With AlphaGo in 2015 we crossed an important bridge in the Turing Test - before AlphaGo, we could be pretty certain that a high level Go game was played by a human. After AlphaGo, we can't be sure.
At this point a display of great competence no longer means that a human did it. Sure; all the parts haven't been completely connected to make self driving cars, general game playing AIs or conversation bots. But on any discrete task it is no longer reasonable to say "computers can't possibly do that" the way I could about a Go game between two 9 dan players in 2010.
I can still say "not commercially viable" and I can point to specific attempts that don't work, but computers are now on the same threshold as humans - [data + time = results]. It may be more data and more time than a human, but that is a big change from [logically modelled domain = results] which is where we were before in AI.
> ...AI systems whose actions cannot be closely scrutinized and explained...
Author is one of the large group of people who are in for a shock when they try to scrutinize and explain a humans crazy actions. I can't even explain why I get the wrong result sometimes when I add numbers in my head.
I've only got a minute, but I think it was helpful for me to look at ML achievements in terms of input, output, and difficulty.
A human brain can learn to recognize cat pictures given a relatively small amount of input versus a convolutional neural net.
A human brain hits a limit somewhere in terms of attention span, though - I used this to understand AlphaGo's accomplishment. AlphaGo is not inherently better at solving complex problems compared to a human brain, but it is demonstrably better at focusing on one complex problem until it has amassed the equivalent of thousands of years of training.
I'd guess that the human inability to concentrate on one problem that long is actually a clue. Humans come up with new go and chess insights continually. Maybe it's because we're bad at evaluating the new ideas objectively for such complex problems, but it may also be a clue that the human brain has a more efficient method to develop new ways of analyzing complex problems.
The best machine-based deep learning approaches still need to crunch through an enormous data set. Especially so for unsupervised learning.
>A human brain can learn to recognize cat pictures given a relatively small amount of input versus a convolutional neural net
To give the ANN a fair go, I think you should pretrain on another dataset while evolving the architecture. Let that chug for, say, 500 million years. Then refine its performance on your cat dataset and see how long it takes for that step.
I think there's another level of representation that needs to happen for NNs to learn the appearance of an object in a similar way that humans do. I've seen a lot of research recently on compressed representations, I think applying that to object recognition might yield some interesting results.
"A human brain can learn to recognize cat pictures given a relatively small amount of input versus a convolutional neural net." is not really accurate, this is an apples and oranges comparison.
A human brain that's already learned to see and recognize objects can learn to recognize cat pictures from a small amount of input data, and a neural network that's trained to see and recognize objects (which explicitly didn't include cats) can also learn to recognize cat pictures from the same amount of input data. One-shot learning or few-shot learning is a thing and gets reasonable results.
And a human brain that's learning to see "from scratch" (i.e. infants) can't learn from a few pictures, it needs huge amounts (many months) of rich image data - i.e. not passive but interactive/experimental, which is known to be neccessary for visual system development in mammals (I recall a bunch of cat experiments) and is known to have much, much better sample efficiency for artificial systems.
So as far as I know there's no good reason to assume that human brain can learn cat pictures from less input than neural networks. What we're seeing and misattributing to the "learning capabilities" here is actually the effectiveness of transfer learning, transferring knowledge and skills everyone's learned as a kid to new (but actually very related) problems.
>> ...AI systems whose actions cannot be closely scrutinized and explained...
> Author is one of the large group of people who are in for a shock
> when they try to scrutinize and explain a humans crazy actions. I
> can't even explain why I get the wrong result sometimes when I add
> numbers in my head.
I was just rereading Asimov's I, Robot, and was struck by this passage:
"The cotton industry engages experienced buyers who purchase cotton. Their procedure is to pull a tuft of cotton out of a random bale of a lot. They will look at that tuft and feel it, tease it out, listen to the crackling, perhaps, as they do so, touch it with their tongue, and through this procedure they will determine the class of cotton the bales represent. There are about a dozen such classes. As a result of their decisions purchases are made at certain prices, blends are made in certain proportions. Now these buyers can not yet be replaced by the machine.
"Why not? Surely the data is not too complicated for it?
"Probably not, but what data is this you refer to? No textile chemist knows exactly what it is that the buyer tests when he feels a tuft of cotton. Presumably there's the average length of a thread, their feel, the extent and nature of their slickness, the way they hang together and so on. Several dozen items, subconsciously weighed out of years of experience. But the quantitative nature of these tests is not known. Maybe even the very nature of some of them is not known. So we have nothing to feed the machine. Nor can the buyers explain their own judgment. They can only say, 'Well, look at it. Can't you tell it's class such and such?'"
