Jay Shah is a graduate student at Arizona State University focusing on deep learning and computer vision for medical imaging. However, besides his research, he regularly has researchers and engineers on his podcast (links below). He has already had a lot of great people from all of the top institutes and companies as guests which made me feel particularly honored when he reached out to invite me, as well. We chatted for more than 2 hours so that Jay splitted our conversation into two episode, with the first one already on YouTube:
All other episodes from Jay's podcast can be found on the below platforms:
What we talked about
After a short introduction, we started talking about some of my PhD research:
- The trade-off between adversarial robustness and accuracy as well as on-manifold adversarial examples as discussed in our CVPR'19 paper.
- How the insight that adversarial examples tend to leave the underlying manifold led to the development of confidence-calibrated adversarial training in our ICML'20
- Neural network quantization and the robustness of quantized networks against bit errors in weights. This was the topic of our MLSys'21 and TPAMI'22 papers on bit error robustness for DNN accelerators.
We then pivoted to my experience at Google DeepMind, differences between industry and academia and my internship project at DeepMind:
- Conformal training integrates conformal prediction and deep learning to obtain intuitive uncertainty estimates alongside performance guarantees — as detailed in our ICLR'22 paper.
Finally, we talked in depth about some career-focused topics:
- Preparing for research engineering interviews in industry — related to this blog article;
- Finding the right PhD program — and deciding whether to do a PhD at all — as I discussed in this article;
- The importance of citations and other metrics in academia;
- And how to decide between academia and industry.
I had a great time chatting with Jay and I hope listeners can get value out of the podcast. While this was my first podcast appearance, I definitely had a great time and appreciate that Jay got in touch in the first place.