I'd dispute that AI research is stuck or that his proposed answer along the lines of "An international A.I. mission focused on teaching machines to read" is a good one.
Research seems to be cracking along with AlphaGo, self driving cars and the like. Recently DeepMind have been doing interesting stuff with dreams[1], imagination[2] and body movement[3], the last one being a little reminiscent of his daughter inventing a way to exit her chair mentioned in the article.
Re government intervention it's not something like CERN where you need billions in capital and it's not an area where a big government project is likely to be the best use of capital.
There's a LOT of non-Deepmind research as well, some things that have actually been published before Deepmind.
Don't get me wrong, Deepmind puts out a lot of great work, but do look at other research labs as well, especially the ones that aren't as well known. Deepmind markets their research really well, but there's a ton of other labs doing good work as well.
Any tips on how to find the more obscure stuff? http://kurzweilai.net occasionally bubbles up really crazy stuff, but perhaps there is a journal or other curated resource you know of.
Let's just say if the terminator had the brain of a self-driving car, I could kill it with my Prius. Lots of "progress" is being made but you could still fairly call it "stuck".
If you look at inventions in tech that had huge impacts it was precisely huge government investment that made it happen. Are you denying that Chinese government investment in AI is going to pay off? Read the AI literature and you will likely find a Chinese name on the literature. This is because of government investment, not despite it.
These things take time and money. The profit motive is a killer of invention.
I've been down this road. I used to be interested in "embodiment" as a path to AI, but in the sense of things that could move around in the physical world, not fall down, and not bump into stuff. Low-end mammal level AI was the goal - mouse level. I made some progress in that area [1], hit the limits of available simulators, spent several years on improving physics engine technology, and eventually sold that off to a game middleware company.
Boston Dynamics took that much further. But it's the same approach I used - analysis of the dynamics, not AI. It's a complicated problem in dynamics, but it barely needs AI at all. Boston Dynamics, unfortunately, demonstrated that even if you spend $120 million, you're not at a minimum viable product that sells yet. Really cool legged robot prototypes, though.
This was all before machine learning took off. For a while I was looking at adaptive model-based control, which is a lot like machine learning.
Machine learning seems to be getting good at what the front end sections of the visual and auditory cortexes do. This is real progress. But a whole organism is still out of reach.
There's a business case for focusing on language skills, but in a way, it's a distraction. The mammals all have close DNA compatibility, but only humans do language much. If we can get into the entry-level mammal range of AI, we should be getting close. I once said something like that to Rod Brooks when he was promoting Cog (a robot humanoid head with a lot of compute power), and he said "I don't want to go down in history as the man who developed the world's greatest robot mouse."
Reverse engineering biology is going very slowly. See "openworm.org", which is an effort to develop a good computerized model of the C. elegans nematode that runs in simulation. C. elegans has 302 neurons, the wiring diagram is known, and it still doesn't work. This shows how little we really know about nervous systems.
I sense the hand of an editor. Particularly regarding the title.
Embodiment seems to be a branch with low-hanging fruit, when it comes to advancing AGI. I think the economic structural problems are important, but it's possible to over-egg the details and for some lab to stumble on an experimental paradigm with features we didn't realise were implicated a priori. When it comes to other AIs, the idea that we are stuck for pragmatic/practical issues is a little silly.
I'm no expert, just a person with an arm-chair (and too much time on my hands), but I suspect that idealising the feature-space we work with can hide as many things as it reveals - it may turn out that the computational problems are so large because we are mostly attempting to solve them ex nihilo. That is, embedding in an environment plays as much a role in the process of intelligence as a neuronal structure does; genes and evolution provide a mode for translating environmental computation into neuronal computation. The vast scope of what we don't know about the role of glial cells for cognition (and the little that we do) makes me doubt that complex structures of binary mechanisms will be sufficient. But again, that's just my speculation, and perhaps lack of education.
> I sense the hand of an editor. Particularly regarding the title.
Gary Marcus is probably fine with the title. He's been talking down deep learning (and talking up his own more old-fashioned Bayesian flavored ideas and startup) for years now, trying to ignore all the successes like Google's knowledge graph and just omitting the actual research, like when he says
> Such systems can neither comprehend what is going on in complex visual scenes (“Who is chasing whom and why?”) nor follow simple instructions (“Read this story and summarize what it means”).
Which deep learning actually... works pretty well on? Look at Facebook and Google's work on visual & textual question answering using approaches like memory networks.
The work on textual understanding is definitely still early days (though the press release makes it seem like it's understanding the entire LOTR, it's just reading a very structured short version) but already crazy impressive:
https://venturebeat.com/2015/03/26/facebooks-latest-deep-lea...
