Multimodal Research Hub - Vision-Language Models (VLMs)
A living resource for Vision-Language Models & multimodal learning
Pathological Truth Bias in Vision-Language Models
MATS, a behavioral audit for vision language models, identifies systematic failures in spatial consistency and suggests repair paths through activation patching.
SSL-Vision
Implementation of four self-supervised learning (SSL) algorithms; SimCLR, MoCo-v2, BYOL, and DINO on the CIFAR-10 dataset using a ResNet-18 backbone. All scripts are designed to run on Google Colab Free Tier (single GPU, ~12 GB RAM) with CIFAR-10.
Task-Aware Segment Anything with LoRA
Abstract PyTorch pipeline that uses a hypernetwork to generate task-specific LoRA adapters for Meta’s Segment Anything Model from natural language prompts, targeted segmentation on COCO instances and benchmarked mIoU via pycocotools. Github repo You can find the repo here.