dataset = data_info$dataset,
  design  = data_info$design,
  return_predictions = TRUE
)

print(model)
#> Naive Cross-Decoding Model
#>   link_by:             (class labels) 
#>   training levels:     a, b, c, d 
#>   n_train:             80 
#>   n_test:              80 
#>   return_predictions:  TRUE
results <- run_regional(model, roi_mask)
metric_cols <- intersect(
  c("roinum", "Accuracy", "AUC"),
  names(results$performance_table)
)
results$performance_table[, metric_cols, drop = FALSE]
#> # A tibble: 3 × 3
#>   roinum Accuracy     AUC
#>    <int>    <dbl>   <dbl>
#> 1      1    0.2   -0.0292
#> 2      2    0.238 -0.0562
#> 3      3    0.212  0.0558
chance <- 1 / nlevels(data_info$design$y_test)
cat(sprintf("Chance level: %.0f%%\n", chance * 100))
#> Chance level: 25%

With random data, accuracy hovers around chance. With real fMRI data, above-chance accuracy indicates that category representations transfer across domains.

By default (link_by = NULL), prototypes are keyed by class label: all training trials for category A are averaged into one prototype, and each test trial is classified against these category prototypes.

When the same individual items appear in both domains (e.g., the same images during perception and recall), you can form item-level prototypes instead:

model <- naive_xdec_model(
  dataset = dataset,
  design  = design,
  link_by = "item_id"
)

This creates one prototype per item rather than per category, giving finer-grained matching when item identity is meaningful.

How does this relate to ERA partitioning?

Naive cross-decoding answers a first-order question: does each target pattern look most like the correct source prototype? That is often the first result to check because it maps directly onto classification accuracy.

The era_partition_model() analysis uses the same direct-transfer idea as one component of a broader encoding-retrieval analysis. It builds matched source and target prototypes for the item key, then reports:

  • direct transfer metrics such as naive_top1_acc and xdec_Accuracy,
  • first-order variance partitioning via first_order_delta_r2,
  • second-order geometry preservation via second_order_delta_r2 and geom_cor, and
  • optional leakage-free Procrustes alignment metrics.

Use naive_xdec_model() when you need a clean baseline map or ROI table for direct transfer. Use era_partition_model() when your scientific question is whether encoding-retrieval transfer reflects item-specific similarity, preserved representational geometry, or nuisance structure such as block, category, or temporal lag.

Searchlight analysis

Naive cross-decoding plugs directly into run_searchlight():

model <- naive_xdec_model(dataset = dataset, design = design)
sl_results <- run_searchlight(model, radius = 3)

The resulting accuracy map shows where category structure transfers across domains, voxel by voxel.

Troubleshooting

“requires external test set” — Naive cross-decoding needs separate train and test data. Provide external_test = TRUE in gen_sample_dataset(), or construct your mvpa_dataset() with both train_data and test_data.

All predictions are the same class — Prototypes may be too similar. Check inter-prototype correlations; if all are above ~0.9, the training domain does not carry enough discriminative information for the method to work.

Chance-level performance everywhere — This is the expected outcome when there is a large systematic domain shift. Consider domain-adaptive models such as remap_rrr_model() (see vignette("REMAP_RRR")).

Next steps