🤖 AI

EvoBrain: Continual Learning of EEG Foundation Models Across Heterogeneous BCI Tasks

This paper introduces EvoBrain, a dynamic continual learning framework that enables EEG foundation models to adapt across heterogeneous BCI tasks by balancing plasticity and stability through Neuro-Spectral Task Normalization and Response-Affinity Distillation, thereby overcoming the scalability and forgetting limitations of traditional task-isolated fine-tuning.

Original authors: Yangxuan Zhou, Sha Zhao, Jiquan Wang, Shijian Li, Gang Pan

Published 2026-06-02

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yangxuan Zhou, Sha Zhao, Jiquan Wang, Shijian Li, Gang Pan

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 you have a brilliant, super-smart brain that has read millions of books about how the human mind works. This is the EEG Foundation Model. It's a general expert that knows the basics of brain signals.

However, in the real world, we don't just have one job for this brain. We need it to do many different things: help a paralyzed person move a cursor (Motor Imagery), detect if someone is sleeping (Sleep Staging), or guess if they are happy or sad (Emotion Recognition).

The Problem: The "One Job, One Brain" Trap

Traditionally, when scientists wanted this brain to learn a new job, they had to take the original brain, retrain it from scratch for that specific job, and then save a completely new copy of it.

The Analogy: Imagine you are a chef. To learn how to bake a cake, you hire a new chef. To learn how to grill a steak, you hire another new chef. To learn how to make sushi, you hire a third.
The Result: You end up with a massive, expensive kitchen full of different chefs, each good at only one thing. If you want to serve a full menu, you need all of them. This is slow, expensive, and doesn't let the chefs learn from each other.

The Solution: EvoBrain (The "Super-Adaptable Chef")

The paper introduces EvoBrain, a new way to train this brain so it can learn new jobs one after another without forgetting the old ones, all while using just one single model.

Think of EvoBrain as a chef who doesn't just memorize recipes but learns how to adapt their cooking style on the fly. It uses two special "kitchen tools" to make this happen:

Tool 1: The "Spectral Tuner" (Neuro-Spectral Task Normalization)

Different brain tasks speak in different "languages" and use different "frequencies."

The Analogy: Imagine you are listening to a radio. Sometimes you need to tune into a low, rumbling bass station (like sleep signals), and other times you need a high-pitched, fast station (like fast thinking or emotion).
How it works: When a new task arrives, EvoBrain doesn't just force the brain to listen to the new station. It uses a Spectral Tuner to adjust the volume and clarity of specific frequencies. It says, "Okay, for this sleep task, let's boost the slow waves and quiet the fast ones." This helps the brain understand the new task immediately without getting confused by the old settings.

Tool 2: The "Memory Keeper" (Response-Affinity Distillation)

The biggest fear in learning new things is forgetting the old things (a problem called "catastrophic forgetting").

The Analogy: Imagine you are learning to play the guitar. If you start learning the drums, you might accidentally start hitting the guitar strings like drumsticks and ruin your guitar playing.
How it works: EvoBrain uses a Memory Keeper that acts like a strict coach.
1. The Geometry Guard: It remembers exactly how the brain reacted to old tasks (like the specific shape of a chord) and makes sure the brain doesn't distort those shapes when learning something new.
2. The Compatibility Check: It only lets the brain borrow ideas from new tasks if they are similar. If the new task is "sleep" and the old task was "emotion," the coach says, "These are too different; don't mix them up." But if the new task is "another type of movement," the coach says, "Great, borrow those techniques!"

The Result: One Brain to Rule Them All

By using these two tools, EvoBrain achieves something amazing:

It learns continuously: It can go from learning to detect sleep, to detecting emotions, to detecting depression, one after another.
It doesn't forget: It keeps its skills for the first tasks while mastering the new ones.
It saves space: Instead of needing 10 different models for 10 different jobs, you only need one evolving model.

What the Paper Actually Proved

The researchers tested this on six different brain tasks (like sleep, movement, and emotions) using four different types of AI architectures.

They found that EvoBrain was much better at balancing "learning new things" (plasticity) and "remembering old things" (stability) than previous methods.
They showed that this approach works regardless of the specific AI "engine" used (whether it's based on Transformers or Mamba).
They visualized the brain's internal "frequency tuning" and confirmed that the system was actually adjusting to the specific needs of each task, just like a radio tuner.

In short: EvoBrain turns a rigid, single-purpose brain model into a flexible, lifelong learner that can handle a whole menu of brain-computer interface tasks without needing to be rebuilt every time.

Technical Summary: EvoBrain

Problem Statement
Electroencephalography (EEG) foundation models, pre-trained on massive unlabeled corpora, offer a unified framework for universal brain decoding. However, current post-training strategies rely on task-isolated static adaptation (e.g., full fine-tuning for each new task). This paradigm presents two critical limitations:

Hindered Knowledge Transfer: It prevents the synthesis of higher-order, invariant neural patterns across heterogeneous BCI tasks (e.g., sleep staging, motor imagery, emotion recognition), confining the model to narrow task boundaries.
Scalability Bottlenecks: Adapting to a new task requires re-training from scratch and storing separate parameter sets, causing computational and storage overheads to scale linearly with the number of tasks.