On visual QA, Francois Chollet's talk at the TensorFlow dev summit shows how easy it is to get a (again constrained to 1-word answers but still very impressive) video QA system working in like 20 lines of Keras:
https://www.youtube.com/watch?v=UeheTiBJ0Io&vl=en
And to sum it up a bit, the hypothesis is that humanlike AI is as much a product of the experience and reality of being physically (and limitedly!) human as it is any abstract algorithm.
You may also know him from a small company called iRobot (aka Roomba).
Thank you very much for the link, and nice summary. This looks very interesting, I'm grateful: my learning habits mean I'm liable to playing with second-hand scraps of ideas, and missing out on sources.
> Embodiment seems to be a branch with low hanging fruit, when it comes to advancing AGI.
If it were "low hanging" it would have been picked already. Reinforcement learning with AI agents is hard, especially in a dynamic environment with many types of objects.
I think the path towards AGI is to do simulation coupled with deep learning. Simulation would open the door to predicting non-trivial effects that cannot be learned by example because they are so rare that there are no training examples. We can generate artificial training examples to cover all the rare cases.
I'm sorry if 'low-hanging' comes off as disrespectful - I'm just guessing that aspects of embodiment, once understood, will be capable of fairly trivial description and reap large consecutive rewards. I remember your u/n from other posts, we seem to have similar interests but you are vastly more educated in the engineering of AI. What's your background, if you don't mind sharing here?
I am suspicious non-contingent aspects to cognition remain that simulation and deep learning don't necessarily grant, though they might well be sufficient. I'm not smart enough to be sure, and I'm stretching for a description: a child self-reared to adulthood in the wild won't display what we usually consider essential facets for 'humanlike' levels of intelligence or competence. We're hardly trying to build a caveman.
They lack whatever is crucial in socialisation -- the ability to make subtle differentiations between other agents' actions and motivations seems to endow self-awareness, and abstractions for successfully handling novel objects and ordering perception relevance. Successful generality to our degree seems to be better 'outsourced' rather than hard-coded into solo agents, at least in the natural examples. Though I understand that's not necessary, perhaps there are good reasons for it. I feel like the first AGI will actually look a lot more like "multiple similarly 'perspected' AIs interacting with one another leads to each carrying the G in AGI". Essentially I'm suggesting it's hard to have generality and relevance to our proficiency (or better) without a 'culture'.
What I'm thinking seems to boil down to inserting some of Piaget's ideas into the philosophy of AI, which might be a bit much, and I'm open to charges of bullshit.
I am just a hobbyist in machine learning. The current situation in AI is that we can only do limited aspects of perception: like vision and hearing, at a superficial level. We can recognize objects but we can't recognize relations between objects nearly as well. There is no global scene understanding yet. With text, we can do syntax and translation, but we can't do reasoning except on trivial cases. There is no power of abstraction yet, of transferring knowledge between domains, which is essential for advanced intelligence.
So, before we have an embodied agent, we need to solve the reasoning and abstraction part, and my money is on graph signal processing (a kind of neural nets) and simulators (also implemented as neural nets). We need to move from simple object recognition to reasoning and simulation on graphs of objects and relations.
> Not long ago, for example, while sitting with me in a cafe, my 3-year-old daughter spontaneously realized that she could climb out of her chair in a new way: backward, by sliding through the gap between the back and the seat of the chair. My daughter had never seen anyone else disembark in quite this way; she invented it on her own — and without the benefit of trial and error, or the need for terabytes of labeled data.
I really hate the constant comparisons of AIs to babies. The author's 3 year old daughter has had 3 YEARS of sensory data obtained through moving and trying to fit through things. That is terabytes worth of data! I would expect an AI to be able to generalize once as well.
Yes terabytes.... Probably petabytes or more. And all of it is unlabeled data.
If you fed video and sensory data to a deep net for 3 years and somehow were able to come up with an activation function that modeled "survival", I still highly doubt that anything at all would come out that remotely resembles human intelligence. There's no way that i'm aware of to label reality in real time.
One way to label is by expectation. If the outcome is what you expected, then it gets one label. If it didn't it gets another label. Deepmind has had some success with this approach.
Not to mention that kinds don't even know how to _see_ when they are born, and for several months thereafter. That's petabytes of visual stimuli going through that brain so that it learns what's statistically more important and infers semantics around it all.
I think an engineering based mindset of perfection creates brittle models that get stuck in local maximums. Reverse gravity and you can think of current AI as small cars getting stuck in pits. The desire to maximize leaves no room for the human equivalent of refactoring. Giving up the good and going through pain in order to reach new heights.
Instead of an anti fragile model that can continuously be run, current models need to be scrapped at error states. Hyperparameter tuning is just random guessing. Getting good data also doesn't work because of the high dimensionality of it for non-trivial tasks.
The idea that ctrl+z is good should be re-examined. Perfect memory like block chains have isn't the answer either though. Perhaps something similar to the non forgetting yet imperfect human mind.
Most data is garbage and even more eventually becomes garbage. (unless you exist in a finite defined world like Go) Is there any sort of neural net that that find or creates "core" memories with weaker supplementary memories?