Asimov wrote this in the 1940's and in this passage he tried to illustrate how unquantifiable and impenetrable human judgment was compared to artificially intelligent robots, which he saw as ultimately rational calculating machines. Ironically, today AI techniques such as neural networks are criticized for making much the same sort of impenetrable judgments.
Asimov also assumed we wouldn't have AI at all if we couldn't mathematically prove certain things about it, and instead we are handing over deadly tasks to AI without any knowledge of how it works.
Yes Asimov assumes AI would be based on logic, he was not aware of neural networks which I think had not been discovered yet. It should have been clear though that if we grow brains in vats they might resemble human brains greatly.
Well, these things would still need to be represented as features, or some sort of input to perform feature extraction on, and we might still not know what that input is.
"USDA’s classing methodology is based on both grade and instrument standards used hand-in-hand with state-of-the-art methods and equipment. The system is rapidly moving from reliance on the human senses to the use of high-volume, precision instruments that perform quality measurements in a matter of seconds."
The key part here is the statement "So we have nothing to feed the machine" which, as we now know, has turned out to be false.
We don't need people to be able to tell you how and why they made that decision, we can treat the human decision-makers as a 'black box' and learn to effectively copy the behavior of that black box without necessarily needing to understand how and why it makes these decisions. We don't need textile chemists to know exactly what it is that the buyer tests when he feels a tuft of cotton if we can replicate the process of "subconsciously weighed out of years of experience" by having a machine look at enough expert decisions that amount to literal years of experience.
This is called being blindsided by ego. There is nothing special about the human brain. It's a mechanical device cycled through evolution, which has probably happened countless times in the universe. If we make an AI which is to the human brain as a fighter jet is to walking, we would just contributing to that evolution.
I’m not sure that ML’s success in AlphaGo generalizes well.
For example, Transformer architectures may be able to generate seemingly realistic text, but after a couple of paragraphs a human can pick out the incoherence.
Furthermore ML still struggles with out of distribution (ood) samples which humans are able to navigate easily.
> ... but computers are now on the same threshold as humans
This parity is still extremely limited and far from general. Computers were developed in the first place to handle fast computation with ease. What we are seeing with ML is an extension of that computation by encoding complex information (such as images) extremely well. This is still in the same computation plane as calculators (performing operations on floats and integers), but forgive my pun, only deeper.
Human cognition is much more than computation. What we have today with machine learning is still very much computation.
We're not sure, but it's a hypothesis that's still a long way from having solid evidential support.
And there's no begged question about appreciated sunsets. The link between emotional and aesthetic perceptions is one of the key problem areas.
Because aesthetic perception includes immediate sense data, but also links it with cultural contexts, personal memories, and sometimes a degree of artistic improvisation to communicate all of the related experiences, memories, and associations.
It's precisely the difference between recognising a sunset with an image classifier and experiencing it with (say) another human that is far beyond current systems.
There is so much to human cognition that researchers haven’t or have been unable to address in AI.
We can tell from our memories that when we experience something, it is not simply raw perception. In fact our System 1 minds are filtering out most of our sensory input so that even if two people are watching the same sunset, they’ll notice it differently.
And on top of that, our memories are infused with so many other details that are not sensory input but artefacts from our lived experiences, personalities and cultural heritage. So that now even if two people are noticing similar things, they’ll still experience the sunset differently.
And next time you are looking at a sunset, try to gather of all your thoughts and explain precisely how you are experiencing it. That is hard.
So yes even a simple begged question like that shows the differences between human and computer cognition enough.
>With AlphaGo in 2015 we crossed an important bridge in the Turing Test - before AlphaGo, we could be pretty certain that a high level Go game was played by a human. After AlphaGo, we can't be sure.
So? That was true for Chess at one time, and Checkers or tic tac toe back in the 1900s. It's just a more computationally advanced version of a game with simple rules, not a general intelligence...
The Turing Test has nothing to do with chess or any other game. It's a conversational test between two parties. It doesn't refer to human-level performance on any specific task. Please check Wikipedia.
> Sure; all the parts haven't been completely connected to make self driving cars, general game playing AIs or conversation bots.
