Swapnil

Senior AI Scientist @Intuit | AI Sec + Alignment Researcher | Open Source @ vLLM | prev @ : IBM, NYU

[Swapnil @SwapnilDreams10]

@namanambavi Congratulations bro!

[Swapnil @SwapnilDreams10]

@niraj_yagnik Love this!

[Swapnil @SwapnilDreams10]

@Dr_Faruk_Kazi Congratulations sir!

[Dialogue & Discourse @dnd_journal]

Vol.14 No.1: Automatic Essay Scoring Systems Are Both Overstable And Oversensitive: Explaining Why And Proposing Defenses,
by @YamanKSingla, @SwapnilDreams10, @someshsingh22, @jessyjli, @RatnRajiv, @cchangyou
https://t.co/6u324aS40S

[Swapnil @SwapnilDreams10]

Big thanks to @someshsingh22,@YamanKSingla,@KbalajiTweets,@cchangyou,@RatnRajiv for the opportunity, and for making the process super enjoyable and educative!

[Swapnil @SwapnilDreams10]

Really exciting news! My first research work accepted to #AAAI2022 on data-free adversaries. We successfully attack and fool SOTA NLP models on a myriad of tasks like Sentiment Classification and Language Inference. And the best part, no actual data needed, just the model!

[Somesh @someshsingh22]

No need for data to craft adversaries anymore! 🧐 Check our paper "MINIMAL: Mining Models for Data Free Universal Adversarial Triggers" @RealAAAI #AAAI2022. Extremely grateful for the 4 accept reviews as this will be my first A* publication in #NLProc!

https://t.co/pDkDCyWNfI

[arXiv Daily @Arxiv_Daily]

AES Are Both Overstable And Oversensitive: Explaining Why And Proposing Defenses
https://t.co/Qr9pYcnb9c
by Yaman Kumar Singla et al. including @RatnRajiv
#DeepLearning #Computation

[arXiv Daily @Arxiv_Daily]

MINIMAL: Mining Models for Data Free Universal Adversarial Triggers
https://t.co/h5WTzzfnoH
by Swapnil Parekh et al. including @RatnRajiv
#Computation #Language

[Swapnil @SwapnilDreams10]

Our new work w @YamanKSingla and @midasIIITD on novel training data-free NLP adversarial triggers which are dangerous phrases which can flip model predictions for all inputs of a particular class! Additionally they transfer well across architectures and datasets.

[arXiv Daily @Arxiv_Daily]

MINIMAL: Mining Models for Data Free Universal Adversarial Triggers
https://t.co/h5WTzzfnoH
by Swapnil Parekh et al. including @RatnRajiv
#Computation #Language

[Yaman Kumar Singla @YamanKSingla]

Here we show algorithm for mining universal adversaries from just models w/o any requirement of data. So for universal adversarial attacks, now you just need to make a few API calls and you can find triggers which perturb output of *any* input to the model.
#NLProc

[Rajiv Ratn Shah @RatnRajiv]

Unable to register my startup with #startupindia. It's more than two weeks since I am getting "CIN verification Service is not available currently."
Without this, no startup can get tax benefits which is due by 31 March. Could you please help @startupindia @PiyushGoyal @DIPPGOI?

[Swapnil @SwapnilDreams10]

@that_laminu @TimIrven @CanaanInc1 +1 would love to get more insight into the economics

[Swapnil @SwapnilDreams10]

@justice1996V @YamanKSingla @jessyjli @RatnRajiv @mile_research yes, this article was one of our key inspirations! thanks

[Yaman Kumar Singla @YamanKSingla]

Read about how essay scoring models treat essays as "word-soups".
Selecting just 30% words of a human-written essay score original grade(+-1)

Joint work with @jessyjli , @SwapnilDreams10, @RatnRajiv , @mile_research

[Swapnil @SwapnilDreams10]

@niraj_yagnik Was looking for it thanks

[Swapnil @SwapnilDreams10]

@EdwardDixon3 @huggingface Isn't t5 already included in hugging face?

[Swapnil @SwapnilDreams10]

@listenMrUtkarsh Probably a sign you don't like it very much. Im sure u are coding even if u make errors. Idk

[Swapnil @SwapnilDreams10]

@NAILsocial @garyvee Man people will take any excuse to put others down, even if it makes 0 sense. You lack context dude

[Swapnil @SwapnilDreams10]

@listenMrUtkarsh iisc blog when

[Swapnil @SwapnilDreams10]

@listenMrUtkarsh You are in Mumbai?