From Caffe to AutoML: Jia Yangqing’s Journey and Insights on AI’s Future

01. Life Experience

Alibaba Interviewer: When did you start programming and what did you study at university, graduate, and PhD levels?

Jia Yangqing: I started programming relatively late, in middle school when I first encountered a computer. I wanted to study computer science at Tsinghua but the score was too high, so I entered the automation program. My undergraduate courses were standard; I only began learning machine learning during my graduate studies. I later pursued a PhD at UC Berkeley focusing on computer vision, and my first serious C++ project was Caffe.

02. Deep Learning Technology

Alibaba Interviewer: What motivated the development of Caffe and what were the biggest challenges?

Jia Yangqing: We needed a general platform for research similar to cuda-convnet, which was hard to extend. Caffe was built to solve our own pain points in training and designing deep networks, and we open‑sourced it because many others faced the same problem. The toughest issue was debugging; initially we lacked tests, but after adopting testing frameworks we could run code and verify it quickly.

03. AI Research Directions

Alibaba Interviewer: Will AI achieve breakthroughs in common‑sense reasoning, and how do knowledge graphs and graph computing fit in?

Jia Yangqing: I’m not a graph‑computing expert, but I believe knowledge graphs (or expert systems) represent a promising direction for AI. While perception tasks like image, speech, and language understanding have advanced, integrating logical reasoning remains unsolved. Graph computing reminds me of probabilistic graphical models from my PhD, which I think are worth investing in.

Alibaba Interviewer: Do you see a future for deep learning models that require no large training sets?

Jia Yangqing: Yes, especially in domains like medical imaging where data is scarce. Small‑data learning is still largely research‑focused, but it holds potential for real‑world applications.

Alibaba Interviewer: Should technical peers invest in deep learning?

Jia Yangqing: Treat deep learning as a tool. If you can experiment on weekends, it’s worthwhile; otherwise, focus on areas where you can add value, such as compiler development or data processing.

04. Career Growth

Alibaba Interviewer: How should algorithm engineers choose their development direction?

Jia Yangqing: AutoML emerged around 2016 and initially required massive resources. A realistic research direction is to find the best model under limited resources, essentially improving search efficiency.

Alibaba Interviewer: What differences do you see between Chinese and Silicon Valley engineers?

Jia Yangqing: Chinese engineers are very diligent, often working late, while Silicon Valley engineers may leave earlier and are more willing to experiment with risky ideas. Learning from their tooling and code‑review practices can boost efficiency.

05. Future Outlook

Alibaba Interviewer: What are the main challenges for AI to impact humanity?

Jia Yangqing: The biggest challenge is that many AI researchers stay in the “ivory tower” and don’t focus on deployment. Real‑world data is messy and large‑scale, requiring robust engineering and software practices to bring algorithms to production.

Alibaba Interviewer: How should one choose a long‑term research direction in the knowledge‑explosion era?

Jia Yangqing: Follow your genuine interests; popularity doesn’t guarantee fulfillment. Even if a field isn’t lucrative, passion can sustain you.

Alibaba Interviewer: Any book recommendation for developers?

Jia Yangqing: Dale Carnegie’s “How to Win Friends and Influence People” – it teaches logical thinking about human interaction, which is valuable for technical professionals.