Research Portfolio

🎓 Master’s Thesis & Extracted Journal Papers

Grade: 20/20

Developing Transformer-based Deep Learning model for Signal Decoding in BCI

Thesis Overview: Pioneered the development of neurophysiologically robust architectures to decode motor intent from EEG signals. This research exposed the "Clever Hans" effect in SOTA models and led to two major journal publications (Under Review at Biomedical Signal Processing and Control - BSPC):

  • Paper 1: The Illusion of Accuracy in Rehabilitation BCI. An Explainable AI (XAI) investigation using Grad-CAM, Saliency Maps, and RSA to prove that high-performance CNNs often learn high-amplitude ocular artifacts (EOG) rather than true neural intent.
  • Paper 2: A Robust Benchmark for Rehabilitation BCI. Validating Convolutional Transformers (CvT) on neurophysiologically isolated EEG signals to ensure safe and reliable intent decoding for clinical use.
Explainable AI (XAI) Convolutional Transformers (CvT) RSA Analysis Artifact Robustness Optuna

🔬 Independent & Applied Research

Active

Speech BCI: Imagined Speech Decoding

Independent Project: Developing an end-to-end Spectro-Temporal Transformer architecture to classify phonemes and words directly from EEG signals using the KARA ONE dataset. This work focuses on building robust silent communication interfaces independent of my thesis research.

Spectro-Temporal Modeling KARA ONE Dataset Assistive Tech SHAP
Completed

Bone Age Estimation (Medical Computer Vision)

Implemented a Vision Transformer (ViT) for regression analysis on pediatric hand X-ray images. This project achieved a lower Mean Absolute Error (MAE) compared to traditional CNNs and was presented at the AIMS and IOAS conferences (Tehran & Germany).

Vision Transformer (ViT) Medical Imaging Regression Analysis

🛠 Technical Stack & Methodologies

AI & Modeling

Python PyTorch Transformers (CvT, ViT) Optuna XAI (SHAP, Grad-CAM)

Neuro-Engineering

MNE-Python EEGLAB fNIRS/EEG Recording ASR / ICA RSA Analysis

Methodological Proficiency

Deep Learning (PyTorch & Transformers) 95%
Neural Signal Processing (EEG/fNIRS) 90%
Explainable AI & Model Validation 85%