⚡ electrical engineering

GPT2MEG: Quantizing MEG for Autoregressive Generation

This paper introduces GPT2MEG, a quantized Transformer-based framework that repurposes language model architectures to autoregressively generate realistic, high-rate multichannel MEG time series, outperforming WaveNet variants and linear baselines in reproducing neural statistics and enabling conditional task-evoked simulation across multiple subjects.

Original authors: Richard Csaky, Mats W. J. van Es, Oiwi Parker Jones, Mark Woolrich

Published 2026-01-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Richard Csaky, Mats W. J. van Es, Oiwi Parker Jones, Mark Woolrich

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your brain is a massive orchestra with hundreds of musicians (sensors) playing simultaneously. Magnetoencephalography (MEG) is like a super-sensitive microphone recording every note they play, second by second. The problem is, this recording is incredibly complex, messy, and different for every person.

The paper introduces a new AI system called GPT2MEG. Think of it as a "brain music composer" that learns to write new, realistic brain recordings from scratch, just by listening to old ones.

Here is how it works, broken down into simple concepts:

1. The Big Idea: Turning Brain Waves into "Words"

Standard AI models for language (like the one you are talking to right now) work with words. They predict the next word in a sentence. But brain waves are continuous, smooth signals, not distinct words.

To fix this, the researchers invented a way to quantize (or chop up) the brain waves.

The Analogy: Imagine taking a smooth, flowing river and turning it into a staircase of 256 distinct steps.
The Process: They take the raw brain signal, squash it into a manageable range, and snap it to the nearest "step." Now, instead of a smooth wave, the brain signal looks like a sequence of numbers (tokens), just like a sentence is a sequence of words.
The Result: They can now use a language model (specifically a GPT-2 style Transformer) to predict the next "step" of the brain signal, just like predicting the next word in a story.

2. The Two Contenders: The "Soloist" vs. The "Conductor"

The team tested two different types of AI composers to see which could write the best brain music:

The WaveNet Variants (The Soloists): These models are like musicians who practice their own instrument in isolation. They look at one sensor at a time and try to guess the next note. Even when they try to listen to each other, they struggle to capture the complex harmony between all the sensors.
GPT2MEG (The Conductor): This model is the star of the show. It doesn't just look at one sensor; it understands the whole orchestra. It uses special "ID cards" (embeddings) to know:
- Which sensor is playing (Channel Embedding).
- Who the person is (Subject Embedding).
- What task they are doing (Task/Condition Embedding).

3. What Did They Find?

The researchers didn't just ask, "Did the AI guess the next note correctly?" (Because even a broken clock is right twice a day). Instead, they asked, "Does the music feel real?"

The Rhythm (Spectral Fidelity): Real brain waves have a specific "hum" or rhythm (like alpha and beta waves). The GPT2MEG model reproduced these rhythms almost perfectly. The other models (WaveNet) and simple linear models sounded a bit flat or robotic in comparison.
The Dynamics (The HMM Test): They used a statistical tool (Hidden Markov Models) to see how the brain switches between different "states" (like switching from thinking about a cat to thinking about a dog).
- The simple models produced brain waves that were too boring and repetitive.
- GPT2MEG produced brain waves that switched states with the same natural, chaotic variety as a real human brain.
The Reaction (Task-Evoked Responses): When the researchers told the AI, "Pretend the person just saw a picture," the GPT2MEG model generated a brain response that looked exactly like a real human reacting to a picture. The other models failed to capture this specific reaction.

4. The "Group Chat" Feature (Scaling to Many People)

One of the coolest tricks GPT2MEG pulled off was learning from 15 different people at once.

Usually, AI models get confused when you mix data from different people.
GPT2MEG used "Subject Embeddings" (digital name tags) to learn the unique style of each person while still finding the common patterns.
The Result: It could generate a brain recording for a specific person that sounded like them, even though it was trained on a group.

5. Why Does This Matter? (According to the Paper)

The paper highlights one major practical use: Data Augmentation for Decoding.

