Building AI That Learns Like How A Capable Human Learns

With the advent and rise of artificial intelligence (AI), humans have created and taught numerous AI models to master a variety of tasks. Ranging from differentiating dogs and cats, to generating images and texts for us. In the field of neurorehabilitation, scientists are even capable of building AI to recognise a human’s brain signals in order to move objects with just a thought.

However, conventional AI has a strong limitation: it tends to do well only on what it has seen before but does poorly when things deviate too far from the training. Imagine you would like to train a robot to recognize dogs. You feed it a large amount of dog images, instructing the robot that these are images of dogs. That’s all good, until you realise that there are many things out there that looks like a dog and now you have to re-train your robot. Continuously undergoing these re-training iterations can be exhausting and is not feasible on a large scale.

Similarly, when we train AI models to recognize a human’s brain signal, we now know thanks to previous studies that the human brain doesn’t always stay the same. How we imagine moving our left hand now, might not be exactly the same as how we imagine it say a few days later. This time variation causes the AI model to become confused, and decoding performance drops as a result. Thus, there’s a need for us to teach AI models to become self-learning so that they can automatically update themselves towards the latest state of the subject’s brain.

Adaptive and meta-learning strategies have been previously applied towards various applications such as image recognition and robotics manipulation. However, brain-computer interfaces (BCI) often use electroencephalography (EEG) signals to read motor intention due to the benefit of it being non-invasive. These signals are more abstract in nature as compared to images and therefore the models suffer more greatly whenever there are significant deviations in brain states, making such strategies paramount to overcoming this challenge. Yet, such techniques are rarely translated towards BCI studies as they are originally optimized towards other fields of research.

Subject-Adaptive Meta-Learning

This is where subject-adaptive meta-learning [1] comes into play. Subject-adaptive meta-learning is inspired from previous methods applied in image classification tasks. One key difference is that in image recognition, the model aims to quickly learn how to classify a new class. For instance, from classifying dogs and cats to dogs and lions. In the context of brain motor imagery decoding, we seek to redefine the new task as a new subject to be decoded instead. In order to further understand this method, we can break it down into two main parts: (1) Subject adaptation and (2) Meta-learning.

Subject adaptation: Subject adaptation is derived from the simple idea that taking the newest and most up-to-date data from our task and re-training the model should ideally improve our model’s performance. This is because it has the lowest time deviation and should be a good representative of the current brain state. By taking our model and fine-tuning the parameters to better suit to these up-to-date data, our model should be able to significantly reduce the drops in accuracy and may even boost the model’s overall decoding ability.

Meta-Learning: While fine-tuning the model using latest data sounds straightforward enough, can we go one level deeper yet? Yes, we can. In recent years, researchers have found a surprisingly simple way to improve our model’s ability to learn and adapt to new tasks. In short, we simply train our AI to learn just like how a human does. This is done by shaping the way we structure the initial training of the AI model to look identical to how we expect the fine-tuning stage would look like. Through this, we have achieved meta-learning by teaching the model to learn like how a human learns. The model learns parameters that best encourage the model to quickly adapt towards the fine-tuning data, just like how humans after undergoing countless of trainings, can quickly learn what is the most optimal and best way to adapt towards a new task.

In the context of training a BCI model, the AI model is trained using data from multiple subjects different from the target subject. A simple application of a fine-tuning strategy would be to train the AI model with the other subjects’ data, followed by updating the model parameters whenever target subject data is obtained. However, subject-adaptive meta-learning is done by treating each subject in the training phase as a pseudo-target subject.

Through this, the model is trained first on the remaining data, and attempts to fine-tune towards the pseudo-target. Then, how well it has managed to adapt toward the pseudo-target is provided as feedback to the model as a meta-loss, allowing it to learn its level of adaptation. Repeating this iteration across many times gives us a model that has learnt the best ways to quickly adapt towards an unseen target subject. This may then be applied to the actual target subject by fine-tuning it with the actual target data. Since our model has already learnt how to learn, it is able to quickly adapt to the target as compared to conventional fine-tuning.

To further ensure the strong generalisability of the learnt model, we take it a step further by backpropagating the normal loss of its decoding ability on top of the previously identified meta-loss. Through this, the model still learns the ability to differentiate between the base classes, while simultaneously attempting to learn the best way to learn given its current known knowledge. This closely mirrors the way humans learn.


In summary, we have achieved an AI model that learns how to learn and adapts with the target’s ever-changing brain state. This builds an important foundation towards co-adaptive models, whereby both AI models and humans learn and adapt around one another over time. We achieve this by treating each training subject as a pseudo-target subject and meta-training the model to learn the nature of the subsequent adaptation task. With this, state-of-the-art decoding performance may be achieved in EEG-based motor imagery decoding. A strong benefit to this model is that the subject-adaptive meta-learning framework is model-agnostic and may be applied to virtually any EEG-based task on any AI model as long as there exists multiple subjects.

[1] Ng, H. W., & Guan, C. (2024). Subject-independent meta-learning framework towards optimal training of EEG-based classifiers. Neural Networks, 106108.