Transformers are here to stay - DSBoost #17
Discover more from DSBoost
Transformers are here to stay - DSBoost #17
Welcome to the 17th issue of DSBoost, the weekly newsletter where you can discover interesting people in the ML/AI world, get the main takeaways of a relevant podcast, and stay up to date with the latest news in the field!
💬 Interview of the week
This week we interviewed Sairam Sundaresan, who is a Research Scientist. Enjoy:
What did you study/are you studying (if your background is different from DS, how did you end up in the field)?
I graduated with a Master's degree in EE: Systems from the University of Michigan. I specialized in Signal Processing and Computer Vision. At the time, Data Science as a field wasn't as well defined as it is today. My professors encouraged me to explore Machine Learning as well since a lot of the research I did in Computer Vision had overlap with ML. That's how my journey began.
What are your favorite resource sites and books (ML/AI)?
Resources are only as good as how much we use them. Also, different people learn better through different media. So with that caveat, here are my favorites:
Books:
Hands-on Machine Learning by Aurelien Geron
Machine Learning, a Probabilistic Perspective by Kevin P. Murphy
Courses:
Practical Deep Learning for Coders from fast.ai
CS 231n and CS 224n from Stanford
Sites:
kaggle.com (forums)
various substacks for ML news
What got you into your current role (portfolio, certification, etc.)?
During graduate school, I got the opportunity to intern at Qualcomm to build a computer vision system. After graduation, the team asked me to come join them full-time. That's how I got my break into the industry. Since then, I've had the opportunity to build many ML and CV models for mobile phones. For the past few years, I've been working as a research scientist at Intel Labs – With a heavy focus on Deep Learning.
What do you enjoy the most in your work?
The fact that nothing ever stands still. If outsiders think keeping up with the field is hard, I'm here to tell you that folks who work on it day-in-day-out also feel the same way. It's incredible how fast things are moving. The challenge that what you work on today might be outdated tomorrow is a good motivator to think deeply about problems and not follow trends blindly. That's what I love.
What tools do you use the most / favorite tools?
Notion as a second brain, Readwise Reader for note-taking, Git for version control, VS Code for coding, and Weights & Biases for experiment management. I'm pretty sure I'm forgetting a few, but these are the ones that come to mind.
Do you use ChatGPT or other Al tools during your work? If so, how do they help you? Do they change your approach to problems?
No, since the work I do is proprietary to the company, I don't use ChatGPT or other AI tools during work. However, outside of work, I love using ChatGPT and Copilot. Together, it's like having a rubber ducky with a cape. I also use ChatGPT for creative brainstorming and Midjourney to generate images for my newsletter when I don't have a doodle to share myself.
What is your favorite topic within the field?
Might be obvious by now, but Computer Vision is my favorite field. In Deep Learning terms, generative models have been amazing to follow. Currently, I'm fascinated by multimodal models, i.e., models that can work with more than one type of input (images, text, video, speech, etc.).
Which one of the recent AI/ML models will have the most significant impact on the industry, in your opinion?
Hard to speculate, given how fast the field moves, but my personal opinion is that Transformers are here to stay. They've been remarkably effective in NLP, and now in vision tasks as well. I feel there's a lot of interesting work to come out of this architecture yet.
What are you currently learning or improving (topics you are interested in nowadays)?
I've been learning about graph models, diffusion models, and how prompting and prompt engineering work. I've also been revisiting some fundamentals as I'm writing an illustrated book for beginners.
What is the biggest mistake you've made? (preferably DS related)
Two mistakes come to mind. Once I joined the industry, I stopped reading research papers. That was a costly mistake and took me a while to correct. The second mistake that comes to mind is that, in my early years, I wanted to use the latest and greatest models for my work. But, the success of a real-world product hinges on a number of other factors – How fast it is, how much power it consumes, a good user interface, and an ability to adapt to the user's needs. The choice of the model we use depends on how these other factors are affected. It's also heavily affected by how much data we have to train the model. I'm glad I made these mistakes early on and learned from them.
What is your most significant achievement? (preferably DS related)
My work on 3D reconstruction was featured in Forbes magazine. Technically, I should say my team's work since a number of us worked closely together to pull this project together. I also mentored a team of space scientists to build a NeRF model to reconstruct our Sun. This work got a shoutout from Alphabet CEO Sundar Pichai and NVIDIA CEO Jensen Huang.
Can you share a fun fact about yourself?
I've written about 40k words in the past three weeks and drawn about 70 doodles. Random right? I'm working on a completely hand-illustrated book on AI for the rest of us. This project started because I love drawing and making AI accessible to a wider audience.
🤖 What happened this week?
Drag Your GAN: Interactive Point-based Manipulation on the Generative Image Manifold
Sam Altman, the CEO of OpenAI is facing a senate hearing over the safety and regulation of A.I
Initial ideas for governance of superintelligence, including forming an international oversight organization for future AI systems much more capable than any today.
🧵 Featured threads
