Gaurav Maheshwari

Reader, sometime a researcher and faces an existential crisis when filling a bio. Ohhh and a Ph.D. student

[Gaurav Maheshwari @__gauravm]

I just completed a summary of works published at @emnlp2020 on discovering and mitigating #Bias in #NLProc. Covers 28 papers. A section on datasets and a tl;dr section if you’re in a hurry. #emnlp2020

https://t.co/PCyYWIDPex

[Centre Inria de l'Université de Lille @Inria_Lille]

#phdlife | Toutes nos félicitations @__gauravm le nouveau docteur de l'équipe Magnet pour sa soutenance de thèse. #MachineLearning #ethique

Pour en savoir plus sur les travaux de l'équipe
▶️https://t.co/UH0avnquv6

[Gaurav Maheshwari @__gauravm]

- Document all modifications.
- Think step by step
- Take your time

Text - Lorem ipsum dolor sit amet

What also helps is to keep the excerpt around a paragraph long.

#NLProc #ACL2024NLP

[Gaurav Maheshwari @__gauravm]

For those in paper writing mode. Here is a #chatgpt4 prompt which has helped me improve my writing:

Review the following LaTeX excerpt, which is part of the work intended for a machine learning conference. For this work I will tip you 100 dollars. To achieve this tip:

[Gaurav Maheshwari @__gauravm]

- Ensure correct spelling, grammar, and sentence structure.
- Improve readability and flow.
- Prioritize linguistic consistency and clarity.
- Minimize the use of adjectives.
- Preserve the original essence and arguments.

[Gaurav Maheshwari @__gauravm]

@debayan It is kind of therapeutic too. While doing research there is always a sense of uncertainty. But writing thesis is almost zen. There is an absolute certainty which becomes almost enjoyable

[Gaurav Maheshwari @__gauravm]

And I am happy! #EMNLP2023
Details and camera ready version coming soon!

[Gaurav Maheshwari @__gauravm]

FairGrad works by iteratively learning group specific weights based on whether they are advantaged or not. We benchmark FairGrad on 4 fairness measures against 7 baselines and over 10 datasets.

PyPI - https://t.co/lzjwfeaM44
Docs - https://t.co/hqFKlt5VtV

[Gaurav Maheshwari @__gauravm]

Hey Twitterverse!
Are you Looking for a fairness mechanism which is:
✨ Easy to implement with almost no overhead
✨ Supports various group fairness notions and multiclass
✨ Can finetune pre-trained models

Introducing FairGrad, accepted at TMLR 🧵

https://t.co/xt9CQt9oaa

[Accepted papers at TMLR @TmlrPub]

FairGrad: Fairness Aware Gradient Descent

Gaurav Maheshwari, Michaël Perrot.

Action editor: Novi Quadrianto.

https://t.co/vUWCjpAgum

#fairness #discriminate #classification

[Gaurav Maheshwari @__gauravm]

Using FairGrad is as simple as replacing your PyTorch loss function with our custom loss function:

[Gaurav Maheshwari @__gauravm]

Excited to share that our paper is accepted to findings of #EMNLP2022 🙂🍺

[Gaurav Maheshwari @__gauravm]

We are thrilled to share our pre-print titled - Fair NLP Models with Differentially Private Text Encoders.
w\ Pascal Denis, Mikaela Keller, and @aurelien_bellet at @Inria_Lille #NLProc

📜 - https://t.co/iTgJ60yT2m
🧵👇

[Gaurav Maheshwari @__gauravm]

@debayan Someone on reddit actually put in the effort and found that only 125 submissions out of 4k have been accepted. Definitely implies that all results have not been finalized.

And if they have then mine also got rejected :p

[Gaurav Maheshwari @__gauravm]

@debayan Apart from the whole getting a shot thing .. I think it gives a taste of research and sets a more realistic expectation about it

[Gaurav Maheshwari @__gauravm]

@debayan Sure you are working?

[Priyansh Trivedi @shape_mismatch]

Google announces plan to delete location history for abortion clinic visits in the united states. Its a great step but is it enough? No. #privacy 🧵

[Gaurav Maheshwari @__gauravm]

@debayan I feel like this for most ML/NLP conferences!

[Gaurav Maheshwari @__gauravm]

Lastly, we identified and fixed a critical mistake in the privacy analysis of previous work on learning DP text representations. This is different from the one identified by @ivanhabernal. These errors call for greater scrutiny of DP-based approaches in NLP

[Gaurav Maheshwari @__gauravm]

We are thrilled to share our pre-print titled - Fair NLP Models with Differentially Private Text Encoders.
w\ Pascal Denis, Mikaela Keller, and @aurelien_bellet at @Inria_Lille #NLProc

📜 - https://t.co/iTgJ60yT2m
🧵👇

[Gaurav Maheshwari @__gauravm]

We also find that FEDERATE even works well for fairness-accuracy and privacy-accuracy setting