The quintessential experiences would only be dislodged with an influx of contradictory data. Initial cores could be initialized via mother-child like training. The training data would be tiered and weighted. There would be an internal system that passed judgement on new ingestion sources. New data would be a necessity. Old data passed in as new would be like a monotonous life digging in cores preventing them from having meaningful change. Almost all data would be labelled as garbage initially unless vouched for somehow. Pure good data would be bad as well because there isn't enough quality differentiation to see what is core and what isn't.
> I really hate the constant comparisons of AIs to babies
Constant comparison? First time I see this comparison to a child.
And btw, it's about that we always equal AI with machine learning, pattern matching, etc.
To be as smart as a 3 year old, we need something entirely different, something way beyond machine learning and all the as AI classified techniques we are aware of today (I think this is what the author meant).
An agent can be intelligent without it learning how to read human language. Look around, most organisms in our world communicate using extremely simple binary language or don't communicate verbally at all. Yet, they are intelligent enough to do very complicated tasks which current robots fail to do. Intelligence is an easier problem than language, and thus should be solved before language.
What a sigh of relief to read a refreshing take on the real progress of AI. Yes, it's stuck, and that's the real problem of AI, that we haven't been able to do anything significant after perception. However, unlike the author, I don't think the solution is to nationalize AI research (we're not close enough for that), but to fund more non-deeplearning research for 5-10 years, and then we might see some progress in non-perception tasks.
> Look around, most organisms in our world communicate using extremely simple binary language or don't communicate verbally at all.
Yes.
> Yet, they are intelligent enough to do very complicated tasks which current robots fail to do.
True.
> Intelligence is an easier problem than language, and thus should be solved before language.
Wrong.
This is the classic mistake everybody makes, including people in Computer Science.
Because if that were so our robots would already be clambering backwards through chairs (per the metaphor in the article).
You have to think of deep evolutionary history. It took centuries to come up with advanced mathematics, so in some strange to humans sense, this isn't that hard. Same with language, it only took tens of thousands of years.
For Nature to learn how to develop a nervous system capable of flexibly interacting with the environment, culminating in our brains, took hundreds of millions of years.
This isn't an claim that we have to wait that long to re-engineer such powers, but it is to point out that if the possibility space for developing a nervous system was much larger than for the same organisms to learn language...
tldr; Walking is hard.
We have been conflating what is easy for us, with what is objectively easy, because we don't appreciate the Deep Time that Nature has been working with. I suspect we will develop EMs (brain emulations, a sort of short cut) before we understand what we are doing but I hope that is wrong.
If I am following you correctly, you are arguing that walking is a harder problem than language because it took much longer to evolve.
This seems to assume that a facility for language and advanced mathematics is independent of the existence of a nervous system capable of flexibly interacting with the environment, but it seems plausible, indeed probable, that language, consciousness and math depend heavily on the prior neural infrastructure, and their development was the most recent step in a process that has been going on since the evolution of the first synapse.
On the other hand, I am skeptical of the somewhat popular view that the key to generalized AI is to make robots that interact more thoroughly with their environment, and that they will then find their own way to language and consciousness. Partly, this is because I do not think that if you intentionally pursue the robotic goal, you will necessarily create the sort of infrastructure that is generalized enough to be the basis for the emergence of language.
> If I am following you correctly, you are arguing that walking is a harder problem than language because it took much longer to evolve.
This is a thorny subject. So I am saying that in some objective way, walking is harder than language because Nature took millions/billions of years to traverse the solution space. Then... once we had a huge number of preconditions existing, then we had the development of language.
I am not saying that this means if it takes 10 years to develop language with some artificial means that it will take 100,000 years to develop walking.
What I am pointing to is that we ought to appreciate that if even blind natural selection took that long, then the possibility space to develop a nervous system must be much larger than we have anticipated.
As evidence of this: consider how (at least in popular culture, but also in comp sci in the old days) we developed chess playing computers and it was broadly assumed that breakthroughs in getting robots to walk and talk would soon follow through. That did not happen. It was a natural assumption but it was wrong.
> This seems to assume that a facility for language and advanced mathematics is independent of the existence of a nervous system capable of flexibly interacting with the environment, but it seems plausible, indeed probable, that language, consciousness and math depend heavily on the prior neural infrastructure, and their development was the most recent step in a process that has been going on since the evolution of the first synapse.
I don't know the answer to that. On different days I think one or the other is true. On Day #1 I think Nature obviously required walking before talking, but we could develop them differently, just has we didn't need to develop better horses to produce cars. On Day #2 I think to myself there's a deeper sense in which you really do require walking before talking because otherwise why didn't Nature develop biological microlife which evolved communications ability long before it developed legs. So...
> On the other hand, I am skeptical of the somewhat popular view that the key to generalized AI is to make robots that interact more thoroughly with their environment, and that they will then find their own way to language and consciousness.
