Hacker News Reader

Comment by mgalgs | Hacker News Reader

> Not just image recognition, obstacle avoidance, rule following, etc.

I would argue that the examples you gave are just image recognition, etc. What is your brain doing if not image recognition and obstacle avoidance?

We're not there with today's technology but eventually our machines will perform better at those uniquely human tasks you mentioned. After all, we humans and our cognition is just the output of these giant meat computers inside our skulls pulling in sensory data and running pattern recognition algorithms.

The $65,000 question is whether or not today's technology is sufficient (however barbaric) to be acceptable for the job.

View on HN

> I would argue that the examples you gave are just image recognition, etc. What is your brain doing if not image recognition and obstacle avoidance

Planning, anticipation, reading body language, etc? Is that person looking like he’s going to jaywalk? That’s not image recognition. Yes, once he steps into the road the self driving car will respond quicker. But the human driver might already be slowing down and watching for hints of initiating movement. Image recognition is just one tiny piece of it.

> That’s not image recognition.

How do you do it? Telepathically? You're still just receiving visual input and processing it.

"Image recognition" is a small subset of "receiving visual input and processing it." We know how to do image recognition with machine learning. Planning and anticipation is a much harder problem.

Human knowledge and intuition. Deductive reasoning.

Knowing there is a location ahead without a crosswalk that kids constantly cross at.

Or knowing that the car in front of you is going too fast, implying that they're unaware of the railroad tracks they're about to jump.

You don't need telepathy to make useful predictions. And humans don't need to be accurate, they just need to be prepared to take a measured response.

> Knowing there is a location ahead without a crosswalk that kids constantly cross at.

That's not based on reasoning. Most people driving by probably don't even know. And it might be in the car's database.

> Or knowing that the car in front of you is going too fast, implying that they're unaware of the railroad tracks they're about to jump.

If I'm behind them, and I'm going much slower because I know the layout of the road... I'm not sure how knowing this matters?

Simply track all the pedestrians all the time, and you'll know in advance where the common crossing points are, and when any pedestrians are moving towards a road where your car is moving. Apple, Google, Facebook, have every advantage here.

Tesla track where every Tesla goes, all the time. Will Apple make a car before Tesla makes a 'smartphone'?

Actually there are powerful techniques to predict what a person will do based on current position, velocity, body language and other cues. They can be trained by trying to predict future actions from past state, and self driving cars apply these models to all objects - people, cars, other things on the road.

Even if (and that's a BIG if),...

There are still a nearly endless number of things the car can't get from on-board sensors, but that people will know, or at least be able to learn.

No, it is 'model if the world' vs. image recognition. There will be no self-driving cars before the former is available, i. e. for a very long time.

Reply on news.ycombinator.com