Imagine you are trying to teach a computer to read minds (decode what a person is thinking), but you only have a tiny amount of data (like 20 trials). It's hard to learn from so little.

The researchers used GPT2MEG to generate fake but realistic brain data to add to their training set.
They trained a decoder on this fake data first, then fine-tuned it on the real data.
The Outcome: The more fake data they generated, the better the decoder got at reading the real brain signals. It's like practicing on a flight simulator before flying a real plane; the simulator didn't replace the real plane, but it made the pilot much better.

Summary

GPT2MEG is a new way to teach AI to understand the brain. By turning brain waves into a language the AI can read, it learned to compose new, realistic brain recordings that capture the complex rhythms, the natural switching of brain states, and the specific reactions to tasks better than previous methods. It acts as a powerful tool to generate extra practice data, helping scientists build better tools to understand how our brains work.

Technical Summary: GPT2MEG: Quantizing MEG for Autoregressive Generation

Problem Statement

Magnetoencephalography (MEG) and related electrophysiological recordings provide millisecond-resolved measurements of large-scale brain dynamics but are difficult to model due to high dimensionality, strong temporal structure, and significant variability across subjects and recording setups. While recent self-supervised foundation models have successfully learned transferable representations from unlabelled neural data using masked or contrastive objectives (e.g., for classification or regression), autoregressive generative foundation models capable of simulating realistic, continuous multichannel neural time series remain comparatively underexplored, particularly for MEG. Existing approaches often lack the ability to support data augmentation, uncertainty-aware decoding, or mechanistic hypothesis testing through realistic simulation of task-evoked activity.

Methodology

Core Approach

The authors propose GPT2MEG, an autoregressive Transformer model adapted for sensor-level MEG. The central challenge addressed is applying a discrete-token language model architecture (GPT-2) to continuous, high-rate, multichannel time series.

Quantization and Tokenization

To bridge the gap between continuous MEG signals and discrete language model inputs, the authors employ a simple quantization scheme:

Preprocessing: Each channel is scaled to $(-1, 1)$ .
Companding: A $\mu$ -law companding transform is applied to skew the distribution, allocating more quantization levels to smaller magnitudes.
Discretization: The transformed signal is uniformly binned into $Q=256$ discrete tokens.
This discretization allows the model to be trained using cross-entropy loss on next-token prediction, avoiding the mean-prediction bias often associated with MSE regression.

Model Architecture

The authors evaluate two classes of autoregressive models trained purely on next-step prediction:

Modified Multichannel WaveNet Variants:
- WavenetFullChannel (WFC): Treats channels as a batch dimension. A shared causal dilated-convolution stack is applied independently to each channel, with channel identity provided via per-channel embeddings.
- WavenetFullChannelMix (WFCM): Adds a learned mixing projection to the skip representation to share information across channels.
GPT2MEG (Transformer):
- Adapts the GPT-2 decoder-only architecture.
- Input Embeddings: The model receives tokenized inputs ( $X$ $X$ ) augmented with three types of embeddings added element-wise:
  - Positional Embeddings ( $W_p$ ): For temporal structure.
  - Channel Embeddings ( $W_c$ ): Distinct learned vectors for each of the $C$ sensors to handle channel-specific characteristics.
  - Subject Embeddings ( $W_o$ ): To scale the model across multiple subjects.
  - Task/Condition Embeddings ( $W_y$ ): Time-aligned labels (e.g., stimulus identity) to enable conditional generation.
- Channel Independence: Notably, the base GPT2MEG model treats channels as a batch dimension (univariate per channel) during the forward pass, relying on embeddings to distinguish channels rather than explicit cross-channel mixing layers, though the authors acknowledge this as a limitation regarding cross-channel dependency modeling.

Evaluation Metrics

The authors argue that one-step forecasting accuracy is insufficient for evaluating generative models. Instead, they emphasize long-horizon generation metrics:

Spectral Fidelity: Power Spectral Density (PSD) comparisons.
Multivariate Dynamics: Summary statistics (fractional occupancy, mean lifetime, switching rate) of a 12-state Hidden Markov Model (HMM) fitted to the generated vs. real data.
Task-Evoked Structure: Trial-averaged evoked responses under task conditioning.
Downstream Utility: Decoding performance when classifiers are trained on simulated trials (transfer learning).

