Elon Musk, whose track record of failed forecasts makes fortunetellers look good by comparison, recently made one of his silliest predictions yet: that advances in AI will enable Tesla's (TSLA) autonomous cars to handle "all modes of driving" by the end of 2019. By all modes, he means drivers will be able to literally sleep at the wheel as their cars take them to their destination.
As a machine learning practitioner, whenever I see Musk make such wild prognostications about AI, two things come to mind. Either 1) this guy doesn't understand this technology very well or 2) he does and is purposely overhyping it to attract investors. Whatever the case may be, this article will attempt to clear up some common misconceptions about what AI can and can't do. I'll try to keep this as non-technical as possible.
Will Tesla's cars be able to handle all modes of self-driving by the end of next year? Elon Musk certainly seems to think so.
Unfortunately, this is simply not possible, at least not within Musk's overly optimistic timeline.
AI is still in its infancy - it can't compete with the human brain, which is many orders of magnitude more efficient, flexible, and adaptable.
This idea was discussed in more depth with members of my private investing community, Small-Cap Research.
Artificial neural networks are just spreadsheets on steroids.
- Computer Scientist John Launchbury
What current AI can do is perform very simple tasks such as Facebook's (FB) facial recognition, Google's (GOOGL) language translations, and Netflix's (NFLX) movie recommendations. While these technologies do exhibit some human-like intelligence, there's no consciousness behind them. They're literally nothing more than sophisticated pattern recognition algorithms.
These algorithms, more commonly called artificial neural networks, are very loosely based on our primitive understanding of how neural connections in our brains work. A neural network with multiple "hidden layers" is called "deep," which is where the now popular term "deep learning" comes from. To gain some intuition about how deep learning works, take a look at this picture:
image input -> neural network -> desired outcome (delta in spatial position)
Neural networks are easy to fool, which is a huge safety issue with self-driving cars. Changing a couple pixels on an image of a stop sign, for example, can trick a state-of-the-art convolutional neural network into labeling it as a speed limit sign. A human driver would never make such a goofy mistake.
The ultimate point is this - until we're able to close the wide gap between AI and human intelligence by developing more efficient, flexible, and adaptable learning algorithms, Musk's dream of fully autonomous driving will remain just that, a dream. Optimistically speaking, developing such algorithms will take not years but many decades. According to one of the world's foremost experts in this field, Yann LeCun, right now we're still "a long way from machines that are even as intelligent as rats."
https://seekingalpha.com/article/415917 ... of=44&dr=1