Quote from the transcript:
Q: What have been the most interesting AI challenges that have been solved through this process?
A: The whole self-driving industry is very deeply rooted in machine learning and AI. I remember even in the earliest days of the DARPA Grand Challenges, there was some machine learning (ML) that was happening on the Stanford vehicle for terrain classification.
There was, of course, the big transition where deep learning really took off around 2012–2013, and we greatly benefited from those early breakthroughs on our project. At that time, Google was arguably the only company in the world that was investing heavily in both self-driving cars and modern ML. When convolution neural networks and deep learning really took off, we had some researchers work with our colleagues at Google Brain to adapt some of their work on convolution networks to the task that we had of pedestrian classification. It was amazing the amount of performance gained in a very short period of time. The error rate dropped by about a factor of 100.
Since then, we’ve been seeing results in other areas beyond perception, including prediction, understanding intent, understanding how people interact with each other, decision making, and simulation. Nowadays, there’s hardly any part of our system where we don’t use deep learning. As the state-of-the-art in machine learning and AI has evolved, we’ve been adapting our system to use the most advanced algorithms and pushing on the state of the art in many areas on our own.
One thing we’ve learned is that as you bring in new, more advanced algorithms, having a system that already works really well can be to your advantage. You can bootstrap it on the previous system, and it gives you great training data. It gives you a great baseline for comparison, and allows you to iterate much faster.
I’m interested in the line: “You can bootstrap it on the previous system, and it gives you great training data.” What might that mean in practice?