Ernst Dickmanns, a German scientist who developed self-driving cars in the 1980s
politico.eu
politico.eu
By the early 1990s, they had one that more or less worked, a Pontiac Firebird with self-driving comparable to Tesla's today.
One of the features envisioned at the time:
> It also featured a sophisticated guidance system intended for use with "the highway of the future," where an electrical wire embedded in the roadway would send signals that would help guide future cars and avoid accidents.
So the idea has been around for a long time, approached from different directions.
I wouldn't be surprised if at some point someone tried to train animals to do the driving and achieved modest success.
Somewhat relevant (animals, real time) but I think it failed:
There's also this: https://www.technologyreview.com/s/401756/rat-brained-robot/ Maybe someday horse neurons will control our cars!
This is the first link I found... I could have sworn remembering an article about them teaching sheep dogs to drive their masters home after a late night at the pub.
https://en.wikipedia.org/wiki/Bus_driver
Given Moore's law, I'd be surprised if we ever see another AI summer after the current one peters out.
I feel this exact same thing is happening with autonomous cars - yes, it's possible to get them to be extremely good at recognising the road and surroundings - but the last few percent, those crucial few percent that make the technology actually usable, I don't see those happening for another 50 years at least.
Then after the war, people's risk tolerances get reset because quibbling over a 1/10,000 chance of death seems ridiculous when people have been actively trying to kill you for the last 5 years and a countable percentage of your friends are now dead.
The way global politics is going we probably don't even have to wait 10 years for this.
Now, a new Cold War, that might push technological competition, although you won't get that "reset" of people's attitudes that you're after.
https://www.army-technology.com/features/feature77200/
"[...] every 55,702 barrels of fuel burned in Afghanistan by the US military forces corresponded to one casualty."
So... that would make them exactly the wars in Iraq and Afghanistan. If that were the case we wouldn't have to imagine, we'd have the results in the field already.
So that war must be of a different kind. Not the kind where you already have technical superiority and have 0 incentive to develop it because you can already make truckloads of money by supplying current generation equipment to the front lines. It has to be a war where developing the new tech is the difference between your country existing or not a decade or more from now.
That's a special kind of war. It could be a cold war but it's unlikely to have one of those in the same way the one from the 20th century unfolded. And if that kind of "hot" war is the only one that can bring these improvements I'd rather stick to driving my own car and labeling my own photo library :).
Someone I know frequently relays an anecdote. They were designing a new military helicopter. One of the requirements was that if the pilot was injured the aircraft should be able to return to base and come to a hover. They were trying to figure out how to cut weight. My friend said, well, if we just remove the pilot and all the equipment needed to support them, we'll lose a lot of weight, and the vehicle will be more aerodynamic. His suggestion wasn't taken seriously. That was before Iraq/Afghanistan.
That project got canceled, my friend retired. I have another friend working at the same company. They are building an autonomous helicopter.
*Edit: ok looks like someone else already noticed this: https://news.ycombinator.com/item?id=17600849
Something like a WW or a cold war where you question whether your city will be the next Hiroshima brings you a jump: the nuclear bomb, the ICBM, man in space and on the Moon, and so many other Sci-Fi tech. That's what I meant.
Yes, today we have slightly better autonomous vehicles than a decade ago but this is natural evolution and it relied on progress in so many other (not necessarily war driven) improvements: computers, electronics, etc.
I would rather not see the war that brings you the AI for autonomous machines.
I think he's envisaging a big peer war with mass mobilization ala WW1/WW2, but he imagines such a war would be like a big Iraq/Afghanistan. It wouldn't.
One observation about the Syrian Civil War (and Iraq and Afghanistan) is that large areas of the country became extremely dangerous, because everyone was fighting over them and you often couldn't see the adversary or know who you were fighting. Another is that supply lines were quite vulnerable; forces could hole up in a military base and be relatively safe, but to continue operations in the field, they needed food/ammo/fuel, all of which needed to be transported at significant risk. A third was that whichever force brought security to a region and stopped the fighting there often won political power, because a majority of people don't care who rules them, they just want to not die.
