UV-JEPA Research · in progress
Building a latent-predictive (JEPA) vision-language model for Urdu, a low-resource language, under data-constrained training on a 59K-image / 319K-caption corpus. Implementing dual encoders (ViT-Small, RoBERTa) with an EMA-updated target text encoder and a cross-attention predictor optimized via normalized latent regression, avoiding autoregressive decoding errors. Supervised by Dr. Syed Muhammad Naqi at QAU.