This implies that connecting these systems is just a detail to be worked out. But the systems have nothing in common ontologically---they share no inputs, outputs, or intermediate concepts. The fact that they both use neurons doesn't mean much. You may as well say we can connect snails and dogs.
> Connectionism is at heart a correlative methodology: it recognizes patterns in historical data and makes predictions accordingly, nothing more. Neural networks do not develop semantic models about their environment; they cannot reason or think abstractly; they do not have any meaningful understanding of their inputs and outputs.
And the symbolists make the same mistake they did the first time. Imbuing programs with nicely-named symbols and hard coded logic does not possess them with understanding, it arrests away their ability to learn it.
Simple programs do not understand themselves, they have no more awareness of the logic they are running than a mouse has awareness of its neurons. Symbols and their programs represent understanding only as much as they map to concepts in the programmer's mind, not their own. Understanding must be something that partakes in computation, not the definition of the program.
Classical chess AI, built with perfect chess simulators, idealized search and expert heuristics, despite flowing through interpretable and semantically sensibile programs that humans have built, are entirely isolated from the semantics of their programs—for the names and the layout of the data structures are not properties that the program itself has any access to.
Despite the limitations of machine learning, this cannot unreservedly be said of neural networks, which are demonstrably extracting semantically meaningful latent spaces from highly complex inputs as part of their computation. ML is still not self-reflective in any useful sense (it cannot hear itself think; it does not see itself learn), but at least it is handling the first level of the task on top of which we might conceivably build understanding. To MuZero, a game of chess is an aspect of its network that it could perceive, at least within a given branch of its search. And we know these networks must be building something at least knowledge-analogous, since how else could a network like GPT-2 answer questions (however imperfectly) across a range of out-of-domain tasks, like knowledge retrieval and translation?
And this is why Gary Marcus' position (besides his repeated telling of false claims) misses the point. Yes, we should embed programs with priors and reasoning beyond brute connectionism—and to that, most people agree;—but this understanding cannot live in the symbols, it must by necessity live in the structure of the computation, and this structure must itself be accessible to the AI. It is this latter thing that the ML community is already doing, in likely the majority of ML papers, of which the convolutions Gary likes so much are just the tip of the mountain. It is this latter thing that explains, much against the grain of Gary's claims, why MuZero is a better network than AlphaZero.
It seems like everyone says the problem is "lacking of understanding" but it doesn't seem like anyone "understands understanding" on the level that this seems to require.
>"Yes, we should embed programs with priors and reasoning beyond brute connectionism—and to that, most people agree;—but this understanding cannot live in the symbols, it must by necessity live in the structure of the computation."
I'd argue it needs to live in the computation and live in the symbols and be able transparently go between these. But how to make that work is still unknown.
I'd argue neural networks need to evolve to a stage where they create symbolic representations and invent a language that allows neural nets to communicate and learn from each other. That would probably require that they also develop a symbolic representation of 'self', as in my knowledge vs. what is communicated to me by other neural beings via some language we all can use to exchange symbolic representations with.
We should view AI in the context of the technological revolution. You all have a good approximation of what might happen. The question is what to do.
I believe that since the population size of intelligent agents will grow exponentially fast and pretty soon, one should consider one of the following options
i) Do pretty much the same as before.
ii) Join a FAANG company.
iii) Do PhD+ level research in ML/CS/AI.
iv) Make money as fast as you can.
v) Some option I didn't think about.
If you pick (i), it seems likely that your kids (if any) will have no money or/and power to remain relevant. Pick (ii)+(iv) or (iii)+(iv). The latter seems preferable but it is harder.
The thing most people don't realize is that advances in AI are going to (at least for the medium term future) produce high powered tools rather than true autonomous agents. The people who rise to the top of the new economy enabled by AI will be those that learn to take full advantage of the tools, and who have those traits that machines lack, i.e. creativity and vision.
>> who have those traits that machines lack, i.e. creativity and vision
I would go further and say that once machines will be able to lie better than humans, we are done.
>> produce high powered tools rather than true autonomous agents
including tools which will help to tell a compelling story and/or lie to people. We can see precursors of these tools appearing during the last few years.
(ii) sounds like a version (iii)+(iv) but it is not because you will not be truly free. If you go (iii)+(iv) via a start-up line, you will not be free either because you will have to deal with customers.
Should come up with something smart and make money. Let's call "be truly smart" (what a smarter AI would do) and not just "academic smart" or "coder smart".