Drones + self-driving supply lines would allow a belligerent to fortify & disperse their own industrial and operations base, well out of contested zones, and then project power at zero risk to their own lives into disputed areas. If the drones are smart enough (i.e. minimize civilian casualties but can easily detect and eliminate belligerents), they're also likely to win political points for eliminating combatants.
In any case, this is unlikely to spell victory for the belligerent who chooses this route. Humans are cheaper than self driving military trucks. You can afford to lose them. Literally-literally.
It's only in our cushy relative world peace situation we find ourselves entertaining the idea of spending a million dollars on a truck is somehow cheaper than losing a ten thousand dollar truck with 6 men on it.
And no, humans are not cheaper than self-driving trucks. They appear so at the beginning of a conflict because wars usually start when there's an excess of humans and a shortage of resources for them all. However, it takes 18 years to grow a human to the point where they can fight in a war, and another year or two to train them. It takes 2-3 years to tool up a factory to produce drones & self-driving trucks (and maybe a decade to get the software right), but once you do, you can produce one every couple days. Assuming you can maintain your industrial & technological base long enough to get that factory up, guess which one is going to win?
The limiting factor for the Japanese in WW2 wasn't planes, it was trained pilots - they had no problem crashing the planes into ships with untrained pilots because those were both abundant resources, but were incapable of fighting a sustainable air war.
https://en.wikipedia.org/wiki/DARPA_Grand_Challenge
The US Army is still putting funding into a variety of autonomous vehicle programs. For example they want to send a convoy of vehicles to resupply a remote outpost without putting human drivers at risk.
Humans have similar problems too. Our intelligence is trained by experience and evolution to operate within certain parameters. More concerning to me is the fact that a human can conceptualize "couch" from a single example. ML algorithims needs to see thousands of couches before they can classify them.
I'm not sure it's true that we conceptualize couch from a single example. We've all seen thousands of couches in many different contexts, homes, schools, doctor's offices, on TV. If you had a person who had never seen a couch or a zebra before, and all you could tell them was right or wrong, it would probably take them (a lot?) more than a single try before they could distinguish between couch and zebra without fail.
The only reason it's easy for us is that we have this giant scaffolding built around zebras as living things that look like a horses and donkeys, and couches as inanimate objects that look like things people sit on and regularly have some pattern on them.
I think people tend to underestimate just how much "training" in the ML sense goes into a human brain. After all, humans spend the first... decade? of life incapable of all but the simplest tasks. That's a decade in which our brains are consuming petabytes of information and processing it constantly.
Watching my children grow up, the way they learn seems remarkably similar to the way computers learn. If you watch a baby learn to move, it is purely an exercise in going too far in one direction and then too far in the other direction, repeated for basically years until they're coordinated enough to move roughly like an adult by the time they're 3 or so. It's the same with words and concepts too, they're just guessing based on things they already know (and the guesses are often waaaay off, because they don't yet know much), but they're constantly filing things away into their frameworks until their frameworks get big enough that this too resembles the way adults learn.
By the time we're adults, the human brain makes ML algorithms look pathetic, but that doesn't take into the account the decades long head start that our brains got.
What project was this?
https://en.wikipedia.org/wiki/Dartmouth_workshop
EDIT: Probably the Summer Vision Project, as Tome said in sibling comment.
The guy the article about says as much.
“I’ve stopped giving general advice to other researchers,” said Dickmanns, now 82 years old. “Only this much: One should never completely lose sight of approaches that were once very successful.”
I really feel for those researchers. Seeing how your life's work is ignored by people who brute-force their ways through the problems you've already solved must be rather depressing.
