Jaisidh Singh

I work on a broad range of problems exploring the limits and applications of what neural networks can learn.

About me

• Master's in ML @  University of Tübingen
• Research @ MPI-IS Tübingen
• Occasional guitarist, bassist, singer, writer
• I enjoy doing (terrible) impressions

Recent work

• (Almost) Free Modality Stitching of Foundation Models ~ pre-print
  Jaisidh Singh, Diganta Misra, Boris Knyazev, Antonio Orvieto.
  → Hypernetworks afford efficient and accurate multi-modal stitching of pre-trained models, shown for models like CLIP and GPT-4.


• Learning the Power of “No”: Foundation Models with Negations ~ WACV 2025
  Jaisidh Singh*, Ishaan Shrivastava*, Richa Singh, Mayank Vatsa, Aparna Bharati
  → We benchmark CLIP via our proposed dataset CC-Neg to show that it does not understand negations in text. Also, CLIP benefits from negation-oriented fine-tuning that improves its negation understanding and overall compositionality.

🔔 Refer to my Google Scholar for a complete list (*: equal contribution).

Pet projects, posts, etc.

• pytorch-mixtures
  A minimalist library for popular MoEs & MoD in PyTorch.
  

• lora-clip
  A library to easily wrap LoRA layer insertion for CLIP.
  

• Tutorial for "DINo: Continuous PDE forecasting with INRs"
  I've recently made an in-depth tutorial on this cool physics-informed ML paper.


• Auto-DEcoder in JAX
  A quick and easy walkthrough of the auto-decoding process with JAX.


• TokenFormer in PyTorch
  A crisp implementation of the TokenFormer Layer in PyTorch.


• Differential Transformer in PyTorch
  A simple snippet for multi-head differential attention in PyTorch.