We cannot be certain consciousness or intelligence are high probability events once you have life. We could be like those French artifact makers who made such exquisite mechanical toys for the aristocracy but ultimately got nowhere whereas the English inventors meddling with water and steam power really kicked off a revolution.
Who is Silicon Valley is genuinely looking at the fundamentals of A-Life or AI? OpenAI? MIRI? Stanford? DARPA?
> For Nature to learn how to develop a nervous system capable of flexibly interacting with the environment, culminating in our brains, took hundreds of millions of years.
Nature never learned anything. Nature is not a force that chooses what features it wants to implement in living things. We evolve in periods of punctuated equilibrium, when the average individual within a population cannot reproduce successfully. Then species change very quickly(sometimes sub-1000 years) to fit their environment. We're not sure why humans evolved to be so intelligent, but one possible reason is that our environment was changing very quickly, so quickly that we had to change our core behaviors within the course of a life time.
It can be very confusing to anthropomorphize Nature. Since Nature never tried to make intelligence, the speed at which intelligence evolved in Nature is pretty irrelevant to the difficulty of the problem of intelligence.
Are you nature? Did you learn? Are you not a force?
It's true nature never "tried" anything except to keep going but I posit it does learn, it's memory is our genes and our own memory, and we are the effect of it's force. I also don't see man and nature as separate. If we created AGI, then nature created AGI. AGI can look back and say the step from biological to machine was akin to single to multi celled organisms.
Yeah that's interesting, it does have some kind of memory with all the genes we have. I mostly just meant that Nature is not an intelligent agent with an agenda (unless you believe that in a spiritual way, which is cool too).
>If we created AGI, then nature created AGI. AGI can look back and say the step from biological to machine was akin to single to multi celled organisms.
I do think it's interesting to consider AGI as a similar step from single to multi celled organism, but I also feel like if we consider everything manmade as part of nature, the term nature doesn't really mean anything. If it refers to all man-made things as well as all non man-made things, it kinda just becomes a synonym for "things"
"Based on replicating nature"? Fiber optic cable? Microchips? Mass spectrometers? Atomic weapons? Even humans first technology, sharp stone tools, isn't really a replication of nature. Maybe some examples would help me understand what you mean.
There seems to be something very different between language (and language-mediated behavior), and almost all of the behavior found in other animals. While the former and at least some of the latter can reasonably be described as categories of intelligence, one can not necessarily argue for claims about the former from facts about the latter.
Perhaps we should make a list of increasingly difficult problems to solve. Image recognition is at the front of the list, and we can cross it off. What is next?
I can't follow the author's logic. First he complains about the limited breadth of AI approaches (bottom up) and then he makes the case for more central coordination of research efforts.
Contrary to applied physics or medicine, AI doesn't require massive capital investment like building a particle accelerator or running clinical trials over years.
So if we already suffer from a lack of diversity, why should we ape the organizational structure of those fields?
The author's general problem is that he seems ignorant (or willfully ommiting) the expert systems period in AI (realized and popular in the 80s, academic foundations discovered in the 60s and 70s).
I agree with the article that GP AI is likely to ultimately be a fusion of bottom-up with top-down systems, and that expert systems seem to be getting short shift after their earlier failures while neural networks are possibly receiving overly optimistic expections.
To be fair, I believe this is the author: https://en.m.wikipedia.org/wiki/Gary_Marcus , and he appears to have a cognitive neuroscience background as opposed to computational AI. So I wouldn't be surprised if he actually was unaware of 1960s-80s CS AI research.
Examples of top-down algorithms, in my opinion (since bottom-up and top-down are debatable concepts in many concrete situations), include:
- Genetic algorithms
- Q learning
In the sense that they learn general behavior first and then learn ever more little "tricks" to be used in particular situations. Both are more effective when combined with ANNs. But when they start they're only aware of very high level goals.
That said, I also have kids, and while they're bigger now, I would argue the idea that humans work top-down from the very beginning doesn't survive caring for a toddler for a few hours (babies can't really move, so they don't make particularly stupid decisions. Toddlers and up to teenagers make idiotic decisions that make sense from particular perspectives. For instance, they exhibit extreme short term decision making (like taking a huge risk of falling down just to get a little piece of candy).
Top-down decision making isn't just something that is eventual emergent behavior, it's learned behavior. Telling a toddler that to get candy he should go to the store, get flour, sugar and ... and follow this recipe doesn't work. They get distracted after 30 seconds. It's not that they're trying to fail, their mind just doesn't let them focus beyond a certain (short) amount of time. Adults have the same limit, just longer time, but they have learned to compensate for it. For instance using TODO lists, or project plans.
I've always had the suspicion that memoization (or lack thereof) is the primary reason really young children do and think a lot of the ways they do.
As adults, it seems like we don't actually experience every sensation of the world anymore. Most of the time it's already high level categorized (e.g. "apple") by the time it hits our conscious mind.