On the other hand, the article notes that today's AI Summer gets its funding from a constant stream of profitable applications (though less dramatic than the blue-sky hype). The field is less reliant on animal spirits of academia, and has a more reliable source of funding.
I still think things like self-driving cars are far in the future, but we're at a point where ongoing research into e.g. machine vision can be funded by less safety-critical commercial applications.
For a historical account of back-progagation, which itself was 'invented' many times, see [3, 4].
It is also not widely known that regular expressions, one of programmers' favourite tools, were introduced by Kleene in [5]: he was interested in characterising the behaviour of McCulloch-Pitts "nerve nets" (early versions of neural nets) and finite automata.
[1] http://www.alanturing.net/turing_archive/pages/reference%20a...
[2] https://arxiv.org/abs/1404.7828
[3] https://www.math.uni-bielefeld.de/documenta/vol-ismp/52_grie...
[4] http://people.idsia.ch/~juergen/who-invented-backpropagation...
[5] S. Kleene, Representation of events in nerve nets and finite automata, see https://www.rand.org/content/dam/rand/pubs/research_memorand...
I sometimes think the contributions of numerical analysis pioneers is undervalued (eg Wilkinson, Kahan, Golub). Backpropagation is really just reverse mode AD, and has been discovered many times, as you point out.
https://en.wikipedia.org/wiki/Automatic_differentiation
EDIT: [4] works now.
I am especially looking forward to Ray Kurzweil eating his hat when the singularity is delayed indefinitely.
"Oh, powerful deus ex machina yet to come, the one above me is lost, take him into the loving arms of your personality review and do not judge your faithful followers by his threshold."
the same reason all other powerful organized religion is distasteful. Rampant Stasi denunciation and a constant attempt to subdue all opposition voices.
Sure most people dismiss most over the top depictions in tv, movies, and games, but it is surprising how many out there still believe governments and the mega rich have something "close".
A while back, I heard a particularly curmudgeonly interview with Noam Chomsky, just commenting on the state of AI and recent AI programs (Watson, Deep blue..).
Chomsky basically called all these statsistical AI programs brute force parlour tricks, nothing to do with computer science or artificial intelligences. He stopped just short of calling the whole thing fraud. Real AI are things like bird flocking algorithims (did he work on something like this, can't remember). I think the jist was that to be an AI program, a program must embody a theory of intelligence. The theory-less nature of modern statistical algorithims seemed to really peeve him. maybe he's right.
Anyway... I don't remember what point 9if any) i was driving at. I think there are abstract values at play, in how we narrate progress. Since we still don't know what inteligence "is," or even if it is a distinct thing, it's hard to know what the real milestones are. Personally, I like the fully realized, running at scale "proofs," but that probably says more about my ability to understand theory than anything else.
The counterargument is our own brains use brute force parlor tricks to solve problems. We learn by a priori (pre-built models), by experience (reinforcement learning) and viewing and weighing features that we observe against prior experience (NNs/ANNs).
The entire point of ML is to produce approximately the same result a human does. Whether it's a parlor trick or a perfect simulation of the brain is irrelevant. There's no advantage to something novel and more "organic" unless it produces better results.
This is a common problem I see - people are always trying to compare human reasoning to artificial reasoning rather than looking at the output. The output is all that matters.
Entire schools of philosophy disagree completely.
There was also a fantastic article here on HN a while ago [1] by Douglas Hofstadter where he ranted against google translate and similar, by showing that the current state-of-the-art statistical brute force word nearness ML cluster doesn't understand the text it's translating, and therefore will always be lacking, always have cases that it simply cannot solve.
He's basically saying that Weak AI can only get to 99% when it comes to translating, but we would need Strong AI to get to 100%, and Strong AI is probably impossible.
I'm sure they do; it's very much a matter of debate.
> brute force word nearness ML cluster doesn't understand the text it's translating, and therefore will always be lacking
Humanlike reasoning doesn't solve this problem. Understanding does not prevent this abstract "lacking." Strong AI is not suddenly perfect.
