Cognitive Science and AI — Neuroimaging and Brain Encoding Portfolio

A curated portfolio of my work in cognitive science and AI, focused on neuroimaging preprocessing, brain decoding, and visual brain encoding using classical machine learning, deep learning, and vision-language models.

Overview

This repository contains two major projects:

• Assignment 1: end-to-end neuroimaging preprocessing and decoding using fMRI and EEG/MEG pipelines
• Assignment 2: visual brain encoding using CNNs, vision-language models, semantic embeddings, and layer-wise transformer analysis

The work emphasizes:

• loading and preprocessing scientific neural data
• building interpretable decoding and encoding pipelines
• comparing model families across ROIs
• evaluating brain-alignment hypotheses using quantitative metrics and visual diagnostics

Start Here

If you're reviewing this repository quickly, see START_HERE.md.

What this repository demonstrates

• Neuroimaging preprocessing: fMRI, EEG, and MEG workflows starting from raw or structured scientific data
• Brain decoding: ROI-based classification using session-aware evaluation
• Visual brain encoding: ridge / linear encoding models from CNN, VLM, and transformer representations
• Representation analysis: layer-wise alignment, prompt ablations, and ROI-specific comparisons
• Interpretation: brain visualizations, topographic maps, correlation-based evaluation, and controlled model comparisons

Repository Structure

• assignment-1/ — foundations for fMRI and EEG/MEG preprocessing, feature extraction, and decoding
• assignment-2/ — foundation models for visual brain encoding using CNNs, VLMs, and transformer layers

Best Places to Start

1. assignment-1/

Focuses on end-to-end preprocessing and decoding across neuroimaging modalities:

• fMRI ROI decoding
• EEG or MEG preprocessing from raw scientific formats
• interpretable classification pipelines
• modality-appropriate visualizations

2. assignment-2/

Focuses on modern NeuroAI and brain encoding:

• CNN-based encoding baselines
• vision-language model embeddings
• semantic prompt ablations
• layer-wise transformer-to-brain mapping

Tech Stack

Python, NumPy, Pandas, Scikit-learn, PyTorch, MNE, Nilearn, Matplotlib, Jupyter Notebook

Why this repo matters

This repository reflects how I approach cognitive AI problems end to end:

1. understand the scientific dataset and task
2. preprocess signals and extract meaningful features
3. build interpretable baseline and advanced models
4. evaluate with scientifically appropriate metrics
5. visualize and interpret model–brain relationships