I agree, it's obvious that human intelligence is >99.9% a behavior copying algorithm. What we call rational thought is in reality restricted to conscious exercise and it is a learned skill. A trick, nothing more, and especially not a core part of our behavior. It is not that different, at a low level, from learning to juggle balls. Rational behavior, firstly, most people just don't have it at all, and secondly even in the people who do behave rationally occasionally, it is only when things are happening slowly enough and they're putting in constant effort toward maintaining that rational behavior, constantly second-guessing themselves and going back in memory every few minutes to evaluate your own actions and formulate a plan.
And if you've been to the third world (or just a large poor part of a large western city), you'll know this is true: billions of people have never learned to act rationally, and only few and far between will ever act rationally. You can do a thought exercise with these people and figure out with them what the rational action is, and the vast majority will simply act anyway.
In 1988 Hubert L. Dreyfus and Stuart E. Dreyfus released a paperback version of their previously published "Mind over Machine" book, in which they mostly spend time debunking the myth that expert systems and rule-based programs are ever going to have "intelligence" on par with human brain.
The book is an interesting read in itself, but what I found remarkable is that in the 1988 release they added a "preface to paperback edition" in which they used a couple of pages to give their views on artificial neural networks, which (though not new) was gaining some steam at the time. The conclusions they reached are as relevant now as they were 3 decades ago.
There have been no new breakthroughs in this area. Most of the research being done is in application of what we have known for decades in specific areas, with minor insights into tweaks and uses of combinations of algorithms to better solve specific problems. The big differences between then and now are: (1) technology is more accessible - data is easier to collect, store and output via many input/output methods; and (2) the hardware is significantly faster - we can now go through more data, make algorithms run faster, and appear to perform better.
This inevitably brought a lot of hype, including many predicting human-like artificial intelligence not too far away. But maybe those with experience in 60s and 70s in the field in USA and Japan can draw a parallel between what's happening now and what has happened few times in the past in this area:
- companies perform neat promising demos with unrealistic implicit or explicit promises
- investors pour money in
- media hype ensues
- after awhile - no new breakthroughs: still can't turn ANN or expert system into a human brain
- outcome is improvements in limited use cases
- hype dies down, but we can repeat the cycle after improvements in hardware
> There have been no new breakthroughs in this area. Most of the research being done is in application of what we have known for decades in specific areas, with minor insights into tweaks and uses of combinations ...
There are 2 huge problems with that:
1) nobody is trying to "embody" an intelligence with any sort of research project behind it. Nobody's even trying to create an artificial individual using neural networks. There are several obvious ways to do this, so that's not really the problem.
Therefore I claim that your implied conclusion, that it isn't possible with neural networks somewhere between premature and wrong.
2) What if the difference between an ANN and our brain is a difference of scale and ... nothing more ? We still do not have the scale in hardware to get anywhere near the human brain, and just so we're clear, the differences are still huge.
Human neocortex (which is roughly what decides on actions to take): 100 billion neurons
Human cortex (which is everything that directs a human action directly. Neocortex decides to throw spear and the target, cortex aims, directs muscle forces, moves the body and compensates for any disturbance like say uneven terrain): another 20 billion neurons.
Various neurons on the muscles and in the central nervous system directly: a few million (mostly on the heart and womb. Yes, also in men, who do have a womb it's just shriveled and inactive). They're extremely critical, but don't change the count very much.
AlphaGo 19x19x48, times 4 I think. About 70000 neurons, and that does sound like the correct number for recent large-scale networks.
A human neuron takes inputs from ~10000 other neurons, on average. A state-of-the-art ANN neuron takes input from ~100, and since it's Google and they've got datacenters, AlphaGo was ~400.
So the state of the art networks we have are on par with animal intelligence of the level of a lobster, ant and honeybee. I think it is wholly unremarkable and understandable that these networks do not exhibit human-level AGI.
What is remarkable is what they can do. They can analyze species from pictures better than human specialists (and orders of magnitude better than normal humans). They can speak. They can answer questions about a text. They can ... etc.
Give it a few orders of magnitude and there will be nothing these networks don't beat humans on.
The author is just spouting off on a topic he doesn't understand. It's just a rehashing of Chomsky's hatred of statistical NLP. He pulls off the neat trick of approximating knowledge of artificial intelligence by hoodwinking the New York Times, but he doesn’t have insight into the topic he's talking about.
Can you describe what you mean by "he doesn't understand"?
His background is in psychology and neuro science so his view on intelligence is probably quite different than person coming from computer science.
For me he puts words onto something I've felt recently, that what we're doing is cool and all, but just doesn't feel like the right way to approach it. We're just putting loads of data and computing power into something that produces results that looks intelligent, but digging deeper bares no resemblence to what a neuro scientist would call intelligent..
