Tejas Duseja

Applied Scientist @ Amazon. He/ His

[Tejas Duseja @TejasDuseja]

@alokbishoyi97 Heyy Alok. Working on post training of SLMs at amazon for agentic use cases. Would love to connect :)

[Tejas Duseja @TejasDuseja]

@BLRAirport I misplaced my blue Tommy Hilfiger jacket (size M) at Wolfgang Puck, Terminal 2 departure area, around 3:00 AM on Oct 31. Please let me know how I can check with Lost & Found or if it’s been turned in. Thanks!

[Teknium 🪽 @Teknium]

I think Meta and Llama-3 is the final nail in the coffin to several misconceptions I've been fighting against for the last year.

Llama-3 Chat was trained on over 10M Instruction/Chat samples, and is one of the only finetunes that shows significant improvements to MMLU. Contradicting several claims:

- That finetuning can't teach a model new knowledge, MMLU is a wide variety huge dataset of knowledge QA, improvement of over 3pts in MMLU strongly shows that it does indeed add new knowledge

- That LIMA paper (ironically by Meta) claim, that "10k" samples is best you can do to teach a model to do things with, completely destroyed by this.

I've been arguing these things to people for a year, with a lot of pushback, but the evidence is clear, though I'd argue it was clear long prior from Hermes work.

[Fermat's Library @fermatslibrary]

"An experiment is a question which science poses to Nature, and a measurement is the recording of Nature’s answer" - Max Planck

[Andrej Karpathy @karpathy]

The Transformer is a magnificient neural network architecture because it is a general-purpose differentiable computer. It is simultaneously:
1) expressive (in the forward pass)
2) optimizable (via backpropagation+gradient descent)
3) efficient (high parallelism compute graph)

[Boris Dayma 🖍️ @borisdayma]

Time to talk about the biggest mistake I made while training DALLE-Mega 😥

[Alex Dimakis @AlexGDimakis]

Someone is trying to scam my PhD student. My student asks to verify their identity 1/2

[hardmaru @hardmaru]

MANGA sounds better than FAANG

[Fermat's Library @fermatslibrary]

George Pólya on teaching John von Neumann

[@emilymbender.bsky.social @emilymbender]

Hey #NLProc (and AI folks working on language), can we just ... not?

https://t.co/Sf5tojVU3X

[elvis @omarsar0]

📘 Probabilistic Machine Learning: An Introduction

I have been looking for a book like this. Kevin Murphy published the 2021 edition of the Probabilistic Machine Learning e-textbook. Love the emphasis on probability and math. It includes code examples.

https://t.co/nK3AYNMP1z

[David Pfau @pfau]

Regardless of anything else, I think being told that I couldn't publish a major paper that I'd been working on for ages, totally out of the blue, and with no reason or recourse, is like a top ten wake-up-in-a-cold-sweat nightmare scenario for me.

[Luca Ambrogioni @LucaAmb]

The dynamics of stochastic variational inference is a balance of three forces.
One pushes the samples towards the likelihood, the other towards the prior and the third pushes the samples apart to avoid the collapse into the MAP estimator.

[Scott Linderman @scott_linderman]

Our first post is by @ZhaoYixiu on Structured VAEs—a paper by @SingularMattrix, @DavidDuvenaud, Alex Wiltschko, @ryan_p_adams, and @Datta_Lab. It's an elegant idea but tough to implement... until now! JAX and TensorFlow Probability made it way easier! 2/3 https://t.co/iSpBLRmGEL

[Ehsan Adeli @eadeli]

Deadlines are amazing! I really do not understand how papers come together in the last moment! 3 days ago, I thought we are not submitting many of the papers we were writing. 10 min before the deadline, the paper looks so good as if we could not do it any better. #CVPR2021

[Itay Evron @itayevron]

Why does L1 regularization induce sparse models?

Many illustrate this using the least squares problem with a norm constraint. The least squares level sets are drawn next to the different unit "circles".

I prepared a cool animation which I believe makes it even clearer 🙂

[MIT CSAIL @MIT_CSAIL]

READ: 130-page cheat sheet outlining key #MachineLearning concepts: https://t.co/JFfnxgY4n6
———
#ML #BigData #AI #DeepLearning #Algorithms #MondayMotivation

[Cristian Garcia @cgarciae88]

Deep Mind recently came out with Long Range Arena, a benchmark suite to pit Efficient Transformers against each other in Long Range Tasks 🏃🏃‍♀️🏃‍♂️.

Do we finally have a winner?

Paper: https://t.co/4Dm3NU11tl

[Stat.ML Papers @StatMLPapers]

Kernel Methods for Policy Evaluation: Treatment Effects, Mediation Analysis, and Off-Policy Planning. (arXiv:2010.04855v1 [econ.EM]) https://t.co/DJ3w0Dirkn

[Arsenii Ashukha @senya_ashuha]

TLDR; BatchNorm prevents BYOL to learn collapsed predictions.

https://t.co/tHiwoG1gNV