Comment by dragontamer | Hacker News Reader

Comment by dragontamer | Hacker News Reader

Parent Full thread

There's absolutely no equivalence between a CNN created today with deep learning, compared to the Neural Network in a human brain.

The problem with the human brain is one of inattention. CNNs / Artificial Neural Networks can remain at attention 100% of the time due to their artificial / machine nature.

But CNNs, despite being at 100% attention the entire time, still have issues determining if that splotch on the screen is the sky or an 18-wheeler.

https://electrek.co/2016/07/01/understanding-fatal-tesla-acc...

--------------

Artificial Neural Networks / Convolutional Neural Networks have a very long way to go before they reach human equivalence. In contrast, sensor systems or LIDAR brute-forces the problem. LIDAR can see things human's can't see, and advanced sensors can tell you (at least, in clear conditions) the location and velocity of virtually every object around the car.

Fake-it till you make it camera-only driverless cars are clearly hype that relies upon a fundamental misunderstanding. Just because CNNs are kinda-sorta like the visual cortex of the human brain doesn't really mean that it works like one.

CNNs have really cool visual learning properties. But I've yet to see one 100% successfully tell you background vs foreground in pictures like a human brain can do. Even in clear weather conditions, the CNN can confuse a truck for the sky and still run full speed into an 18-wheeler.

I explicitly wrote my comment because @Animats said "machine learning system", not "CNN", which I know definitely not enough about to comment on.

Surely the field of "machine learning" includes things which are not even invented yet, the same way maths comprises of theorems not yet discovered.

I was just saying that I think one could "just hook up some cameras to a machine learning system, train it, and you have self driving", maybe not today though.

The "neural network" terminology is really cringeworthy all things considered. ("Deep learning" is much better.)

I dunno. I think "auto-optimization" best describes the process.

Even calling it "learning" is kinda overselling itself. There really are only two camps of "Learning": auto-optimization (against a trained dataset), and auto-categorization (aka self-training).

Its auto-optimization: the algorithm self-corrects itself to try and look more like the training-set's "ground truth". Or auto-categorization, as the algorithm looks for patterns and tries to draw its own categories.

"Learning" implies model finding. Which... strangely enough... I'd argue that 3-SAT solvers are more "learning" based, at least with colloquial use of the word. Those things really do craft new theories and test them through the process of elimination / resolution / etc. etc. "Neural Networks" explicitly DON'T do this however.

Your standards are even higher than mine. I salute you for that.

Have you looked into Hubert Dreyfus yet?

The 100% attention thing I why I have tremendous faith that computers will be out-driving humans in short order. Even really good human drivers have a limited perspective and focus. And even the best human drivers are subject to emotional state.

Yeah, humans suck at attention. But until we can program computers to tell me what is a foreground moving object, what is a shadow, and what is the background... you literally can't solve the "Should the car apply the brakes right now" problem.

https://arxiv.org/pdf/1801.02225.pdf

Just go through this recent paper I found on DDG.gg and see the amount of effort it takes to parse foreground / background data that a self-driving car needs.

You gotta figure out if there's a still-object on the road. And whether or not its a shadow (shadows don't move after-all but its safe to drive over). Like, neural networks can't do that stuff 100% reliably yet. And it may never happen.

Or some researcher next year might come out and discover a method to parse background / foreground / shadows out of pictures. But then there are a whole host of OTHER issues involved.

------------

I think self-driving technology has potential. But you need more than just neural networks hooked up to cameras. I really like Waymo's direction with advanced super-human sensors. Avoid the shadow / background / foreground problem entirely and just have LIDAR give you the precise coordinates of all objects within 100-feet of the car.

IMO, if self-driving technology ever happens, it will be because of advances in advanced LIDAR or other kinds of sensors. Stuff that can avoid the research-problems that the "Deep Learning" community hasn't been able to solve for the last 50 years.

That's why I think the future lies with a mix of (possibly layered) sensing and decision-making technologies. LIDAR + cameras + downloaded maps + machine learning + whatever.

The "until we can program computers to..." is a when question, not an if question. There's nothing about driving, in any situation, that doesn't fall in the face of "assuming infinite computing power and infinitely good sensors". Driving isn't a creative act, it's a responsive one.

I think that's reasonable, although our time estimates for when this problem solved may disagree.

My main issue is that a large number of people seem to think that cameras + machine learning are enough to solve this problem. And while I'm not an expert at machine learning, what I know about it makes me a pessimist. There's just too many unsolved problems in the machine-learning community to apply machine learning to the car-driving problem.

Machine learning probably can solve weird cases people don't expect. IIRC, CNNs are better at recognizing blurred or garbled text than humans these days. So CNNs can read speed limit signs, road signs, and other texts and and at least process that.

Even figuring out if its a speed limit sign, an address, or a route-number probably can be solved by CNNs. But higher level reasoning (is that spraypaint messing up the signpost?? Which was common in some of the areas I drove through) seems like an unsolved problem.

----------

Anyway, Machine Learning + Cameras are IMO, at best... a partial solution to some of the problems. Anyone who thinks that cameras + radar + machine learning is sufficient is probably just a TSLA long who wants to believe in the success of their stock. Otherwise, I think most people are reasonable and recognize the importance of experimenting with a ton of different methodologies to solve this problem.

Humans suck at these classification tasks things too, though, if you set your expectations high. Is it a cat or a bag, or a weird shadow? Doesn't matter, I ran it over before I could figure it out and react. Or maybe it turned out to be a boulder? Now I'm one of 35,000 traffic fatalities this year in this country.

Computer vision has been improving rapidly in the last 10 years, I think it's too soon to rule out the viability of a camera-based solution entirely. Though I do hope improved lidar technology can improve on humans.

Expectations for CNNs with Tesla's tech is currently "Does not run into firetrucks". And they're not exactly succeeding right now.

https://www.wired.com/story/tesla-autopilot-why-crash-radar/

Humans are way, way better than current CNNs on this field. We can talk about cats, shadows, and boulders when CNN-based methods stops crashing into concrete barriers, parked fire-trucks, and 18-wheelers making a left turn.

I don't want to dismiss the work of Deep Learning / Machine Learning specialists. I just want to point out that the problem is incredibly difficult. It is very far away from being a solved problem.

> Traffic-Aware Cruise Control cannot detect all objects and may not brake/decelerate for stationary vehicles, especially in situations when you are driving over 50 mph (80 km/h) and a vehicle you are following moves out of your driving path and a stationary vehicle or object is in front of you instead

This is a known issue, a known pattern and has happened multiple times this year. Its repeatable. CNNs today are not working in this case, and fixing it will require a research effort of mammoth proportions.

You're talking about what the tech does. I'm talking about what it can do, with reasonable-sounding non-magical improvements to tech. Just because there are limitations now does not mean those limitations will still exist in ten years.

Umm, Radar?

Comma.ai is doing vision+radar right now and driving cars with it.

Reply on news.ycombinator.com