> Even Google Translate, which pulls off the neat trick of approximating translations by statistically associating sentences across languages, doesn’t understand a word of what it is translating.
This is just another incarnation of "AI is the thing we haven't done." He's parroting Chomsky's disdain for statistical models and John Searle's fundamental misunderstanding of AI. For the former, Norvig has a fair rundown of Chomsky's complaints (http://norvig.com/chomsky.html).
> bears no resemblance to what a neuroscientist would call intelligent
TensorFlow gets results. The neuroscientist can claim it's a P-zombie, but they need to point to some criteria for accepting something as intelligence. Otherwise we're just moving goalposts.
>> Even Google Translate, which pulls off the neat trick of approximating translations by statistically associating sentences across languages, doesn’t understand a word of what it is translating.
>This is just another incarnation of "AI is the thing we haven't done."
I don't think so - it appears to be an objectively correct assessment of the current state of the art.
> Otherwise we're just moving goalposts.
The first movement of the goalposts was to call '80s technology AI. Now they are drifting back to where they started.
On the other hand, I am surprised by the claim that AI is stuck; my outsider's impression is that progress has accelerated. Perhaps the impression of being stuck comes from more people realizing how difficult a problem it is.
> On the other hand, I am surprised by the claim that AI is stuck; my outsider's impression is that progress has accelerated. Perhaps the impression of being stuck comes from more people realizing how difficult a problem it is.
Deep learning made practical a large number of applications that were previously intractable by neural network approaches. Advances over the last 10 years have pushed the boundaries of what machine learning systems are capable of doing. However machine learning has algorithmic limits to what it can accomplish, and we are starting to hit those limits. A change in paradigm is required to begin making real progress again. Either a change to something new or a regression to older ideas that were temporarily put on the back burner.
That's not a universally held view, but I think it is the sentiment behind this editorialized title.
>>> Even Google Translate, which pulls off the neat trick of approximating translations by statistically associating sentences across languages, doesn’t understand a word of what it is translating.
>> This is just another incarnation of "AI is the thing we haven't done."
> I don't think so - it appears to be an objectively correct assessment of the current state of the art.
How deep an understanding is required to meet the threshold? The skepticism feels like "no true Scotsman" applied to the definition of understanding.
I observe the following in young children when exposed to a new word:
0. First exposure to totally new word used in a sentence with more familiar words.
1. Brief pause
2. Mimic pronunciation 1-2 times
3. Process for minutes, hours, or days.
4. Use the word in a less than 100% correct way
5a. Maybe hear the phrase repeated back with the error "corrected" (hello internet)
5b. Maybe hear more usage of the word in passing from others (with varying degrees of "correctness")
6. Recurse for life.
At what point did the person understand the word? How is AI translation substantially different?
I'm not sure I understand any word in a way that would satisfy AI skeptics.
This is a discussion of the current state of affairs, not, for example, a Searle-like claim that understanding can not and will never be achieved. To substantiate a claim of 'no true Scotsman', I think you have to present an actual case where you think a machine has achieved understanding, but which is being unreasonably dismissed.
Ironically, your last sentence has 'no true Scotsman'-like reasoning, along the lines of 'no true AI sceptic would fairly evaluate a claim of machine understanding.'
BTW, I am not a skeptic of the potential of AI, though I am skeptical of some claims being made.
I think the thing that bothers people like the author and Chomsky is that deep nets can't explain or justify how they make decisions in a way that a human could fit in their brain. There is no book called "How To Play Go as well as I Do" by AlphaGo. This is something we're going to have to live with : The machines are smarter than we are, so we won't be able to understand them except at the most basic levels where we use inductive proof to extrapolate our understanding of very small models to enormously large ones that are beyond our ability to fit into our brains.
We can understand how individual chemical reactions in Einstein's brain work, but that doesn't make us smarter than him.
That's why I like the turing test. That one is already pretty good. Is there something more advanced ? Since by now i would expect that the AI community (or others) have defined a better test to accept true AI ?
I feel like you're conflating sentience and intelligence. Algorithms don't have to reduce to anything special to display intelligent behavior. After all, humans reduce to chemistry. All that is required is the ability to solve problems.
As for sentience, we don't understand it, so the only way we'd recognize it is if it was extremely similar to human sentience.
Your concern about "the right way to approach it" is getting closer to the real conflict which is not in the "way" but in what "it" is. In your mind, what is the desired goal or outcome for AI R&D?
If your objectives are in medicine, cognitive science or philosophy of the mind, you might want simulations which are isomorphic to biological minds. You probably hope that AI work will provide illumination into how the mind works, or why it sometimes fails, or how to improve it.
If your goals are in computing and product engineering, you want predictable, reproducible, and adaptive methods for making smarter tools on time and on budget. You may want the product to have behaviors compatible with humans (as a product feature) but you shouldn't care whether the implementation technique in any way resembles an actual human mind. Behaviorism is all that matters for a product evaluation. The design and marketing teams can take care of imbuing the product with intangible properties imagined by consumers.
