During my summer research internship at NIT Calicut, I focused on decoding cognitive states from raw Electroencephalography (EEG) brain signals using deep learning. I implemented five different deep learning architectures combining Convolutional Neural Networks (CNNs) and Transformer encoders in PyTorch. Additionally, I successfully developed and evaluated a novel Channel Correlation Positional Encoding technique that models electrode-position topology relative to the central Cz sensor. This research directly supports the development of non-invasive Brain-Computer Interfaces (BCIs) by utilizing multi-head self-attention to capture spatial-temporal correlations in neural data.

Research & Development Internship Report: EEG Classification using Transformers

Institution: National Institute of Technology (NIT), Calicut

Project Title: Implementation of Spatial-Temporal Transformer-Based Deep Learning Models for Raw EEG Classification

Role: Machine Learning Research Intern

Date: Summer 2025

• Implemented and evaluated 5 state-of-the-art Transformer-based architectures (s-Trans, t-Trans, s-CTrans, t-CTrans, f-CTrans) in PyTorch for classifying raw 64-channel EEG signals from the PhysioNet Motor Imagery dataset, reproducing a major IEEE Transactions study.
• Developed a custom Channel Correlation Positional Encoding mechanism using 3D spatial electrode coordinates (cosine similarity to the central Cz electrode) to model spatial topology in self-attention layers.
• Built a robust EEG preprocessing pipeline in MNE-Python and NumPy utilizing bandpass filtering (0.5–30 Hz), sliding-window trial segmentation (3s and 6s epochs), Z-score normalization, and random Gaussian noise augmentation ($\alpha = 0.01$) to prevent model overfitting.
• Designed and coded a scalable 5-fold cross-individual cross-validation training suite with automated early stopping, model checkpointing, and GPU acceleration support.
• Executed comprehensive unit testing and validation scripts to verify forward-pass dimensions, gradient propagation, and optimization steps across all 5 architectures.

Overview & Objective

The primary objective of this internship was to implement, test, and package the deep learning architectures proposed in the IEEE TNSRE 2022 research paper:

“A Transformer-Based Approach Combining Deep Learning Network and Spatial-Temporal Information for Raw EEG Classification” (Xie et al.).”

The project utilizes the PhysioNet EEG Motor Movement/Imagery Dataset (EEGMMIDB), consisting of:

• Subjects: 109 individuals.
• Signals: 64-channel raw EEG sampled at 160 Hz.