François Chollet: "There is no realistic amount of data that covers 'everything you can encounter' in the real world."
This is a fundamental question; it’s about whether self-driving cars are possible with the current machine learning paradigm.
If driving is a surfeit of new situations — say, one every 5,000 miles — then self-driving cars aren’t possible (without any fundamental advancements in AI). By “new” I mean situations that an imitation learning or reinforcement learning agent couldn’t generalize to from its training dataset, even if that dataset included 100 billion miles’ worth of situations. (And that no existing hand-coded heuristic could handle.)
If new situations are incredibly rare — say, one every 1 million miles — then self-driving cars are probably possible (without any fundamental advancements in AI).
Humans are able to generalize to new situations, and self-driving cars using present-day ML aren’t. But most crashes are caused by humans making the same relatively few mistakes over and over. So, if self-driving cars can mostly avoid those mistakes, that could more than compensate for their inability to generalize to new situations if new situations are rare enough.
That’s a reasonable expectation. Would it also not be unreasonable to expect an autonomous vehicle that encounters a new situation to fall back on some kind of fail safe behavior. Human drivers often do this instinctively to avoid trouble, like turning away, slowing down, or stopping. It seems that the same kind of response is possible from a fully autonomous vehicle, even if it is not necessarily the optimal response, and might still result in an accident. It’s unreasonable to expect perfection of AVs, but then people can be unreasonable.
Welcome to the forum! I suppose the first worry is that generic fail safe behaviours won’t be enough to avoid danger in all situations. And the second worry is that if self-driving cars go into fail safe mode too frequently — say, once every five trips — then you need a huge number of remote operators, and issues with latency and cell service may limit how much remote operation is even possible.
I’ve been reading the posts on your site. Really informative, and no BS. Wish I could contribute more, but I’m a total novice.
Thanks! I recently wrote an explainer article on Medium you might find helpful. Don’t worry about being a beginner. I’m sort of a beginner myself — there’s a lot of stuff I don’t know about machine learning.
Your article was very helpful. If this competition were a game of chess, Tesla’s strategy and execution would constitute a checkmate.