First round: We just chatted casually at the start — they asked about the project I'd worked on, how big the network was, how many parameters, how big the data was, that kind of thing. I couldn't really remember most of it. Then we moved on to coding k-means. I had to explain what k-means is first, and after I finished writing it, they asked me to write a better initialization method. They gave me a hint — it was probably k-means++ — but I couldn't write it out.
Second round: They started by asking about things on my resume. Then they got into a modeling question. Around the Waymo car there are a lot of agents, like other vehicles, and they asked how you'd predict the positions of those agents two timestamps into the future. The input includes the previous position history, surrounding environment info, and so on, and the training data has real path trajectories.
They asked how you'd model this — modeling the other cars, not the Waymo car itself — and then went into more detail about the model, things like the standard "what is MHA" (multi-head attention) type of question. One question that was pretty hard: I said that during training you use a causal mask, so at each step you only look at the previous ground-truth positions. But during inference, you don't get to see the "authentic" ground-truth results the way you do in the training data, because every prediction carries some error, so training it this way doesn't actually simulate what inference looks like in reality. What do you do about that? I didn't know.
They told me you could use reinforcement learning, or add noise to the training data.
I felt like I was just brought in to fill a quota...
Discussion
Loading comments…