> always have cases that it simply cannot solve.
This is an inherent problem with reasoning, not with strong vs weak AI. Which leads us back to:
> Strong AI is probably impossible.
If we measure "Strong AI" as infallible, then yes. If we measure it by "understanding," then no. Which is why I care more about results than the philosophical debate over understanding/consciousness/humanness. 99.999% is acceptable if 100% is impossible.
We will probably exceed human's ability to come up with reliable training data before we reach strong AI.
It seems like a reasonable extrapolation from modern technology. Think about text to speech producing sounds that are hard to distinguish from humans, deepfake producing images that look like they were actually filmed. Those are all weak AI. They don't need strong AI- just tons of labelled data- to produce things that can fool inexperienced humans.
Alice: “Computers aren’t doing real intelligence, because they don’t think the way we theorize we think.”
Bob: “Computers are doing real intelligence, because we don’t think the way we theorize we think.”
But to say that AI "hasn't really come very far" because its technological achievements outweigh its computational side doesn't really make much sense to me.
Progress is slow. Seeing a story like this and saying "wow, and we're only now seeing this in use in practice today" makes an implicit assumption that once something is proved it should become widespread. The original self-driving cars were not practical; it took advancements in computational power and speed to make it so. That's "coming far." We don't need some novel ML approach to advance AI.
People generally assume, likely because of the astronomical rate of quality of life expansion due to technology and engineering in the 1900s, that things keep "going forward" at some rate. Often they think this rate is exponential. I see people citing AI implementation engineering as their primary examples of this. It's frustrating because it's not true. AI is old. That's fine and good and wholesome. But there is no real progress in AI theory, it's all practical engineering. Unfortunately, theory and engineering are often conflated with "technological progress" without realizing the former limits the latter with a very hard ceiling.
However I can also kind of see the argument here that the deep learning ideas don't seem to have such a fundamental idea behind them. We modeled how biology works at a very simple and abstract level. Now we have more computing power (and especially more data) so we could find a few applications.
But to me at least it doesn't look like a solid theory to build upon. It doesn't deliver insights into how things work like Fourier transforms did with signals and Bayes did with probabilities.
That's the popular story behind neural networks. Like any good popular explanation of a STEM concept, it's as useful as it is inaccurate.
To be fair, our own understanding of "how it works" in practice is not necessarily set in stone. Approximating our belief in how it works is the best case scenario, and expecting it to perfectly mirror human reasoning is unrealistic and frankly not all that appealing from a problem-solving stance.
Human reasoning has a lot of built-in buffers and fault tolerance, and it is, itself, a fairly narrow AI.
It would be nice if you quoted honestly.
Neural networks have taken some inspiration from biology, but that doesn't account for the vast bulk of the work. Biology analogies play a much bigger part in popular explanations than they did in the development of the technology.
(edit) Here, for example, take a look at the original paper on perceptrons. Whole lotta math. This having been released during one of the AI flaps, there was also a fair bit of psychology talk. But not much indication that he was just trying to copy a brain and otherwise didn't understand how it works.
https://blogs.umass.edu/brain-wars/files/2016/03/rosenblatt-...
Similar happens with deep learning. They are grounded in theory, most notably a mathematical proof that deep MLPs are, in principle, capable of learning any mathematical function. You just won't see much of that stuff if you aren't actually reading the papers, because most people aren't that interested in vector calculus. Much easier to explain things by analogy.
Also, another interesting fact about fourier transforms- in the old days, computing FTs was too expensive, so physicists used lenses- which perform approximate FTs- to decompose signals into spectra.
> Thus, Gauss' algorithm is as general and powerful as the Cooley-Tukey common factor algorithm and is, in fact, equivalent to a decimation in frequency algorithm adapted to a real data sequence.