The thing is, "AI" as it exists today in the form of frameworks, libraries, APIs and the data behind it all, sit firmly inside products as features. Sure you can build a platform that is focused on making those features easier to build and implement into products - and sell that as a PaaS, but at the end of the day it will go into a product that people use.
Otherwise it's pure research and effectively a cost center. There is no standalone AI/ML etc... capability that makes money on it's own. It's all where it's integrated with existing consumer or enterprise products.
Understanding that, whomever builds and scales product adoption, are the ones who are really controlling the direction of AI. If you look at who are building and scaling consumer and enterprise products it's really only a handful of players.
So if you're a successful startup with a good product that uses AI driven features, you're either gonna get acquired or a major (FAANG etc...) will just copy and scale your product way faster than you can. So again, all roads lead to ii).
Can symbolism be defined as a reductionist approach? Did symbolism fail because it did not acknowledge computational irreducibility of intelligent systems?
Are connectionist / neural network models commercially more successful than symbolist approaches have ever been? Or is the current AI decade comparable to the 80s culminating in expert systems?
Much more successful. (and still perhaps just as over-hyped. :-). Most Google searches you do were influenced by a neural network prediction. The ads you see from Google and Bing were predicted by statistical machine learning models. When you talk to your phone or Google home or Alexa or Siri, the speech to text is performed by a deep neural network. When you (don't get as much) spam in your inbox? DNN. When you search your iPhoto album for cats? DNN. List goes on and on.
DNNs are one of the best tools we have for bringing uncertain, hugely complex real world audio, text, and visual data into a form where we can manipulate it symbolically or mathematically with traditional programming. In other words, they bring more domains into the scope of automation. (Observe that none of what I just said sounded like "AI" as popularly imagined).
We still have a huge gap in the middle between symbolic AI and machine learning. Basically, what the cerebellum does - manage short-term activity. "Common sense" at the low level.
Not much going on in this area. Which is why robot manipulation in unstructured situations still sucks, after 50 years.
Yeah someone made a big claim about it being increadibly hard for robots to deal with cloth. It's true it is tricky, but the shirt folding problems has been solved for a while.
So in the video, the user still had to manually insert each piece of clothing into the machine individually. I would have just folded it myself at that point and it would have been quicker too. I thought the whole point was for human not to have bother with this meaningless non-value added task.
I consider true AI is possible only when it is able to read through an entire novel or book and then provide an accurate summary of its story or plots that not only makes sense, but correctly conveys the emotional aspect of those books. Remember a lot of meanings or sentiments from the stories in those books are implicit, and so they do require a certain level of emotional processing to interpret it correctly.
It's quite astounding when you think about how much of our life has been automated or made virtual. But basic house hold tasks like cooking, dishes, tidying up, folding clothes are still fully manual. Also shopping for food, vacuum or tending your garden (for the majority of us). Tasks which all together take many hours each week.
Think about all the time going to waste. Yes, there are tasks which some actually want to keep manual (maybe you love cooking) and there are (subpar) online service, which try to replace that labor, but there is no automatic option otherwise. And outsourcing labor is not what I mean. I want a washing machine at home, which is a lot simpler to handle than outsourcing it to a washing business.
I sometimes wonder, if the reason is really a technical one or if there are other factors, social or incentive problems, which hold progress back in those areas...
I mean, I'm not very good at folding shirts anyway.
Maybe the solution is lower standards?
Reminds me of some story about a person doing something and another person saying "that looks painful, what's the trick", with the answer being "not minding that it hurts".
Granted I have the naivety of someone who doesn't work in the field, but I think the gap was recently narrowed, no? The resemblance of the modern 4-layer LSTMs to traditional LL(K) parsers had me on the floor for days.
That's interesting, I did some traditional parsing work a dozen years ago, but had not noticed any obvious similarities like the one you mention, can you elaborate on that resemblance?
I think the recent moves in game playing AI (e.g. AlphaGo, Pluribus) are beginning to bridge that gap. Specifically, the use of concrete tree search algorithms (e.g. MCTS, Minimax) combined with neural networks to intelligently guide the search.
I suspect this sort of algorithm will be applied to robot manipulation in the next couple of years, and would be extremely surprised if DeepMind et al weren't already working on it.
It makes sense. Machine learning can only learn what it has already seen. Combining with connectivism from a neural network, it may come up with brand new concepts through linking existing ideas and elements. But since these new concepts are also new to itself as well, would the machine know how to interpret it correctly? Connectivism alone may help keep generating new links and connections, but to interpret those new elements correctly we would need something more than just a neural network.