Key Contributions

GPT2MEG Architecture: Introduction of an autoregressive Transformer for MEG that utilizes discrete tokenization combined with channel, subject, and stimulus embeddings to enable conditional generation of realistic neural time series.
Benchmarking Beyond One-Step Accuracy: Demonstration that one-step prediction accuracy is weakly informative for long-horizon behavior. The paper establishes that while deep models and linear AR baselines may have similar one-step accuracy, Transformers significantly outperform them in reproducing long-range spectral, dynamical, and task-evoked statistics.
Multi-Subject Scaling: Successful scaling of the model to multiple subjects via subject embeddings, showing that a multi-subject model can generate subject-specific data that improves downstream decoding via transfer learning.
Conditional Generation: Validation that the model can generate task-evoked responses that closely match the amplitude, timing, and trial-to-trial variability of real MEG data when conditioned on stimulus labels.

Results

One-Step Prediction: All models (WaveNet variants, GPT2MEG, and linear AR) performed only moderately better than a naive "repeat previous timestep" baseline in terms of top-1 accuracy. This suggests that next-step prediction accuracy does not correlate strongly with the quality of long-horizon generation.
Spectral Fidelity: GPT2MEG and WaveNet variants reproduced the characteristic $1/f$ shape and alpha/beta band peaks (10 Hz, 19 Hz) of real MEG. However, WaveNet variants showed reduced power at the 19 Hz peak compared to GPT2MEG.
Multivariate Dynamics (HMM): GPT2MEG was superior in capturing the distribution of HMM state dynamics (e.g., switching rates, state lifetimes). Linear AR and WaveNetMix failed to produce data with variable state statistics, while GPT2MEG matched the variance and mean of real data distributions most closely.
Task-Evoked Responses: GPT2MEG successfully generated evoked responses with correct amplitude and timing (e.g., visual responses to stimulus offset) that closely matched real data. In contrast, WaveNet models failed to capture the evoked time-course shape.
Transfer Learning:
- Classifiers trained on GPT2MEG-generated data achieved lower accuracy than those trained on real data initially.
- However, pretraining a decoder on GPT2MEG-generated data (40–60 trials per condition) followed by fine-tuning on a small amount of real MEG data (20 trials) significantly improved decoding accuracy compared to training on real data alone.
- Each additional 20 generated trials per condition improved final decoding accuracy by approximately 2% on real MEG validation trials.
Ablation Studies: Removing channel or condition embeddings degraded performance, confirming their necessity. Models trained with shuffled labels or single labels failed to generate distinct evoked responses, proving the model relies on semantic content of the conditioning, not just temporal templates.

Significance and Claims

The paper claims that GPT2MEG represents a significant step toward building autoregressive generative foundation models for MEG. By repurposing language model architectures through a simple quantization and embedding scheme, the authors demonstrate that Transformers can faithfully reproduce the temporal, spectral, and task-evoked statistics of real neural data better than WaveNet variants or linear autoregressive baselines.

The significance lies in the potential for data augmentation in data-scarce decoding scenarios and the ability to perform mechanistic hypothesis testing by simulating realistic neural dynamics. The authors modestly note that while their experiments are small-scale, the framework has the potential to scale to produce foundational models for brain signals. They explicitly state that the current channel-independent architecture is a limitation and that future work must address cross-channel dependencies and more flexible conditioning for diverse stimulus modalities.

Limitations Acknowledged:

The channel-independent nature of the GPT2MEG model means it does not directly leverage cross-channel information for single-sensor predictions.
One-step prediction accuracy is a poor proxy for generative quality.
The reliance on categorical task labels limits scalability to diverse tasks without robust representation learning for stimuli (e.g., images, audio).
Generated data currently does not fully match the complexity of real neural data in terms of trial-to-trial variability in state activations.