MSc VC at Saarland University, BTech CSE at IIITD (2024). Actively looking for PhD opportunities in theoretical DL, dynamical systems, and optimization.
Delhi
Joined July 2020
- Tweets191
- Following1.1K
- Followers88
- Likes11.2K
Madhava retweeted
Transformer generalization theory provides simple formatting tricks for boosting small models on state tracking (here: GPT-5.6 Luna).
Come meet the authors at COLM: Grand Ballroom Poster Location: #132 on Tuesday, 4:30 PM PDT in Poster Session 2.
Madhava retweeted
Hi My grandmother is admitted in Holy Family Hospital in Delhi (share.google/dRiocgvTuR75K7M…) and is in need of AB- blood. If you or anyone you know is around and open to donating that would be very very helpful.
Madhava retweeted
Can Transformer LLMs hide their reasoning from CoT monitors?
tl;dr: Transformers cannot hide complex reasoning, but they might cryptographically obfuscate their CoTs, making it extremely difficult to extract hidden goals.
Oh wow, great morning surprise. KAMIM (arxiv.org/abs/2407.13873) got accepted to ICASSP'25!
Late post but I'm at Interspeech till Thursday! Looking forward to connecting with people working on low-resource and multilingual ASR.
Madhava retweeted
walked past a man playing Elden Ring on the sidewalk in San Francisco. this is the last shred of art + culture left in this city!!!
Preprint and code released for KAMIM - Keypoint Aware Masked Image Modelling
Key idea: Weighing each patch differently according to keypoint density during MIM boosts linear probing performance.
Arxiv: arxiv.org/abs/2407.13873
Code: github.com/madhava20217/KAMI…
(1/n)
KAMIM uses the density of keypoints (obtained from FAST, SIFT or ORB) for patches of an image to weigh each patch differently during the reconstruction phase. This marginally improves fine-tuning performance and significantly boosts linear probing accuracy with a ViT-B.
(3/n)
Never knew I could focus like this. Spent about 10 hours finishing a paper I had been procrastinating over for a long time. For real, sometimes you need to take a step back to move (more than) two steps forward :)
Attending my first conference, @ieee_re, where my first paper (arxiv.org/abs/2404.17842) got accepted into the RE@Next track. Will be in Reykjavik till next Friday and looking forward to connecting with people here 🫡!
Sry but Delhi is better than Blr even though you get firsthand experience of being a patty in an air fryer and can get hit by a car on the footpath. The traffic, waterlogging and the lack of a largely deterministic and widely connected transport system (i.e. metro) is unbearable.
