When it comes to smart devices, we expect them to deliver results immediately. Think of your smart watch. You wouldn’t want to spend hours on personal calibration before you can use it. Right now, this is the reality for traditional EEG-based systems. They require lengthy calibrations and work well in the lab under controlled conditions for the people they were trained on. But they struggle the moment they meet someone new.

This is one of the biggest challenges for bringing BCI technology into the real world - the generalization problem. And for anyone building products that rely on mental-state data, it's the challenge that determines whether the technology scales or stays stuck in controlled environments.

Why brains don't generalize easily

Even when two people experience the same task (e.g. solving a problem under pressure or navigating a complex interface), their brain signals look remarkably different. There are many factors that influence EEG decoding, including anatomy, brain activity patterns, and even how people mentally approach the same task.  

For years, EEG-based mental state classification relied on relatively simple machine learning pipelines with models being trained from scratch for each specific task and dataset. These approaches depended heavily on handcrafted features and carefully curated data, and they faced a crucial limitation: EEG data is expensive and time-consuming to collect with the need for specialized equipment, controlled environments, and expert supervision. On top of that, labeling the data can be subjective and requires domain expertise. This leads to a tendency for datasets to be relatively small, making it difficult to train robust models. As a result, the models were often trained on one group of subjects, often failing to generalize to new, unseen individuals.  

To resolve the problem of gathering data that incorporates the complexity and variability of neural signals across people, time, and contexts, we require massive amounts of training data that can reliably work for new users and situations, as well as new approaches in how we handle that data.  

The foundation model shift

Something interesting is happening in EEG research right now, and it's something that Alberto, one of our research interns, has been exploring: the introduction of foundation models for EEG.  

The concept will be familiar to anyone following developments in AI. Rather than training a model from scratch for each specific task and dataset, foundation models are pretrained on massive, diverse datasets. The idea is that through scale and diversity, the model develops a general representation of brain activity that can then be adapted to specific applications. So, using a foundation model to generate relevant embeddings for EEG data allows us to focus on generating and training only the classification models for the specific mental state which considerably reduces the data requirements.  

While still an emerging area of research, early results are promising for EEG specifically, with three meaningful implications. First, better generalization across subjects, because the model has already encountered a wide range of individual variability during pretraining. Second, task adaptability, since the same underlying model can be reused across different applications without starting from scratch. Third, reduced dependence on specific hardware configurations, which matters enormously for real-world deployment across different devices and setups.  

How this connects to what we're building

At Zander Labs, generalization stands as the central engineering challenge behind our universal mental state classifiers. The goal is a system that you can put on any user and have it work immediately, without calibration or retraining. We believe that mental state sensing should be no different from other biofeedback tools.

To support this, we're currently recording data from 4,000 subjects across sites in the Netherlands and Germany. We're building an expanding suite of classifiers that capture human cognition through continuous monitoring of naturally occurring neural responses (e.g. workload, focus, error etc.) at a level of generality that holds across individuals, tasks, hardware, and environments.

The emergence of foundation model approaches in EEG research is directly relevant to further accelerate our work. The more the field moves toward large-scale pretraining on heterogeneous data, the closer it gets to solving the problem of making mental state monitoring robust enough to work comfortably and easily in the real world.

As we continue to improve our BCI infrastructure and explore new ways to ensure its scalability across diverse AI systems, foundation models offer new possibilities to scale BCI technology. And as Zander Labs grows the datasets to train these models, the ceiling on what's possible rises with them.

Curious about how mental state data could work within your AI systems? We'd love to explore it with you.

‍