N A Adarsh Pritam

I am a Research Intern at the Centre for Brain Research, Indian Institute of Science (IISc), and also Master's student in Data Science (2024-26) at Alliance University.

Previously, I was a Summer ML Research Intern at the i2CS Research Group, IIIT Kottayam, where I worked with Dr. Kala S on multi-modal alignment of Vision Language Models for Plant Disease Detection. I graduated with a BSc in Mathematics and Statistics from Bengaluru City University.

My research interests span Machine Learning, Computer Vision, NLP, and Generative Models, with applications to Neuroscience and Healthcare.

Email  /  CV  /  LinkedIn  /  GitHub  /  Google Scholar

Water Your Repo, Grow Your Greens!

In this work, we study how visual stimuli can be recosntructed using fMRI signals by mapping neural activity to the latent spaces of pretrained diffusion-based vision-language models (e.g. LDM). Brain2VLM analyzes the structure of the brain-to-latent mapping rather than only focusing on improving reconstruction pipelines. We analyze how neural activity corresponds to structural diffusion latents and semantic embeddings, and how this varies across regions of the visual cortex. This provides a systematic view of brain-to-latent alignment for brain decoding.

In this work, we study how synthetic data can improve melanoma (skin cancer) diagnosis from dermoscopic images. SkinGenBench provides a controlled setup to analyze how generative model choice and preprocessing affect synthetic image augmentation in medical imaging. We compare GAN-based and diffusion-based approaches under different preprocessing conditions and evaluate their impact on image quality and downstream classification performance. This provides a systematic analysis of synthetic data augmentation for melanoma detection.

Developed a modular end-to-end pipeline for translating American Sign Language (ASL) fingerspelling from static images into English sentences, using Google Teachable Machine and a fine-tuned FLAN-T5 model for gloss-to-text generation.

Working on multimodal health data analysis and machine learning for brain aging and cognitive impairment prediction using large-scale cohort and multi-omics datasets.

Conducted research as a part of master's thesis at Alliance University on applying machine learning to neuroscience and healthcare, including brain-to-image reconstruction from fMRI signals (Brain2VLM) and synthetic data augmentation for melanoma diagnosis (SkinGenBench).

Provided deep learning edge to the group by developing a Vision-Language Model (VLM) for descriptive plant disease diagnosis by integrating a CLIP ViT-Large with an InternLM2-7B LLM.

Proposed an interpretable statistical framework for modeling life expectancy using World Health Organization (WHO) data, used VIF-based multicollinearity resolution with domain-aware feature engineering and sequential feature selection. Developed and validated regression models within a reproducible pipeline, and deployed a Streamlit dashboard to analyze the influence of socioeconomic and healthcare variables, highlighting education, GDP, vaccination index, and BMI as key drivers.

Updated Wei Yang's popular pytorch-classification repo by adding a unified training pipeline supporting 20+ architectures and automatic head adaptation for transfer learning. Also built an architecture-agnostic Grad-CAM system for clear interpretability across the entire model zoo.

Implemented a LLaMA-inspired large language model from scratch in PyTorch, building core transformer components including multi-head attention, grouped-query attention (GQA), and rotary positional embeddings (RoPE). Analyzed the role of attention mechanisms and positional encoding in token representation and sequence modeling.

For a complete overview of my open source contributions, check this GitHub search query.