And honestly, if you want a biological mind, we already have techniques to build them: go find a mate, procreate, and raise your offspring. Nobody tasked to deliver a commercial AI product is actually going to want a solution that behaves like real human minds, where individual units off the same assembly line may require psychotherapy, develop self-destructive habits, or worse slip through QA with an undetected sociopathy or psychopathy which creates a manufacturer liability.
Some of us old school engineering types may harbor a disdain for the current neural net renaissance because it feels a little too black box to us. Deep down, we'd prefer a tool-building tool that had more directly visible logic and rules in it, because we tend to believe (rightfully or not) that such a method is more amenable to engineering practices and iterative designs. But, the risk in this mindset is in forgetting that even complex, logical systems can exhibit emergent properties and chaotic behavior. We probably need to engage in more statistical methods whether we like it or not...
great perspective! My desired goals are more in something like cognitive science..
I think most of my grumpiness about this is not that things arent progressing (check out Santa Fe Institut and their work on complexity) it's just that AlphaGo and self driving cars are getting all the attention
I believe he has much more insight than most AI researchers/engineers. What's going on in AI right know looks exactly like what happened in the 1960s, sucking all the resources into something (i.e Deep Learning) without looking at the big picture, without looking at epistemological questions, and not training new grads on solid AI foundations anymore like probability theory and logic. Not asking the right questions in terms of expressivity of models, sensitivity analysis, robustness and manifold learning. The next AI winter is coming.
Yes, but the next AI Winter will be far less bleak and job-free as that of the 80's, IMO. The hayday of DL-based AI is likely to continue for 10+ Years, as we refine the methods and see how far word2vec and improved compositionality of the many existing DL methods takes us. AI is likely to improve rapidly for another decade, IMHO. But I agree that DL AI, if not augmented substantially, will stop well short of AGI and consciousness.
But 10 years is a long time in internet years. I would not be at all surprised if there's another word2vec (or two) arose before then to displace the tectonic plates yet again.
Indeed, DL-based AI is likely to continuue - rightfully so - for non-safety critical domains (translation, games, ...) It's crazy that few people notice that real world applications with safety issues (where 99.9% accuracy is a joke) - ex: self driving vehicles - require much more that what DL has to offer. The AI community's sloppiness when it comes to safety and robustness engineering will lead to the some serious problems (when it comes to real-world industrial applications). But yeah there will still be jobs for people playing around.
The author is a tenured professor at one of the best institutions in the world - what do you have to show for yourself?
If you honestly think the author should be discredited because you don't find is PhD appealing, you need to take a step back and seriously readjust your arrogance level.
> Working at a startup does not make you an expert. Especially when your opinion goes against what pretty much everyone else in the field thinks.
Achievements are perhaps the best yardstick for expertise. Whether a person has worked at a startup or has contrarian viewpoints are both irrelevant.
Gary Marcus was co-founder and CEO of Geometric Intelligence, successfully raised some money and grew a team, [1] then successfully sold the company to Uber, after which he directed Uber's AI lab. He has a PhD from MIT in cognitive science. And he's been a professor of Neural Science at NYU for nearly 20 years. [2]
"Especially when your opinion goes against what pretty much everyone else in the field thinks" -> Most Deep learning people are consumers or hardware manufacturers, very few researchers in the field, and most people that have deep knowledge of it would agree with this guy.
AI is currently about giving us an equivalent of a pocket knife for some tasks that are viewed as being in the "intelligence domain". That's all. Nobody really thinks we are anywhere close to general intelligence. Or like the famous saying "computers are bicycles for brain", current AI is about adding a small electric engine to those bicycles to make them easier for everyone.
> Even the trendy technique of “deep learning,” which uses artificial neural networks to discern complex statistical correlations in huge amounts of data, often comes up short.
That doesn't seem to be very surprising given the limited complexity compared to say a fly's brain. Artificial NNs manage to work because they are highly specialized to a specific task.
The problem we've always had with AI was that most people were trying to engineer it rather than reverse engineer it. Every time there would be a major advance the computational neuroscientists would say: "we knew that, you should have come talked to us 15 years ago." There's some work out there on this, but it's more basic research on how to use developmental and genetic and evolutionary algorithms to grow neural networks. Most AI researchers try to skip this step but it's what's holding back progress.
It's been about 10 years since I left the field. I believe Rodney Brooks did some work on this. Eggenberger was also doing some interesting work about 15 years ago--not sure where that is today. The computational demands for evolutionary & genetic algorithms are significant but there's no free lunch.
> An international A.I. mission focused on teaching machines to read could genuinely change the world for the better — the more so if it made A.I. a public good, rather than the property of a privileged few.
> author: Gary Marcus is a professor of psychology and neural science at New York University.