Remember there is a very fine line between creativity and illusion. Technically there is not much of a difference because both are a distortion of reality. Even we humans have a hard time distinguishing this sometimes as many new inventions used to be deemed crazy at first. It looks like when it comes to concepts that we hadn't seen, we were also susceptible to the same flaw just like those machines.
I don't really know much about symbolic AI but can it really solve this problem? If it can then the human game may really be over!
Excellent historical overview. However, I am astounded with this idea that human intelligence is explainable. It is not. We do not know why people decide how they decide. The explanation is always a narrative created subsequently.
> The explanation is always a narrative created subsequently.
Only in the most tautological sense - explanation has to be made in a form of narrative in order to communicate it. As for the correctness aspect, I recall reading on HN recently that the study which claimed to demonstrate that such explanations are wrong had some bad issues with their data science.
"....The old-fashioned car manufacturers said 'We believe in electric motors too, and we can derive electric motors by injecting petrol into the engine."
There is a persistent lure to symbolic AI, and it is the lure of anthropomorphization, of thinking that AI must be smart in the way humans are smart, by manipulating symbols.
But natural language is not so much a central expression of machine intelligence as it is of human intelligence. Humans confuse linguistic aptitude with intelligence. Symbolic manipulation will be at best an API by which machines can relate to humans as the machines get smarter and smarter. Symbolic tools that conform to the bandwidth limitations of humans, which are constraints that our current machines don't face. (Historically they did, and symbolic AI made more sense then, since neural-net training was unfeasible.)
The lure of symbolic AI puts us in the bondage of old ideas, as Keynes would say. Symbolic AI is the equivalent of replacing our economy's fiat currency with an arbitrary supply of a yellow mineral.
Natural language is high dimensional and hard to do convolutions with. That’s why you need dimensionalaty reducing embedding layers. There isn’t anything magical about symbols, their space just has less structure than simple numbers.
Neural network fads come around every 25 years (1960s - killed on purpose by Minsky by his book Perceptions. Minsky wanted everybody to work on his type of AI, symbolic AI which went nowhere. Late 1980's. 2010s.) They usually die out, leaving no trace behind. Maybe this time it's different, imho because of Google translate. Most neural networks don't actually earn any money. But in general they make super slick demos, that's for sure .
Daniel Kahneman has an interesting thought on this: “I am puzzled by the number of references to what AI “is” and what it “cannot do” when in fact the new AI is less than ten years old and is moving so fast that references to it in the present tense are dated almost before they are uttered. The statements that AI doesn’t know what it’s talking about or is not enjoying itself are trivial if they refer to the present and undefended if they refer to the medium-range future—say 30 years." See edge.org for more.
I think a lot of the confusion comes from the fact that everybody defines "AI" differently.
To me, AI = self consciousness, anything less is just fancy ML.
If you ask the question "are we going to have self-conscious machines in 30 years" I would bet against it. If anyone has any reason to bet _for_ it, I'd like to hear those reasons :)
This article downplays some benefits of symbolic approaches. For example, a set of logical statements can benefit from a SAT solver, and a concrete object can benefit from a relational database or a search through a graph, and theorems can be proved symbolically.
68 comments
[ 1.7 ms ] story [ 168 ms ] threadAt this point a display of great competence no longer means that a human did it. Sure; all the parts haven't been completely connected to make self driving cars, general game playing AIs or conversation bots. But on any discrete task it is no longer reasonable to say "computers can't possibly do that" the way I could about a Go game between two 9 dan players in 2010.
I can still say "not commercially viable" and I can point to specific attempts that don't work, but computers are now on the same threshold as humans - [data + time = results]. It may be more data and more time than a human, but that is a big change from [logically modelled domain = results] which is where we were before in AI.
> ...AI systems whose actions cannot be closely scrutinized and explained...
Author is one of the large group of people who are in for a shock when they try to scrutinize and explain a humans crazy actions. I can't even explain why I get the wrong result sometimes when I add numbers in my head.
A human brain can learn to recognize cat pictures given a relatively small amount of input versus a convolutional neural net.
A human brain hits a limit somewhere in terms of attention span, though - I used this to understand AlphaGo's accomplishment. AlphaGo is not inherently better at solving complex problems compared to a human brain, but it is demonstrably better at focusing on one complex problem until it has amassed the equivalent of thousands of years of training.
