Shorts
Current shorts include: Tap or Bottled Water? - Quick guide on water choices in the U.S.
Sumanth's Personal Website
Sumanth's Personal Website
Latest posts
Everything about Tokenization
This was a fun project I worked on in November 2023. Tokenization is an oft-neglected topic in natural language processing. With the recent blow-up of interest in language models, I thought it might be good to step back and really get into the guts of what tokenization is. I created a repository to serve as a deep dive into different aspects of tokenization. This was initially supposed to be a…
Everything about Distributed Training and Efficient Finetuning
There’s been an insane amount of interest in large language models (LLMs) these days, with a very special open source community of hackers figuring out the best way to finetune, serve and run inference on consumer-grade hardware. A number of excellent open-source codebases have popped up to meet these needs, notably FastChat , Axolotl and LLama.cpp , with the 🤗HuggingFace ecosystem being at…
Notes on "An Observation on Generalization"
Ilya Sutskever, one of the Co-founders of OpenAI, recently gave a talk at the Large Language Models and Transformers Workshop, Simons Institute, 2023. The talk was recorded, and is available on YouTube . The topic of the talk was “An Observation on Generalization”, and Ilya talks about how we can reason about unsupervised learning using the lens of compression theory. I found the talk…
Integrating $(IA)^3$ with HuggingFace's PEFT
This was a simple course project for the CSE 256: Natural Language Processing course, Spring 2023 at UCSD. Parameter-efficient fine tuning (PEFT) methods are all the rage now, given that they enable you to finetune large models like Falcon-7B on just Google Colab ! A new state-of-the-art method called $(IA)^3$ ( Infused Adapter by Inhibiting and Amplifying Inner Activations ) was proposed…
Surfstore - A dropbox-like service
A fun course project I did as a part of the Graduate Networked Systems course at UCSD. We created a cloud-based file storage service called SurfStore in Golang. SurfStore is a networked file storage application based on Dropbox, that lets you sync files to and from the “cloud”. We implemented the cloud service, and a client which interacts with the service via gRPC. SurfStore has two services: the…
Jester- A text-to-meme generation engine
This project was based on a (serious!) mission to empower everyone to be funny by delegating some of the work to AI. Introducing Jester, a text-to-meme generation engine (try it out here ). The idea is simple: if you’re texting with a friend, it should be very easy to quickly make a relevant meme based on some input text. This involves retrieving a meme template for given user text and then…
Quotes to Ponder On
I’m a sucker for good quotes, especially aphorisms. In fact, I would love to read entire books that are really just quotes from the wisest people in history. Here’s a quote to start things off: “It is a good thing for an uneducated man to read books of quotations.” - Sir Winston Churchill. “The most dangerous worldview is the view of those who have never looked at the…
Moral rules from the ground up
“There is no right or wrong, so I’m going to cheat, steal and lie” is a statement that most people would regard as ‘self-serving’. They’ve got it backwards . Moral standards define how you look at actions, events and intentions as right or wrong. My interest in morality was rekindled in my senior year due to a course called Professional Ethics - one every…
Image Restoration for Under-Display Cameras
TL; DR: We created a novel deep learning based model for image restoration, resulting in a publication at RLQ-UDC Workshops, ECCV 2020. We also placed 2nd /150 teams at the Under Display Camera Challenge. Check out my presentation here ! Under Display Cameras present a promising opportunity for phone manufacturers to achieve bezel-free displays by positioning the camera behind semi-transparent…
Adversarial attacks and model uncertainty
In this project, we wanted to understand how to build uncertainty aware neural networks - for classifying an image as a cat/dog, think of this as giving a confidence interval along with a probability estimate. We started out in a theoretical setting: we looked at how dropout can be used as a bayesian approximation ( source ). The application in mind was adversarial attacks - using uncertainty…
Books
What we consume determines who we are. Our interests provide a window into our psyche, and very few things impress upon us like the way a good book can. These are some of the books I’ve enjoyed reading in the past and some I’m making my way through now. Poor Charlie’s Almanack: The Wit and Wisdom of Charles T. Munger (Personal Favourite ❤️) The MindBody Prescription, by John…