| name | brain-to-language-source-attribution |
| title | What Are We Actually Decoding? Source Attribution for Non-Invasive Brain-to-Language Retrieval |
| description | Source attribution framework for MEG-to-audio brain decoding that separates decoding performance into structural shortcuts, stimulus-evoked evidence, and contextual aggregation |
| tags | ["brain-decoding","source-attribution","meg","language-decoding","gcb","group-context-bias","neuroscience","computational-neuroscience"] |
| created | 2026-05-26T00:00:00.000Z |
What Are We Actually Decoding? Source Attribution for Non-Invasive Brain-to-Language Retrieval
arXiv: 2605.24524
Authors: Xinyu Zhang, Sichao Liu, Runhao Lu, Alexandra Woolgar, Lihui Wang
Subjects: Machine Learning (cs.LG); Computation and Language (cs.CL); Neurons and Cognition (q-bio.NC)
Summary
This paper addresses a critical methodological challenge in non-invasive neural language decoding: how to properly attribute decoding performance to neural sources versus confounding factors. The authors recast stimulus-locked MEG-to-audio retrieval as an auditing framework that separates apparent performance into three sources:
- Structural shortcuts — non-neural nuisances like signal duration that can inflate results
- Window-level stimulus-locked evidence — genuine neural evidence at the individual time-window level
- Cross-window contextual aggregation — integration of evidence across multiple windows
Key Contributions
-
Structured Auditing Framework: Separates brain-to-language decoding performance into three distinct sources with diagnostic tests for each
-
Structural Leakage Isolation: Signal-blind Gaussian noise reaches 66.3% Rank@1 under variable-length decoding but collapses to near chance once fixed-duration windows and stimulus-identity splits are enforced — demonstrating that variable-length decoding creates a structural shortcut
-
Group Context Bias (GCB): An inference-time additive logit bias that pools sentence-consistent evidence across windows, making the contextual source measurable:
- R@1 shifts from 44% to 52% on Gwilliams dataset
- R@1 shifts from 22% to 29% on MOUS dataset
- Effect collapses under random-grouping perturbations
- Vanishes when local evidence is attenuated in MEG or near chance in EEG
-
Oracle Diagnostic: 95.7% of Top-1 errors select the wrong sentence, localising the residual bottleneck to sentence-level competition
Methodology
- Framework: Stimulus-locked MEG-to-audio retrieval with controlled source attribution
- Controls: Fixed-duration windows, stimulus-identity splits, signal-blind baselines
- Intervention: Group Context Bias (GCB) — an auditable score-space intervention
- Validation: Two datasets (Gwilliams, MOUS) with MEG and EEG modalities
Implications
- Brain-to-language decoding performance should be source-attributed, not merely reported
- Variable-length decoding without proper controls overestimates neural evidence
- GCB provides a principled way to measure contextual aggregation effects
- Framework applicable to other neural decoding tasks beyond language
Activation Keywords
brain-to-language-source-attribution, meg-audio-retrieval, structural-shortcut-detection, group-context-bias, neural-decoding-evaluation, source-attribution-framework, brain-decoding-methodology