I’ve spent the past 31 days figuring out how I can move forward, not understanding what to do to move forward.
For the past little bit, I’ve been focusing on building DataGAN. TLDR; is that I’m focusing on using AI to create synthetic data that we can use to train self-driving cars. A lot of my time this month was focused on questioning technical specifics and understanding.
p.s. I’m currently looking for someone who could help me with getting access to hardware i.e. a lab, or cloud compute credits. If you’re interested/know anyone in DataGAN and what I’m doing, please do send an email! Here’s my proposal about what I’m doing with DataGAN that you could forward to.
For context, DC-GANs are not really the best way to train an AI model. Mode collapse is still a legitimate problem that’s present, and it’s what I’ve faced as well. GANs are so sensitive to hyperparameter changes that even after spending ~2 weeks tuning it, I didn’t really get anywhere.
I just eventually dropped DC-GANs and looked for different ways to understand other variations of GAN. I came across W-GANs, a quick video explaining them vs. other GANs (tldr; W-GANs use a different divergence metric known as the Wasserstein metric), a tutorial showing a basic implementation, and got right into building this out.
Long story short, I completely skipped the fact and never questioned why W-GANs work. The only understanding that I did have was that we use a different divergence method to better approximate the distribution of a dataset for training along with changing the discriminator into a critic.
Technical understanding is 🔑 and I only realized that as I was drafting out my goals for next month. Rather than focusing on understanding the overarching concept, go deep. Understand the science behind the concept, that’s how you make real progress.
And what if you don’t have the technical competency to understand it? Then focus on learning. I remember talking to Alex + Andrei about technical interviews and programming, and the biggest thing I realized is the fact that,
Even if I’m a good programmer if I don’t have the bandwidth to think and break down problems, I wouldn’t go far.
I now spend lots of time on DMOJ + binarysearch.com with my friends, making the journey 10x more exciting.
My goals for next month will be to actively reflect more + seek understanding. Stop trying to consume surface-level content; question, research, and validate. Let your curiosity drive you.
DataGAN
Lots of progress made with what I’m doing. I’ve doubled down on the goals that I’m trying to achieve + the contributions that I’m making. TLDR; I’m going to be focusing on creating the first end2end, self-driving car fully trained on GAN inputs. I think the overall problem of GANs in self-driving is very unique and can introduce new ways to scale up + generate data at the masses for self-driving cars.
Oh, and I also wrote an article explaining the science behind DataGAN + my goals. It’ll go into GANs + DC-GANs and how I’m using them to solve a big problem in self-driving.
Next steps right now are focusing on adding end2end inputs to the model. Having the GAN images doesn’t contribute to anything. I want to show that we can create a self-driving car, fully trained on GANs that’s deployable + robust. I’ll be working next on outputting steering values using Deep Convolutional Networks + Transfer Learning and then work my way to other control methods i.e. speed.
other highlights
- did a 24hr AI hackathon sprint w/ Aahaan, Amy, and Richa. We’re focusing on automating the vaccine development pipeline using AI x AlphaFold.
- attended the scale.ai conference and wrote some notes about Waabi and how they’re solving self-driving using SOTA technology + simulation
other reflections
This is the 13th newsletter that I’ve written. This means that it’s been a little over a year since I started this journey on writing newsletter and sharing my progress w/ people. I’ve learned so much more in the last year than the last 14 combined. It’s insane to know that there’s so much in the world that you don’t know and that you’ve only explored the tip of the iceberg. Am super grateful for the people who’ve been following along my progress as a person + going from 0-100, and can’t really wait to see where I’ll be in the next 10 years!
here’s my email, twitter, and website to learn more about me. if you’d like to have a chat with me, here’s my scheduling link as well!