I'd guess that the human inability to concentrate on one problem that long is actually a clue. Humans come up with new go and chess insights continually. Maybe it's because we're bad at evaluating the new ideas objectively for such complex problems, but it may also be a clue that the human brain has a more efficient method to develop new ways of analyzing complex problems.
The best machine-based deep learning approaches still need to crunch through an enormous data set. Especially so for unsupervised learning.
To give the ANN a fair go, I think you should pretrain on another dataset while evolving the architecture. Let that chug for, say, 500 million years. Then refine its performance on your cat dataset and see how long it takes for that step.
A human brain that's already learned to see and recognize objects can learn to recognize cat pictures from a small amount of input data, and a neural network that's trained to see and recognize objects (which explicitly didn't include cats) can also learn to recognize cat pictures from the same amount of input data. One-shot learning or few-shot learning is a thing and gets reasonable results.
And a human brain that's learning to see "from scratch" (i.e. infants) can't learn from a few pictures, it needs huge amounts (many months) of rich image data - i.e. not passive but interactive/experimental, which is known to be neccessary for visual system development in mammals (I recall a bunch of cat experiments) and is known to have much, much better sample efficiency for artificial systems.
So as far as I know there's no good reason to assume that human brain can learn cat pictures from less input than neural networks. What we're seeing and misattributing to the "learning capabilities" here is actually the effectiveness of transfer learning, transferring knowledge and skills everyone's learned as a kid to new (but actually very related) problems.
"The cotton industry engages experienced buyers who purchase cotton. Their procedure is to pull a tuft of cotton out of a random bale of a lot. They will look at that tuft and feel it, tease it out, listen to the crackling, perhaps, as they do so, touch it with their tongue, and through this procedure they will determine the class of cotton the bales represent. There are about a dozen such classes. As a result of their decisions purchases are made at certain prices, blends are made in certain proportions. Now these buyers can not yet be replaced by the machine.
"Why not? Surely the data is not too complicated for it?
"Probably not, but what data is this you refer to? No textile chemist knows exactly what it is that the buyer tests when he feels a tuft of cotton. Presumably there's the average length of a thread, their feel, the extent and nature of their slickness, the way they hang together and so on. Several dozen items, subconsciously weighed out of years of experience. But the quantitative nature of these tests is not known. Maybe even the very nature of some of them is not known. So we have nothing to feed the machine. Nor can the buyers explain their own judgment. They can only say, 'Well, look at it. Can't you tell it's class such and such?'"
Asimov wrote this in the 1940's and in this passage he tried to illustrate how unquantifiable and impenetrable human judgment was compared to artificially intelligent robots, which he saw as ultimately rational calculating machines. Ironically, today AI techniques such as neural networks are criticized for making much the same sort of impenetrable judgments.
It hasn't aged that badly.
"USDA’s classing methodology is based on both grade and instrument standards used hand-in-hand with state-of-the-art methods and equipment. The system is rapidly moving from reliance on the human senses to the use of high-volume, precision instruments that perform quality measurements in a matter of seconds."
[1] https://www.cottoninc.com/cotton-production/quality/classifi...
We don't need people to be able to tell you how and why they made that decision, we can treat the human decision-makers as a 'black box' and learn to effectively copy the behavior of that black box without necessarily needing to understand how and why it makes these decisions. We don't need textile chemists to know exactly what it is that the buyer tests when he feels a tuft of cotton if we can replicate the process of "subconsciously weighed out of years of experience" by having a machine look at enough expert decisions that amount to literal years of experience.
For example, Transformer architectures may be able to generate seemingly realistic text, but after a couple of paragraphs a human can pick out the incoherence.
Furthermore ML still struggles with out of distribution (ood) samples which humans are able to navigate easily.
> ... but computers are now on the same threshold as humans
This parity is still extremely limited and far from general. Computers were developed in the first place to handle fast computation with ease. What we are seeing with ML is an extension of that computation by encoding complex information (such as images) extremely well. This is still in the same computation plane as calculators (performing operations on floats and integers), but forgive my pun, only deeper.
Human cognition is much more than computation. What we have today with machine learning is still very much computation.
Are we sure? And don’t say “computers can’t appreciate a sunset” or any begged questions like that.
And there's no begged question about appreciated sunsets. The link between emotional and aesthetic perceptions is one of the key problem areas.
