04/29/2024 Update
Introduction
Our goal for this project was to create a probability distribution for EEG and spectrogram data on various categories of harmful brain activities such as seizures.
Progress
We initiated the project by preprocessing EEG data, focusing on feature extraction through Fast Fourier Transforms (FFTs) and converting these signals into a format suitable for convolutional neural network (CNN) processing. Sean managed this intricate preprocessing phase, ensuring our data was aptly prepared for the subsequent modeling stages.
Simultaneously, Amit and Aviral developed and implemented the initial models for our project. The architecture consists of a transformer to capture the temporal dynamics of EEG signals and a CNN model designed to process EEG data represented as images. These models form the core of our analysis pipeline.
Challenges
We tried various combinations of multiple advanced models, including a transformer, LSTM, CNN, and a Vision Transformer (ViT) which presented substantial challenges, particularly in maintaining efficient data flow and processing. Optimizing the models to prevent overfitting and achieve high precision required continuous adjustments and parameter tuning.
In addition, it was very difficult to train the model due to the size of the dataset. We found an online implementation (cited in our writeup) which develops a way to segment and batch the dataset and save the weights of the model. This allows us to train the model efficiently.
Next Steps
The immediate focus is on completing the integration of the Vision Transformer and conducting extensive testing to validate and refine our system. If everything works out, we plan to transition to using feature extraction to use multiple types of inputs and make our model truly multimodal. The final stages will involve compiling our findings and methodologies into a comprehensive report and preparing for the final presentation. The work we completed so far has positioned us well to meet our project goals.
Log in or sign up for Devpost to join the conversation.