I think autonomous driving advocates would do well to look at the history of computer handwriting recognition, an easier technical problem with lower consequences that received significant investment over decades. But it has never gotten good enough to succeed in the marketplace against alternatives.
Why? It never exceeded consumer expectations, which are extremely high for automated systems. Even a correctness rate of 99.9% means multiple errors per day for most people. Consumers expected approximately zero errors, despite not being able to achieve that themselves, sometimes even with their own handwriting!
Because handwriting is made by humans, there is some percentage of it that simply cannot be reliably recognized at all. But people hold that against computers more than other people because computers are supposed to be labor saving devices.
Likewise because roads are made by people, and other cars are driven by people, so a self driving car will never be able to be perfectly safe. But that is essentially what advocates are promising.
That’s especially true if people expect the same level of convenience, especially in terms of time. People speed and take risks all the time when driving, in the name of saving time. I think it’s likely that an autonomous car optimized for safety would also be a car that just takes a lot longer to get anywhere with.
Speed matters. It’s a big reason we all use touch keyboards on our phones instead of handwriting recognition.
Excellent point, stealing that. I work in automotive and an engineer, traveled around the world, think the realistic possibility of self driving cars without major changes in how we make roads, everywhere, is extremely low.
My error rate in recognizing handwriting would be much higher than a free, open-source recognizer. I am very bad at recognizing other people's handwriting. I don't understand your need for 0 errors. Real life is full of errors and imperfections, there is chaos everywhere. Seems like you expect unachivable.
I doubt it. My handwriting is at least average neatness, and stroke based recognition systems still make multiple errors per sentence. It's just a frustrating waste of time and now that we have touch screen keyboards there's no longer any point to handwriting recognition.
The only handwriting recognition system which ever worked correctly with a low error rate was Palm Graffiti. It forced the user to learn a new shorthand writing style designed specifically to avoid errors.
The secret to Palm Graffiti's market success was that it hacked user expectations.
Because it asked users to learn a new way of writing, when the recognition failed, users were more likely to blame themselves, like, "Oh, I must have not done that Graffiti letter right, I'll try again."
But when it came to recognizing regular (i.e. natural) handwriting, users believed inherently (i.e. somewhat unconsciously) that they already knew how to write, and the machine was new, so mistakes were the machine's fault.
While we're sharing anecdotes, my handwriting is remarkably terrible, and the iPadOS Notes app does a good job of transcribing it.
I think this supports the grandparent's point about using the actual strokes, including angle and azimuth, to reconstruct intent.
I was also fairly proficient with Graffiti, back in the day, but I consider that an input method, not handwriting recognition. I was facile with T9 as well.
Analyzing the individual strokes works flawlessly with Chinese and Japanese, where the stroke order is fixed (occasionally with a few variants). If you have the stroke information and the user writes correctly you can recognize characters that even humans would fail to read from the finished glyphs.
That's great, but I would wager that nearly 100% of all writing ever done in human history was done without capturing the strokes while writing. Therefore, while this added accuracy is great, it is virtually useless for most written work.
Isn't that a bit irrelevant? If we are talking about patterns that work well for the user, clearly writing everything traditionally and then going back and taking pictures of everything is a cumbersome process. Writing on iPad or similar is clearly the medium in which this shines, at which point you do capture the strokes.
That only works if you can assume that everybody using the system you're desiging has access to the underlying technology. Sure, if you're desiging some new system (like an autonomous vehicle on a closed loop, controlled system / system purpose built to perform digit recognition as it is written on it, but why wouldn't you just have the user directly input on a keypad) then you'll get a better result, but in the general, real world, case (autonomous vehicle on city streets with other vehicles / recognizing digits from scanned input without the stroke data) then your special case optimization are impossible and for all general practical purposes do not apply, so appealing to their assistance in increasing accuracy doesn't actually do anything to help the system perform better.
While that's true, having the ability to capture strokes now allows machine-learning models to better determine what potential strokes were used to make a specific shape. Just because we didn't have it for everything doesn't mean it's not useful for adding accuracy to the past.
> "handwriting recognition... has never gotten good enough to succeed in the marketplace against alternatives."
