Source-linked AI summary
Mechanistic Interpretability Reveals Shared Causal Subspaces in Brain-to-Speech Decoders
Maryam Maghsoudi, Ayushi Mishra, Sanghamitra Dutta
TL;DR
Covert speech is harder to decode, and how decoders internally process different speech conditions remains unclear. The study uses same-sentence activation patching to examine cross-condition transfer, finding that it relies on neuron groups and that recurrent representations overlap across source modes. It suggests shared representations, especially those learned from vocalized speech, as a route toward data-efficient covert decoding.
Problem
Covert speech is harder to decode than vocalized speech, while little is known about how decoders internally process different speech conditions.
Source: §1 · p. 1Method
The study patches ranked neuron groups between paired same-sentence source and target conditions in a decoder trained on the target mode.
Source: §2 · pp. 2–3Results
Takeaways & Limitations
Shared representations, especially those learned from vocalized speech, could support data-efficient covert decoding through objectives that encourage cross-modal transfer.
Source: §1 · p. 2, §4 · p. 4The study covers one participant and one Conv1D–BiGRU architecture; generalization across participants and language-model-based decoders remains for future work.
Source: §4 · p. 4
Abstract
from arXiv · showhide
Decoding covert speech, such as mimed or imagined, from brain activity is harder than decoding vocalized speech. Cross-modal transfer, where information from one speech form helps decode another, is a promising remedy; yet how a decoder internally represents and processes brain activity from different speech forms remains unclear. In this work, we ask: which internal neurons of a decoder carry cross-modal information, and are these neurons shared across different speech forms? To answer these questions, we leverage mechanistic interpretability, using recordings of the same sentences in vocalized, mimed, and imagined input pairs for activation patching. We insert the decoder's internal activity for a sentence in one condition into its processing of the same sentence in another and measure the change in decoding accuracy. We find that no single neuron drives this benefit; instead, it arises from small groups of neurons, with vocalized speech as the most useful source. These groups are largely condition-specific in the early stage of the decoder but overlap in the later stage. These findings point toward more data-efficient covert speech decoders through training objectives that encourage shared later-stage representations learned mainly from vocalized data.
1. INTRODUCTION
Covert speech is harder to decode than vocalized speech, and little is known about how decoders internally process different speech conditions. Using same-sentence activation patching, the study finds that transfer depends on groups of neurons and that vocalized activity is especially useful for covert decoding.
1. INTRODUCTION
- Covert speech decoding remains harder because its neural responses are weaker, noisier, and harder to collect in large amounts.Prior work found shared structure across vocalized, mimed, and imagined responses and showed that decoders can transfer between conditions.Source: §1 · p. 1
- The study uses activation patching to test whether individual neurons or groups carry cross-condition information and whether the same neurons recur across sentences and source conditions.It replaces selected internal activity from one condition with activity from the same sentence in another and measures the decoding change.Source: §1 · p. 1
- VocalMind provides sEEG recordings of the same sentences in vocalized, mimed, and imagined forms, allowing condition-related activity differences to be examined with sentence identity held constant.The analyses target an early stage extracting short-term features and a later stage integrating information over longer spans.Source: §1 · p. 1
- Cross-condition transfer comes from small neuron groups rather than single neurons; vocalized activity is most useful for mimed and imagined speech, with early neurons condition-specific and later neurons reused.The authors suggest shared later-stage representations, learned mainly from vocalized data, as a route toward more data-efficient covert decoders.Source: §1 · pp. 1–2
2. PRELIMINARIES AND APPROACH
The approach compares same-sentence neural activity across speech conditions within a decoder trained on the target condition. It patches source activity into target processing, then tests group effects and neuron reuse across samples and source modes.
