Yann LeCun on Autonomous Driving: Deep Learning is Obviously Part of the Solution | AI Podcast Clips
00:00:00.000 | (gentle music)
00:00:02.580 | - Elon Musk is confident that large scale data
00:00:11.100 | and deep learning can solve the autonomous driving problem.
00:00:14.160 | What are your thoughts on the limits,
00:00:16.480 | possibilities of deep learning in this space?
00:00:19.000 | - It's obviously part of the solution.
00:00:21.160 | I mean, I don't think we'll ever have a self-driving system
00:00:24.100 | or at least not in the foreseeable future
00:00:25.860 | that does not use deep learning, let me put it this way.
00:00:28.700 | So in the history of sort of engineering,
00:00:32.060 | particularly sort of AI-like systems,
00:00:36.080 | there's generally a first phase
00:00:37.440 | where everything is built by hand.
00:00:38.560 | Then there is a second phase,
00:00:39.800 | and that was the case for autonomous driving,
00:00:41.840 | you know, 20, 30 years ago.
00:00:44.000 | There's a phase where there's a little bit of learning
00:00:46.400 | is used, but there's a lot of engineering
00:00:48.600 | that's involved in kind of, you know,
00:00:50.580 | taking care of corner cases and putting limits, et cetera,
00:00:54.160 | because the learning system is not perfect.
00:00:55.920 | And then as technology progresses,
00:00:59.160 | we end up relying more and more on learning.
00:01:01.600 | That's the history of character recognition,
00:01:03.480 | the history of speech recognition,
00:01:04.680 | now computer vision, natural language processing.
00:01:06.960 | And I think the same is going to happen
00:01:09.560 | with autonomous driving,
00:01:11.720 | that currently the methods that are closest
00:01:16.240 | to providing some level of autonomy,
00:01:18.640 | some, you know, decent level of autonomy,
00:01:20.480 | where you don't expect a driver to kind of do anything,
00:01:24.080 | is where you constrain the world.
00:01:26.400 | So you only run within, you know,
00:01:28.140 | a hundred square kilometers or square miles in Phoenix,
00:01:30.900 | but the weather is nice and the roads are wide,
00:01:34.140 | which is what Waymo is doing.
00:01:35.780 | You completely over-engineer the car
00:01:38.820 | with tons of lidars and sophisticated sensors
00:01:43.060 | that are too expensive for consumer cars,
00:01:44.820 | but they're fine if you just run a fleet.
00:01:46.860 | And you engineer the thing,
00:01:50.260 | the hell out of the everything else.
00:01:51.940 | You map the entire world,
00:01:53.500 | so you have complete 3D model of everything.
00:01:55.920 | So the only thing that the perception system
00:01:57.700 | has to take care of is moving objects
00:01:59.740 | and construction and sort of, you know,
00:02:03.220 | things that weren't in your map.
00:02:05.040 | And you can engineer a good, you know,
00:02:07.700 | SLAM system and all that stuff, right?
00:02:09.140 | So that's kind of the current approach
00:02:11.380 | that's closest to some level of autonomy.
00:02:13.020 | But I think eventually the long-term solution
00:02:15.180 | is going to rely more and more on learning
00:02:18.960 | and possibly using a combination
00:02:20.580 | of self-supervised learning and model-based reinforcement
00:02:24.860 | or something like that.
00:02:26.840 | (upbeat music)
00:02:29.420 | (upbeat music)
00:02:32.000 | (upbeat music)
00:02:34.580 | (upbeat music)
00:02:37.160 | (upbeat music)
00:02:39.740 | (upbeat music)
00:02:42.320 | [BLANK_AUDIO]