There is little market demand for handwriting recognition, and thus little active research goes into it. Not because it is a difficult or problematic technology, but because better alternatives exist that make it irrelevant.
Even if someone were to come up with an absolutely perfect handwriting recognition system, most people wouldn't use it. Why? because the advent of multi-touch screens means that most people can type much faster than they can hand-write anyway.
This is all true today, but it was not true in the past. There was a time in the computing industry when everyone believed that pen interfaces with handwriting recognition would be a crucial enabler for highly mobile computing. Both Apple and Microsoft built major product launches around this idea in the early 1990s.
Oh, absolutely. I remember that era well. But I'm talking about today, of course.
What changed was that touch screens became better. The old capacitive touch screens were clunky, slow, inaccurate. You could put a keyboard on them, but the lag and poor accuracy meant you couldn't really touch type comfortably. Then multitouch came along and made on-screen keyboards much more responsive and accurate.
But also, Blackberry and (pre-smartphone) phones with SMS made people more comfortable with the idea of using keyboards for text entry on handheld devices. And crucially, auto-correct and predictive text entry covered up for accuracy errors and made text entry by keyboard even more attractive.
I wouldn't bet on it but I also wouldn't call it indefensible. Fully autonomous driving is a very complex problem with a very long tail. Being able to drive semi-reliably on American highways doesn't mean that you're almost done, not even close.
An other handicap for self-driving cars is that the problem is effectively harder at the start when the majority of the traffic will still be operated by human drivers who are a lot harder to predict reliably than an other autonomous vehicles.
Beyond that, I strongly believe that software engineering is still ridiculously immature and unable to deliver safe, reliable solutions without strong hardware failovers. We have countless examples of this. We simply don't have the maturity yet, we're still figuring out what type of screwdrivers we should use and whether a hammer could do the trick.
Isn't that just "image de-obfuscation" though? Seems like narrow AI will be able to out-class humans at that in no time. You can generate as much training data for that as you want. Doesn't really require human-type intelligence. Though I guess you might mean that the obfuscation makes the edge cases even harder, which makes sense.
There's like fifty caveats that go along with this statement, but this is the internet so I'm just going to skip all that.
Something like half your human brain is devoted to visual processing.
There's a tendency to think that things like language is what makes the human brain special, or our ability to plan or think abstractly, and we talk about things like "eagle eyes", but the truth is humans are seeing machines with most everything else as an afterthought.
The reason your cat will attack paint spots on glass for hours and flips the hell out about laser pointers is because their visual systems are too simple to distinguish between those and the objects that actually interest them, like insects.
I think it is, actually. Going from raw pixels to objects is the (relatively speaking) easy part. It's the next part (using that for planning and common-sense reasoning) that's the hard part. Machine learning has already advanced past humans in this regard for many classes of problems - which is part of the reason why captchas are getting so hard.
Since the internet places no weight on things like "common knowledge to anyone in the field" or "I took a bunch of classes on the brain in college", here's a random quote from someone at MIT: http://news.mit.edu/1996/visualprocessing
I have a car (Honda Pilot) where the company decided to make the lift gate window too high, probably to accommodate mounting the spare tire inside the cabin. This design makes you dependent on the rear camera for most reverse use cases.
It probably made a lot of sense in the Southern California design center. In Upstate New York, that camera is covered in road spray and salt, and my brain cannot see anything or act effectively without cleaning it. Even after doing that, it will get dirty again after a few minutes of driving.
I’d guess that a least a few dozen people will hurt by this decision.
Take this problem to the self-driving car and things get even worse. You’re going to have a lot of problems with sensor effectiveness that cannot be magically fixed with software.
It sometimes seems as if half the purpose of assistive driving systems serves to compensate for the absolutely horrible sight lines in a lot of newer vehicles.
And I have heard rumors of lobbying going on to get the requirement for rear view mirrors dropped when video feeds are provided to replace the functionality.
Here's my go-to example about the challenges of driving in a Canadian winter.
I was waiting for my bus to work one morning after a large snowfall. The snow clearing crews were hard at work, but the street was effectively blocked by piles of snow, men, and machines.
Yet, my bus arrived on time *driving down the sidewalk".
I am not sure how any self-driving system could have figured that out :)
