MEG-to-MEG Transfer Learning and Cross-Task Speech/Silence Detection with Limited Data
summary
The gist
Data-efficient neural decoding is a central challenge for speech brain-computer interfaces <ref:2602.18253#pg2> The paper presents the first demonstration of transfer learning and cross-task decoding…
In short
The study demonstrated that pre-training a Conformer model on 50 hours of single-subject listening data significantly improved performance when fine-tuned on minimal, subject-specific speech perception and production tasks. Transfer learning yielded consistent gains, especially in cross-task decoding between perception and production. This suggests learned neural representations capture shared speech processes rather than just task or subject specifics.
Key concepts
- Transfer Learning
- This technique involves training a machine learning model on a large dataset (pre-training) and then using that knowledge to quickly adapt the model to a new, smaller dataset (fine-tuning). In this study, the model learned general speech features from hours of listening data and applied those features to decode specific perception or production tasks with very little new data.
- Cross-Task Decoding
- This refers to using a neural model trained for one type of task (e.g., listening) to successfully perform a different, related task (e.g., speech production). The research showed that models pre-trained on listening could reliably decode speech production, proving that the underlying neural representations are shared across these different speech modalities.
- MEG-to-MEG Transfer Learning
- This is a specific application of transfer learning where the model is trained using Magnetoencephalography (MEG) data from one set of subjects or tasks and then applied to decoding tasks involving other subjects or different types of speech. The core finding is that large-scale listening data allows a model to generalize effectively to new, limited subject-specific fine-tuning data.
- Shared Neural Representations
- This concept posits that the brain uses common neural patterns for different speech functions, like perceiving sound and producing sound. The study found this because models trained on production could successfully decode passive listening above chance level, indicating that the learned features reflect fundamental neural speech processes rather than just motor planning or task-specific actions.
Terminology used across episodes
This episode discusses
- MEG-to-MEG Transfer Learning and Cross-Task Speech/Silence Detection with Limited Data · Paper Radio
- Scaling Law in Neural Data: Non-Invasive Speech Decoding with 175 Hours of EEG Data
- Decoding individual words from non-invasive brain recordings across 723 participants
- Scaling laws for decoding images from brain activity
- Transfer Learning from ImageNet for MEG-Based Decoding of Imagined Speech
- MEGConformer: Conformer-Based MEG Decoder for Robust Speech and Phoneme Classification · Paper Radio
- EEG Foundation Challenge: From Cross-Task to Cross-Subject EEG Decoding
- MEGState: Phoneme Decoding from Magnetoencephalography Signals
- Brain-to-Text Decoding: A Non-invasive Approach via Typing
The paper
MEG-to-MEG Transfer Learning and Cross-Task Speech/Silence Detection with Limited Data · Read on arXiv
Xabier de Zuazo, Vincenzo Verbeni, Eva Navas, Ibon Saratxaga, Mathieu Bourguignon, Nicola Molinaro
HiTZ Center, University of the Basque Country – UPV/EHU, Spain · Basque Center on Cognition, Brain and Language – BCBL, Spain · Ikerbasque, Basque Foundation for Science, Spain · Laboratory of Functional Anatomy, Faculty of Human Motor Sciences, Universite libre de Bruxelles (ULB), Belgium · Laboratoire de Neuroanatomie et Neuroimagerie translationnelles (LN2T), ULB Neuroscience Institute, Universite libre de Bruxelles (ULB), Belgium · WEL Research Institute, Belgium
Data-efficient neural decoding is a central challenge for speech brain-computer interfaces. We present the first demonstration of transfer learning and cross-task decoding for MEG-based speech models spanning perception and production. We pre-train a Conformer-based model on 50 hours of single-subject listening data and fine-tune on just 5 minutes per subject across 18 participants. Transfer learning yields consistent improvements, with in-task accuracy gains of 1-4% and larger cross-task gains of up to 5-6%. Not only does pre-training improve performance within each task, but it also enables reliable cross-task decoding between perception and production. Critically, models trained on speech production decode passive listening above chance, confirming that learned representations reflect shared neural processes rather than task-specific motor activity.
DOI: 10.21437/Interspeech.2026-439
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: We're looking at the paper: "MEG-to-MEG Transfer Learning and Cross-Task Speech/Silence Detection with Limited Data".
Jane: Here's the gist of the paper. Data-efficient neural decoding is a central challenge for speech brain-computer interfaces <ref:2602.18253#pg2> The paper presents the first demonstration of transfer learning and cross-task decoding for MEG-based speech models spanning perception and production <ref:2602.18253#pg2> Transfer learning yields consistent improvements, with in-task accuracy gains of 1-4% and larger cross-task gains of up to 5-6% <ref:2602.18253#pg2> Not only does pre-training improve performance within each task, but it also enables reliable cross-task decoding between perception and production <ref:2602.18253#pg2> Critically, models trained on speech production decode passive listening above chance, confirming that learned representations reflect shared neural processes rather than task-specific motor activity <ref:2602.18253#pg2>.
Tom: The core methodology involves a transfer learning pipeline applied to MEG speech decoding tasks.