While Continual Learning (CL) offers a theoretical solution, applying it to EEG foundation models is uniquely challenging due to multi-level heterogeneity. Different BCI paradigms exhibit distinct acquisition settings, label semantics, signal morphologies, and neuro-spectral patterns. Standard CL methods often fail to address the specific need to recalibrate task-dependent spectral responses while preserving the response geometry of previous tasks, leading to catastrophic forgetting or negative transfer.

Methodology: EvoBrain
The authors propose EvoBrain, a dynamic, task-aware continual learning framework designed to evolve a single pretrained EEG foundation model across a stream of heterogeneous tasks. The framework balances learning plasticity (adapting to new tasks) and memory stability (retaining old knowledge) through two complementary components:

Neuro-Spectral Task Normalization (NSN):
- Goal: Enhance plasticity by aligning incoming tasks with historical statistics while recalibrating task-specific spectral responses.
- Mechanism:
  - Prototype-guided Temporal Normalization (PTN): Uses historical task prototypes (exponential moving averages of task-wise data means) to stabilize cross-task temporal statistics. It employs learnable moment-gating parameters to dynamically rescale and shift current task data, reducing distributional drift without rigidly forcing a match.
  - Task-conditioned Spectral Residual Modulation (SRM): Explicitly models task-specific spectral patterns. It computes the real-valued Fourier spectrum of the normalized input, applies learnable soft band gates for standard EEG frequency bands ( $\delta, \theta, \alpha, \sigma, \beta, \gamma$ ), and modulates the spectral residual. This allows the model to enhance or suppress specific frequency bands based on the task's neuro-spectral signature.
Response-Affinity Distillation (RAD):
- Goal: Enhance stability by preserving old-task response organization and promoting selective knowledge transfer.
- Mechanism:
  - Time-Dependent Replay: Samples historical tasks with higher probability for older tasks (which are more vulnerable to forgetting) to provide replay data.
  - Response-Geometry Distillation (RGD): Preserves the relational geometry of replayed samples by matching the off-diagonal entries of the sample-wise relational matrices between the current student model and a frozen teacher model (the model state after learning the replayed task). This maintains the structural relationships of old-task responses.
  - Spectral-Affinity Alignment (SAA): Aligns response statistics (mean, variance, covariance) between current and replayed tasks only when they exhibit compatible spectral signatures. This is achieved by computing a "signed spectral affinity" based on the similarity of the task-conditioned band gates. This prevents indiscriminate alignment between unrelated paradigms, mitigating negative transfer.

The overall objective function combines the task-specific classification loss with the RAD loss, weighted to balance plasticity and stability.

Key Contributions

Pioneering Cross-Task CL in EEG: This work is the first to explore cross-task continual learning for EEG foundation models, moving beyond single-paradigm or single-subject continual learning.
Novel Framework (EvoBrain): Introduces a plug-and-play framework that enables a single unified model to generalize across diverse EEG tasks without restarting task-isolated fine-tuning.
Architectural Agnosticism: The framework is validated on four diverse foundation backbones, including Transformer-based (CBraMod, CSBrain, DeeperBrain) and Mamba-based (EEGMamba) architectures, demonstrating its efficacy regardless of the underlying backbone.
Mechanism Design: Proposes NSN and RAD to specifically address the plasticity-stability trade-off in the context of neuro-spectral heterogeneity, rather than relying on generic distribution alignment.

Experimental Results
The authors evaluated EvoBrain on six distinct BCI tasks (Sleep Staging, Motor Imagery, Emotion Recognition, Imagined Speech, Mental Disorder Diagnosis) using four foundation models under three different task arrival orders.

Performance: EvoBrain consistently outperformed state-of-the-art continual learning baselines (including EWC, LwF, DER++, AGEM, and AdaptFormer) across all backbones.
Stability-Plasticity Trade-off: The method achieved a low Relative Forgetting Rate (RFR) (typically below 4.0%), indicating strong retention of previous tasks while maintaining high accuracy on new tasks. In some cases, performance on earlier tasks improved after learning subsequent tasks, suggesting positive transfer.
Robustness: The framework demonstrated robustness across different task sequences and backbone architectures.
Ablation Studies: Removing either NSN or RAD significantly degraded performance, confirming their complementary roles. Specifically, RGD was identified as the key component for preserving old-task relational structures.
Comparison with Static Paradigms: EvoBrain retained 91.3% of the performance of isolated single-task fine-tuning while using only one shared backbone, offering a scalable alternative to the linear cost growth of static fine-tuning.

Significance and Claims
The paper claims to advance the realization of a "one-for-all" EEG decoding system. By formulating downstream adaptation as a cross-task continual learning problem, EvoBrain addresses the fundamental limitations of static, task-isolated post-training strategies. The authors position their work as a necessary step toward scalable, general-purpose neural decoding systems that can incrementally acquire knowledge from heterogeneous BCI tasks without catastrophic forgetting. The study emphasizes that successful continual adaptation in EEG requires not just generic normalization but specific neuro-spectral recalibration and affinity-aware response preservation.

The paper concludes that while the current study uses a controlled task-incremental protocol, the proposed framework provides a practical post-training direction for scalable EEG foundation models, paving the way for future exploration in more flexible online settings.