MIRAZUL HAQUE

I might be late to the autoresearch party, but I wanted to try a fun experiment applying it to @numerai data. I built a simple agent loop for the Numerai tournament: propose one change, retrain, evaluate, keep the win, revert the loss. A few findings that stood out: •CORR Sharpe went from 0.84 → 1.56 •Moving from the small to medium feature set was the biggest gain •Using all eras hurt vs a more selective training window •A 2-target ensemble captured most of the benefit •Era-wise ranking of predictions was a bad idea Used just my Claude Pro account + local training.

@IAmAYadavally @sigsoft @FSEconf @davidlo2015 Congrats Aashish!

Happy to share that our paper received the "ACM @sigsoft Distinguished Paper Award" at @FSEconf! A lot of people to thank -- my advisor Prof. Tien Nguyen, the reviewers, PC Co-Chairs @davidlo2015 and Prof. Lin Tan, and the awards committee😇

@berraksismann Congratulations Professor!

Very excited to share the news that I received the National Science Foundation (NSF) CAREER Award! Many thanks to my amazing mentors and students. Without them, this wouldn't have been possible. 😊 nsf.gov/awardsearch/sh…

@IAmAYadavally @davidyoung8906 @UTDJonsson Thanks! Aashish.

@davidyoung8906 @mirazece2015 @UTDJonsson Congratulations, @mirazece2015! :)

Some more group photos from the ceremony. Congratulations to @mirazece2015 again! Mirazul also get the doctoral dissertation award from @UTDJonsson .

@davidyoung8906 @UTDJonsson Thank you, Professor. Would not have been possible without your guidance.

@davidyoung8906 Thank you for all the support and guidance, Professor @davidyoung8906. It has been a pleasure to work with you.

@mirazece2015 is leaving for his full-time job at JP Morgan AI Lab. Wish him best of luck for his future endeavors!

@saikatdutta2012 I opted for PP after waiting for more then 1 month. This year, both approval and the card generation procedures are very slow.

People who applied for F1 OPT this year, how long did it take for approval? Did you opt for premium processing?

@saikatdutta2012 @cs_cornell @CornellCIS Congratulations Saikat!

Super excited to announce that I will be joining Cornell as an Assistant Professor of CS (@cs_cornell, @CornellCIS) starting Fall 2024. I am truly grateful to my advisor, mentors, friends, and family, without whom this journey would not be possible!

If you are attending @coling2022, please attend the poster session featuring our work about automatic testing of NLP systems. Brilliant work by @guanqun_yang.

Had an amazing experience attending #CVPR2022. Got to know about novel works on AI robustness and privacy. Simin also presented our work NICGSlowdown (Efficiency attack on caption generators). @CVPR

Attending and presenting at @ICSEconf have been a really great experience. Hopefully the short discussions can lead to future collaborations. Looking forward to @CVPR now. @davidyoung8906

@ayanmaj1992 @IValeraM Congratulations @ayanmaj1992. I will go through the paper. Please let me know if there is any recorded talk regarding this work.

How can we make #ML-based decision-making systems fair, optimal, and stable? Check out our latest #FAccT work showing how unlabeled data can help us out! Joint work with Miriam Rateike, Olga Mineeva, Krishna Gummadi, and @IValeraM. Preprint: arxiv.org/pdf/2205.04790 🧵👇🏽

NICGSlowdown(CVPR 2022): For generative models, we propose this novel attack on Image Caption Generator models to decrease the efficiency of the model. We add perturbation to an image to significantly increase the caption length of the image. Link: arxiv.org/abs/2203.15859 (n/n)

However, both ILFO and EREBA focus on classification models and it is needed to evaluate energy-based robustness on generative models too. (5/n)

I am glad to share that our work NICGSlowdown has been accepted to CVPR 2022. This work has been led by Simin, and the work is advised by Dr. Wei Yang and Dr. Cong Liu. This is our lab’s third acceptance in top-tier conferences on this topic. Here is a research overview (1/n).