![karthik viswanathan Profile](https://pbs.twimg.com/profile_images/1883239673360744448/Gt6CM6T8_x96.jpg)
karthik viswanathan
@nickinack1
Followers
189
Following
29K
Statuses
161
🧑💻 ai @M42Health, ex: cs @iiit_hyderabad, @AdobeResearch, @IRIC_umontreal, @iitmadras, foundation models + genomics. avid badminton fan 🏸
abu dhabi
Joined October 2020
@suzuki2001_ we did finetune on both of these models, but the results were below par. moreover, in models like NT, extensive task-wise probing was required, which almost defeats the purpose of foundation models, if some of our random models can perform better in some tasks
0
0
2
RT @BrandesNadav: New preprint claims that most existing DNA language models perform just as well with random weights, suggesting that pret…
0
65
0
our new work @M42Health with @kirill_vish
Genomic Foundationless Models: Pretraining Does Not Promise Performance 1. This study challenges the paradigm of pretraining in Genomic Foundation Models (GFMs), revealing that randomly initialized models often match or surpass pretrained ones in fine-tuning tasks. 2. Despite significant computational investments in pretraining, the study finds marginal gains (typically 2-3%) for pretrained GFMs, questioning their cost-effectiveness compared to simpler baselines. 3. The analysis highlights that randomly initialized models like Caduceus and DNABERTv2 outperform larger pretrained models in critical genomic tasks, including functional element classification and mutation sensitivity. 4. Pretrained GFMs show limited sensitivity to sequence variations, such as single nucleotide polymorphisms, undermining their utility for clinical applications like pathogenicity prediction. 5. The findings advocate for a shift in focus from scaling pretraining to designing biologically informed tokenization and task-specific architectures, akin to AlphaFold’s targeted success in protein structure prediction. 6. This work underscores the need for new benchmarks that emphasize clinically relevant tasks and deeper biological understanding, steering genomic model development towards practical and impactful applications. @kirill_vish 💻Code: 📜Paper: #Genomics #MachineLearning #FoundationModels #Bioinformatics #ComputationalBiology
0
0
3
congratulations 🇮🇳🇮🇳 thank you sagar, amruta and harshit; wholesome chess streaming for the past two weeks
22nd September, 2024 is a red letter day for Indian Chess - this is one of the biggest moments in Indian Chess History. Team India wins Double Gold medals in the Chess Olympiad 2024! The Indian Open and Women's Team both secured Team Gold medals, with many more Individual Board medals to come. Team India (Open) needed just a draw today for Team Gold, but they defeated Slovenia 3.5-0.5 anyway. Gukesh, Praggnanandhaa and Arjun scored wins on the top 3 boards, and Vidit had a long and fighting draw on board 4. The Indian Open Team secured first place with a whopping 21/22 Match points! Team India (Women) had to win against Azerbaijan today, and they delivered with full force. Harika, Divya and Vantika won, while Vaishali drew her game. India defeated Azerbaijan 3.5-0.5! If Kazakhstan won against the USA today, there would've been tiebreaks involved - but that match ended in a draw. The Indian Women's Team secured first place with 19/22 match points! A huge congratulations to all the players, coaches and the entire team involved. What a day! Graphic: Anmol Bhargav #chess #chessbaseindia #ChessOlympiad
0
0
4
@BadmintonJust we do not need a WD coach who used to think WD was boring and started watching WD only after they retired
4
2
10
@Anmolkakkar27 @BadmintonJust @ankhitweets @BAI_Media i’m prepping yall @ankhitweets @BadmintonJust for kal ka match.
2
0
1
@kyleshew The deductions on executions over the years for successive CoPs have been increasing, and there is little to no incentive for a gymnast who has flawless execution, and this is concerning. We like seeing higher scores and lowering the average score at every CoP, affects the sport.
0
0
1
@saynotomodern @riebum wdym it should be “reserved for elite players”? elite players are created by winning such medals. this comment completely contradicts itself in all possible ways.
0
0
4