Jane: The study pre-trained a Conformer-based model on 50 hours of single-subject listening data and fine-tuned it on just 5 minutes per subject across 18 participants performing speech perception and production tasks <ref:2602.18253#pg2> The pre-training dataset, LibriBrain, provides over 50 hours of within-subject MEG recordings from a single participant during naturalistic English speech listening <ref:2602.18253#pg2> The fine-tuning phase involves evaluating models on Bourguignon et al.'s dataset, which comprises 18 healthy adult participants performing listening, playback, and speech production tasks <ref:2602.18253#pg2> To manage the limited data per subject and task in the fine-tuning set, lightweight modifications were introduced to the model architecture <ref:2602.18253#pg2> These modifications included selecting checkpoints using validation loss rather than F1-macro, introducing RollAugment temporal augmentation, and using soft targets instead of hard 0/1 labels <ref:2602.18253#pg2>.
Tom: The experimental setup compares in-task performance against cross-task generalization.
Paper summary: Tom: So we’ve covered the main claim of this paper, which is that transfer learning using a Conformer based model can significantly improve MEG speech decoding performance when you only have minimal data for the actual tasks you want to test.
Jane: The paper sets up a comparison between pre-training on fifty hours of single-subject listening data and fine-tuning on just five minutes per subject across eighteen participants doing both perception and production tasks <ref:2602.18253#pg1,on 50 hours of single-subject listening data and fine>.
Lu: They are using LibriBrain for pre-training, which is this large dataset of within subject MEG recordings from a single participant during naturalistic English speech listening, which provides over fifty hours of data <ref:2602.18253#pg3>.
Meng: So the fine-tuning part involves using Bourguignon et al.'s dataset with eighteen healthy adult participants, and they used some lightweight modifications to the model architecture to handle the limited data per subject and task.
Tom: They actually introduced a few things like selecting checkpoints based on validation loss instead of F1 macro, adding a RollAugment temporal augmentation, and using soft targets instead of hard zero one labels during fine-tuning <ref:2602.18253#pg3>.
Jane: And the results show that in-task accuracy gains are around one to four percent, but the larger cross-task gains between perception and production can be up to five or six percent <ref:2602.18253#pg3>.
Lalam: The critical finding they point out is that this pre-training enables reliable cross-task decoding between perception and production, which is something previously underexplored <ref:2602.18253#pg2>.
Tom: So, essentially, the paper shows that pre-training on listening data helps the model learn representations that generalize across different subjects and even different types of tasks like production.
Jane: That’s a big deal because it means we don't need massive amounts of per-subject data to build useful speech decoders for brain computer interfaces.
Lu: The paper argues that this transfer learning approach addresses the gap where previous work showed differences in neural representations between passive listening and overt speech, and the role of motor activity versus shared speech representations was limited <ref:2602.18253#pg2>.
Meng: It seems like they are tackling the data efficiency issue head-on by showing that this transfer learning pipeline works across these two contrasting datasets: one large pre-training set and one smaller fine-tuning set.
Conclusion: Tom: So, to wrap up, the authors of "MEG-to-MEG Transfer Learning and Cross-Task Speech/Silence Detection with Limited Data" present a demonstration of using transfer learning from large scale single subject MEG data to enhance speech decoding across new subjects and tasks.
Jane: They show that pre-training on LibriBrain consistently improves in task performance, but more importantly, they see substantially improved cross-task generalization between speech perception and production.
Lu: What this means for the field is that the Conformer model learns representations that capture core neural processes underlying speech, rather than just patterns specific to a single subject or a single task.
Meng: For deployment purposes, it suggests that we can train effective MEG speech detection models with very limited per-subject data, which is critical when long calibration sessions are simply not feasible in clinical settings.
Tom: The paper confirms that representations learned from large scale listening data actually generalize to new subjects and tasks even when you only fine-tune on minimal data <ref:2602.18253#pg2>.
Jane: And they also confirm that decoding relies on shared neural speech representations rather than solely task specific motor activity, as models trained on production can decode passive listening above chance <ref:2602.18253#pg3>.
Lalam: So the overall contribution is establishing a new way to handle data efficiency in MEG speech decoding by showing that bidirectional cross-task transfer between listening and playback tasks shows comparable performance, suggesting those neural mechanisms are shared across different modalities <ref:2602.18253#pg3>.
More episodes
- 2603.28595-Optimistic Actor-Critic with Parametric Policies for Linear Markov Decision Processes
- 2603.29765-Training-Free Dynamic Upcycling of Expert Language Models
- 2605.00778-Observable Performance Does Not Fully Reflect Adaptive System Organization: A Multi-Level Analysis of Gait Dynamics Under Occlusal Constraint
- 2605.06906-TraXion: Rethinking Pre-training Frameworks for Mobility and Beyond
- 2605.09165-Sparse Layers are Critical to Scaling Looped Language Models
- 2605.09439-Inverse Design for Conditional Distribution Matching
- 2605.10313-Signature Approach for Contextual Bandits with Nonlinear and Path-dependent Rewards: Regret Bound and Statistical Inference
- 2605.10878-Minimum Neural Weight Norm = Kolmogorov Complexity
- 2605.11617-MIST: Reliable Streaming Decision Trees for Online Class-Incremental Learning via McDiarmid Bound
- 2605.15085-From Data to Action: Accelerating Refinery Optimization with AI