Because aesthetic perception includes immediate sense data, but also links it with cultural contexts, personal memories, and sometimes a degree of artistic improvisation to communicate all of the related experiences, memories, and associations.
It's precisely the difference between recognising a sunset with an image classifier and experiencing it with (say) another human that is far beyond current systems.
There is so much to human cognition that researchers haven’t or have been unable to address in AI.
We can tell from our memories that when we experience something, it is not simply raw perception. In fact our System 1 minds are filtering out most of our sensory input so that even if two people are watching the same sunset, they’ll notice it differently.
And on top of that, our memories are infused with so many other details that are not sensory input but artefacts from our lived experiences, personalities and cultural heritage. So that now even if two people are noticing similar things, they’ll still experience the sunset differently.
And next time you are looking at a sunset, try to gather of all your thoughts and explain precisely how you are experiencing it. That is hard.
So yes even a simple begged question like that shows the differences between human and computer cognition enough.
So? That was true for Chess at one time, and Checkers or tic tac toe back in the 1900s. It's just a more computationally advanced version of a game with simple rules, not a general intelligence...
> Sure; all the parts haven't been completely connected to make self driving cars, general game playing AIs or conversation bots.
This implies that connecting these systems is just a detail to be worked out. But the systems have nothing in common ontologically---they share no inputs, outputs, or intermediate concepts. The fact that they both use neurons doesn't mean much. You may as well say we can connect snails and dogs.
For example...
And the symbolists make the same mistake they did the first time. Imbuing programs with nicely-named symbols and hard coded logic does not possess them with understanding, it arrests away their ability to learn it.
Simple programs do not understand themselves, they have no more awareness of the logic they are running than a mouse has awareness of its neurons. Symbols and their programs represent understanding only as much as they map to concepts in the programmer's mind, not their own. Understanding must be something that partakes in computation, not the definition of the program.
Classical chess AI, built with perfect chess simulators, idealized search and expert heuristics, despite flowing through interpretable and semantically sensibile programs that humans have built, are entirely isolated from the semantics of their programs—for the names and the layout of the data structures are not properties that the program itself has any access to.
Despite the limitations of machine learning, this cannot unreservedly be said of neural networks, which are demonstrably extracting semantically meaningful latent spaces from highly complex inputs as part of their computation. ML is still not self-reflective in any useful sense (it cannot hear itself think; it does not see itself learn), but at least it is handling the first level of the task on top of which we might conceivably build understanding. To MuZero, a game of chess is an aspect of its network that it could perceive, at least within a given branch of its search. And we know these networks must be building something at least knowledge-analogous, since how else could a network like GPT-2 answer questions (however imperfectly) across a range of out-of-domain tasks, like knowledge retrieval and translation?
And this is why Gary Marcus' position (besides his repeated telling of false claims) misses the point. Yes, we should embed programs with priors and reasoning beyond brute connectionism—and to that, most people agree;—but this understanding cannot live in the symbols, it must by necessity live in the structure of the computation, and this structure must itself be accessible to the AI. It is this latter thing that the ML community is already doing, in likely the majority of ML papers, of which the convolutions Gary likes so much are just the tip of the mountain. It is this latter thing that explains, much against the grain of Gary's claims, why MuZero is a better network than AlphaZero.
>"Yes, we should embed programs with priors and reasoning beyond brute connectionism—and to that, most people agree;—but this understanding cannot live in the symbols, it must by necessity live in the structure of the computation."
I'd argue it needs to live in the computation and live in the symbols and be able transparently go between these. But how to make that work is still unknown.
I believe that since the population size of intelligent agents will grow exponentially fast and pretty soon, one should consider one of the following options
i) Do pretty much the same as before.
ii) Join a FAANG company.
iii) Do PhD+ level research in ML/CS/AI.
iv) Make money as fast as you can.
v) Some option I didn't think about.
If you pick (i), it seems likely that your kids (if any) will have no money or/and power to remain relevant. Pick (ii)+(iv) or (iii)+(iv). The latter seems preferable but it is harder.
I would go further and say that once machines will be able to lie better than humans, we are done.
>> produce high powered tools rather than true autonomous agents
including tools which will help to tell a compelling story and/or lie to people. We can see precursors of these tools appearing during the last few years.
Always a good way to ensure you've covered everything
Should come up with something smart and make money. Let's call "be truly smart" (what a smarter AI would do) and not just "academic smart" or "coder smart".