And incremental. Sometimes you'll hear people say that deep learning is nothing new, because the first deep neural networks were developed many decades ago, or because CNNs and LSTMs were developed in the late 90s, etc. That mindset glosses over a lot of the incremental developments (e.g., getting a decent handle on transfer learning) that were necessary to get from the initial academic demonstration of the idea to its successful commercial application.
* Out of the lab
* On roads that aren't level
* When the obstacles move
* etc
So computers driving cars on a one lane road could probably do it more efficiently than humans do today on a 3 lane one.
The idea isn't to give cars more space but rather to give them less and use it so efficiently that you actually get even better results then you had before.
I assume you are referring to the often declared demise of Moore's Law. Reports of such demise have been greatly exaggerated: https://ourworldindata.org/grapher/transistors-per-microproc...
This reminds me of a remark, possibly made by P. M. S. Blackett, which went something like this:
"We've run out of money, now we'll have to start to think."
It is possible that stagnation in computing performance will be just the stimulus that the field needs.
In 2016 Google made a system that was better at any person at playing Go. Go has roughly 130,000 moves possible at any point. That's approximately equal to 400 * 2^9.
Moore's law states that the number of transistors (which is a proxy for computing power) in CPUs doubles every year. 2016 - 1996 = 20 years of computing power growth.
In short, after our computing power has increased roughly by a factor of 2 ^ 20 someone made a system that plays a game that's 2 ^ 9 time more complicated than chess. Why is this seen as some giant, surprising leap forward?
For example, a brute force search of a 19x19 Go board to a depth of 20 would yield on the order of 361^20 = 1.4E51 game states. With a reduction in search depth and better algorithms, state-of-the-art engines might cut this down by ten orders of magnitude, but can still be beaten by rank amateurs.
Deepmind's approach to board game engines blows all previous approaches out of the water. The claim that their success is incidental to Moore's law is categorically false.
We had 20 years of doubling of computing power before Lee Sedol match. In those 20 years there were many other AI programs that have beaten various world champions at other board games (and no one cared). There were other good Go engines before Alpha Go. They would beat most human players in the world.
Why AlphaGo of all other programs is seen not as increment, but as some giant leap forward? It doesn't solve a new class of problems and it doesn't use any fundamentally new algorithms.
It's embarrassing that I never knew about PROMETHEUS given that it was a €749m pan-European collaboration on autonomous driving.
>Our vision system relied (heavily, not exclusively) on sensing prominent horizontal features symmetric about a common centerline. (Cars and trucks, especially at the time, have a lot of horizontal lines: bumpers, window top/bottom, valence, etc.)
Thread here: https://news.ycombinator.com/item?id=10328687
Bridges also have many hoizontal features symmetric about a common centerline. I would expect many false positives driving near bridges.
The van from the vid came years after the two mercedes s cars. I think there was no breaking for obstacles, exiting the highway, or even using interchanges. It was simply holding a lane (maybe change it when told to) and holding a set speed.
Correct me if I am wrong, as I did not read into the experiment.
Not interesting, and frankly, takes away from the dramatic results of recent self-driving cars.
https://news.ycombinator.com/item?id=10333126
The had auto-breaking, car detection and lane merging the least.
I think it's safe to say that while this was pioneering work, none of the technology it used was acceptable in terms of safety or generalizability.
I pointed that those functionalities existed and presumably sometimes worked, as opposed to OP that said they did _not_ have them. At the very least, the false positive proves they had emergency breaking.
What I find amazing is what they managed to do with much less powerful HW than what we have today, with much more primitive camera systems. To put it into context, the project is older than I am and they managed to achieve things that most of my live until pretty recently I would have categorized somewhere between amazing and impossible.
My concern is that this article massively overstates the results that Dickmanns had in a way that implies his work had any real chance of being used in a production environment. There is a substantial difference in the sorts of Probabilistic systems that Thrun and others build today.
Sad that new AI is not doing the deeds on checking prior art..