2. PRELIMINARIES AND APPROACH
- Paired inputs are recordings of the same sentence in source and target modes, each passed through the decoder trained on the target mode.The setup defines source and target activations at a selected layer and compares the patched prediction with the unpatched target-mode baseline.Source: §2 · p. 2
- Activation patching replaces selected target activation trajectories with corresponding source activations, continues the target-mode forward pass, and measures the change in Pearson correlation.Positive ∆PCC means the source activity improves target-mode decoding.Source: §2 · p. 2
- In Vocalized→Mimed patching, single-neuron effects cluster near zero and are more broadly distributed in the convolutional layer than in the RNN.The distribution is pooled across neurons and folds.Source: Figure 2 · p. 2
- For mimed and imagined targets, top-k curves rise and then degrade; vocalized targets show no improvement, while vocalized-source patching yields the largest covert-target gains.The curves plot change relative to k=1 against the number of patched neurons, with RNN and convolutional layers distinguished.Source: Figure 3 · p. 3
- The decoder uses a Conv1D encoder, a three-layer bidirectional GRU, and a linear mel-spectrogram predictor; analyses use six-fold cross-validation on held-out samples.The convolutional output is the early site and the recurrent output the later site; folds aggregate mean and standard deviation of ∆PCC.Source: §2 · p. 3
- Top-k patching ranks neurons by mean single-neuron effect and jointly patches the k highest-ranked units; a peak followed by degradation can indicate interference from added neurons.The study also identifies each sample’s strongest individual-patching neuron and measures reuse through winner counts and top-k coverage.Source: §2 · p. 3
- To compare neuron reuse across source modes for a fixed target, the approach ranks neurons by winner count and evaluates top-k set overlap against a permutation null.The null draws two neuron sets repeatedly and reports its mean and 95% confidence interval, accounting for layer dimensionality.Source: §2 · p. 3
3. RESULTS
Cross-modal effects are weak at the single-neuron level but emerge from groups, with recurrent neurons showing substantial overlap across source modes. Convolutional contributions are more source-specific, while recurrent neurons are reused across samples and conditions.
3. RESULTS
- Single-neuron patching effects average around 10^-4 in mean ∆PCC, and no individual neuron separates from the population.The positive-effect fraction varies by transfer direction, reaching 0.94 for Vocalized→Mimed at the RNN layer.Source: §3 · pp. 3–4
- RNN top-k neuron overlap is roughly 5–8× chance across targets, while convolutional overlap is at or below chance.Mimed targets reach ∼0.8 overlap at small k; a causal subspace is defined as activation directions that directly affect model output.Source: Figure 4 · p. 4, §3 · p. 4
- For Mimed and Imagined targets, patching initially improves performance before larger neuron groups degrade it; Vocalized targets show no improvement.Convolutional curves generally peak at larger k than RNN curves, suggesting more distributed convolutional contributions and more compact recurrent subsets.Source: §3 · p. 4
- Influential RNN neurons are reused across samples more consistently than convolutional neurons across source–target pairs (p < 0.05).For Vocalized→Mimed at the RNN layer, 69 unique winners span 200 samples, and the top five account for 35% of samples.Source: §3 · p. 4
4. DISCUSSION AND LIMITATIONS
Cross-modal transfer is carried by small neuron groups, with convolutional representations more source- and sample-specific than recurrent representations. Mimed speech’s intermediate position may explain strong Vocalized→Mimed transfer and high source-mode overlap when mimed speech is the target.
4. DISCUSSION AND LIMITATIONS
- Cross-modal transfer is carried by small neuron groups, while recurrent representations contain a compact shared causal subspace reused across source modes and samples.Convolutional representations are more source- and sample-specific.Source: §4 · p. 4
- Mimed speech’s intermediate position may explain strong Vocalized→Mimed transfer and high source-mode overlap when mimed speech is the target.Source: §4 · p. 4
- Shared representations, especially those learned from vocalized speech, could support data-efficient covert speech decoding through objectives that encourage cross-modal transfer.Source: §4 · p. 4
- The study is limited to one participant and one Conv1D–BiGRU architecture, so generalization across participants and language-model-based decoders remains to be tested.The controlled setting isolates cross-modal representations; future work should also relate causal subspaces to electrode anatomy.Source: §4 · p. 4