Development of an Open-Source EEG Foundation Model using Model Merging

[ZhNuren]

Development of an Open-Source EEG Foundation Model

Good day @zeydabadi and everyone,
My name is Nuren Zhaksylyk and I am currently pursuing MSc in Computer Vision at MBZUAI, Abu Dhabi, UAE. Currently, I am the Graduate Student Researcher at BioMedIA Lab here at MBZUAI under Dr. Mohammad Yaqub. Currently my ongoing research is application of model soups in medical image tasks and now we are pushing our findings to MICCAI 2024.
As I understood, you want to build foundation model tackling EEG data. I want to list possible issues you might encounter during this process and suggest solutions and show how my experience and skill set is going to help you.

• Data scarcity. This problem is common in medical domain and we might not get enough data to build meaningful foundation model from scratch. What we do in this case? We rely on Transfer learning and will try to finetune models trained on natural signals such as speech recognition. I have huge experience in transferring knowledge from natural domain to medical domain as I was working on Medical Souping paper which we submitted to MICCAI 2024 recently which is available at FissionFusion.
• Noisy signals. As we are about to work with unlabeled data, it will have a lot of noise. Right now I am taking Music AI taught by Prof. Gus Xia where I learned technique of disentanglement. I will be helpful in cleaning the data via signal and noise disentanglement. Also, I already had experience in motor-imagery task classification from EEG data in my bachelors where I utilized mutual information to extract features and SVM for classification.
• Long training time. As you will be training on big amount of data it might take a lot of time which is essential during research to get ahead of competitors and publish your paper first. What we can do is divide our data into several parts, train different models and use model merging which was proven recently to work by Sakana AI. Or we can utilize power of different architectures which we train each of them on all data and then merge blockwise or kernelwise which we did in our Biomedia Lab (MedMerge) and it worked. These are potential ideas for publication which will make your paper much stronger.
• Computational power. As a part of BiomedIA Lab I will be able to use our clusters for training once we establish connection between our lab and yours. It is great opportunity for collaboration between our labs and future papers.

I am ready to spend as much time as needed as I see this project interesting and might extend it further for my MSc thesis. I have whole year ahead to do various projects and aim to publish papers in CVPR and MICCAI.
I am very proficient with PyTorch and building pipelines for training and testing the models. Also, I am mainly working in medical domain and have deep understanding of many concepts and have publication in Nature Scientific Data. We have collected ARCADE dataset for benchmarking deep learning models focusing in tackling CAD.

Contributor:

Nuren Zhaksylyk

• email: nuren.zhaksylyk@mbzuai.ac.ae
• github: ZhNuren
• linkedin: Nuren Zhaksylyk
• CV: NurenZhaksylyk_CV.pdf

Potential Mentors

• Mahmoud Zeydabadinezhad (@zeydabadi)
• Babak Mahmoud

Kind regards,
Nuren