Otherwise it's pure research and effectively a cost center. There is no standalone AI/ML etc... capability that makes money on it's own. It's all where it's integrated with existing consumer or enterprise products.
Understanding that, whomever builds and scales product adoption, are the ones who are really controlling the direction of AI. If you look at who are building and scaling consumer and enterprise products it's really only a handful of players.
So if you're a successful startup with a good product that uses AI driven features, you're either gonna get acquired or a major (FAANG etc...) will just copy and scale your product way faster than you can. So again, all roads lead to ii).
You can do trading, no need to have clients in that case.
DNNs are one of the best tools we have for bringing uncertain, hugely complex real world audio, text, and visual data into a form where we can manipulate it symbolically or mathematically with traditional programming. In other words, they bring more domains into the scope of automation. (Observe that none of what I just said sounded like "AI" as popularly imagined).
Not much going on in this area. Which is why robot manipulation in unstructured situations still sucks, after 50 years.
https://foldimate.com/
https://www.theverge.com/2018/1/10/16865506/laundroid-laundr...
So not quite solved just yet. And I'll give you long odds against them working out the kinks in the next five years.
https://youtu.be/OpiFp9i6owY?t=209
2. Human operators are doing the heavy lifting (quite literally) by loading the pieces one by one in a standardized initial state.
I'm talking about a machine that will fold men's shirts that are initially sitting in a pile just as they came out of the dryer.
Think about all the time going to waste. Yes, there are tasks which some actually want to keep manual (maybe you love cooking) and there are (subpar) online service, which try to replace that labor, but there is no automatic option otherwise. And outsourcing labor is not what I mean. I want a washing machine at home, which is a lot simpler to handle than outsourcing it to a washing business.
I sometimes wonder, if the reason is really a technical one or if there are other factors, social or incentive problems, which hold progress back in those areas...
Try folding a shirt with mittens on some time to get some idea of what a robot with a standard manipulator would have to deal with.
[1] https://youtu.be/YdStbvguvwU
Maybe the solution is lower standards?
Reminds me of some story about a person doing something and another person saying "that looks painful, what's the trick", with the answer being "not minding that it hurts".
I suspect this sort of algorithm will be applied to robot manipulation in the next couple of years, and would be extremely surprised if DeepMind et al weren't already working on it.
Remember there is a very fine line between creativity and illusion. Technically there is not much of a difference because both are a distortion of reality. Even we humans have a hard time distinguishing this sometimes as many new inventions used to be deemed crazy at first. It looks like when it comes to concepts that we hadn't seen, we were also susceptible to the same flaw just like those machines.
I don't really know much about symbolic AI but can it really solve this problem? If it can then the human game may really be over!
Only in the most tautological sense - explanation has to be made in a form of narrative in order to communicate it. As for the correctness aspect, I recall reading on HN recently that the study which claimed to demonstrate that such explanations are wrong had some bad issues with their data science.
https://twitter.com/tabithagold/status/1070736319901519876 https://twitter.com/tabithagold/status/1071189769499996161
"....The old-fashioned car manufacturers said 'We believe in electric motors too, and we can derive electric motors by injecting petrol into the engine."
There is a persistent lure to symbolic AI, and it is the lure of anthropomorphization, of thinking that AI must be smart in the way humans are smart, by manipulating symbols.
But natural language is not so much a central expression of machine intelligence as it is of human intelligence. Humans confuse linguistic aptitude with intelligence. Symbolic manipulation will be at best an API by which machines can relate to humans as the machines get smarter and smarter. Symbolic tools that conform to the bandwidth limitations of humans, which are constraints that our current machines don't face. (Historically they did, and symbolic AI made more sense then, since neural-net training was unfeasible.)
The lure of symbolic AI puts us in the bondage of old ideas, as Keynes would say. Symbolic AI is the equivalent of replacing our economy's fiat currency with an arbitrary supply of a yellow mineral.
I could go on, but I'll just link to this page, for those who are interested: https://pathmind.com/wiki/symbolic-reasoning
Sounds like you are not from the field.
It's like you haven't seen the advancements in CV or NLP (not involving translations) made capable entirely due to fancy neural network arcitectures
To me, AI = self consciousness, anything less is just fancy ML.
If you ask the question "are we going to have self-conscious machines in 30 years" I would bet against it. If anyone has any reason to bet _for_ it, I'd like to hear those reasons :)