Not sure what he has in mind. There are already a lot of smart people building Q&A systems. We need tests to establish if a system can read. Once you have those then you can throw a competition up on Kaggle with a big purse.
An example: The city councilmen refused the demonstrators a permit because they [feared/advocated] violence.
When you switch between "fear" and "violence", the meaning of 'they' change. There are many more examples like this.
The best performance in the first round of the 2016 challenge was 58% by a neural network based system. Random guessing would yield 44% (some questions had more than 2 choices). Human performance was 90.89% with a standard deviation of 7.6%.
It's important to note that Winograd Schemas don't really test if the system understands those sentences, they essentially test the system has appropriate "common sense" knowledge/experience about how our world and society works, i.e., it tests whether the system understands whatever other data sources are usable to find out about this topic.
To give the proper answer in the example you use, a human (or a system) needs to know how such permits are issued and what are the common reasons for refusing such permits. As such, a sufficiently sophisticated pattern matching system is perfectly sufficient to answer such questions - there's a simple pattern difference that fearing violence causes you to refuse permits but advocating violence causes you to get refused. It's worth thinking about where do humans learn this? For the Winograd schemas like putting a trophy in suitcase, it's the basic childhood experience of putting stuff in boxes that we all share, but a machine won't (unless it's raised as a child-robot). For schemas like this one, it's understanding how our society works learned by participating in our society for years, which we all share, but a machine won't (unless we allow machines to participate in our society). I.e. it's not so much a measure of intelligence as a measure of shared background experiences. A human from a hunter-gatherer tribe wouldn't be able to answer the councilman-permit schema, but that doesn't mean he/she isn't intelligent.
The difficulty there is caused mainly by the need to have domain-specific knowledge in a wide range of domains - we will perceive systems as "dumb" unless they share the same background knowledge that most humans have gained by being part of our society and basic schooling, and since the machines won't do that (yet), we're looking for "unnatural" ways of getting common sense knowledge without the direct experimentation and participation that we do.
Click-baity. AI tech isn't stuck. There are many forthcoming breakthroughs, particularly in medicine, which should really benefit humanity. Radiology is poised to let CNNs make radiologists a lot more efficient. We just need to build the labeled datasets.
If we invest heavily in some AI tech, let it be to produce huge medical datasets. The software and hardware is ready. We're only lacking sufficient data to make more diagnoses with super-human accuracy.
Yeah, where I sit, papers from two years ago are considered ancient, and the advances if the last few years means I can train things on my laptop in for hours that likely would have been a week of gpu time in 2014. This means we can experiment more easily, and try things out with fewer resources, which in turn leads to faster innovation.
So AI isn't stuck. It's also mostly working on well defined, targeted problems.
Life, on the other hand, works towards a very ill-defined objective function (survive, collectively) over millions of years; all of the emergent behavior we're astonished by is maybe just side effects of working on that objective. (This is a crass viewpoint, but let's stick with it for the sake of argument.)
We mostly aren't working on such objective functions, partly because it's hard to compare results, partly because there aren't clear milestones for success (indeed the goal posts for AGI shift as far as AI advances) and partly because skynet.
In fact, we are consistently surprised by the AI we already have. It finds ways to exploit our fitness functions constantly, and fine tricks and heuristics to gain a couple points on the final score constantly. Click bait comes to mind: we want to surface good content, use clicks as a proxy for quality, and get what we see for instead of what we wanted. Which somehow takes us directly to president trump... (Sure your kid can find a cool way to get out of a chair, but call me when she inadvertantly threatens the basis of the US democracy in the process. And then we can talk about the pressing need for AGI.)
> Life, on the other hand, works towards a very ill-defined objective function (survive, collectively) over millions of years; all of the emergent behavior we're astonished by is maybe just side effects of working on that objective
Agreed. Plus think about all the data and processing power that went into evolution. And some folks think because a system beat a human at Go that we're nearer solving life's age old question, that is, the essence of intelligence.
> There are many forthcoming breakthroughs... We just need to build...
Not sure how you don't see the irony. This has probably been said thousands of times for many scientific areas throughout history. Example:
There are forthcoming breakthroughs in humanity being an interstellar civilization. We just need to build faster-than-light engines and terraforming equipment. Nothing major, right?
> There are forthcoming breakthroughs in humanity being an interstellar civilization. We just need to build faster-than-light engines and terraforming equipment. Nothing major, right?
Building a dataset is easy and not something you would compare to faster-than-light engines. Believe it or not, some major breakthroughs are held back by simple lack of funding, and lack of awareness.
To make a dataset you need to pay radiologists to label enough data for the system to do its job well. This could be thousands, or hundreds of thousands of images. It is technically speaking very doable, but also very expensive. Then there are data privacy issues stopping you from sharing data. These are social issues, not engineering issues.
Your reply, while overall correct, still underestimates the problem of "how do we invent AI?". Data-sets shouldn't be "tuned" or "refined" (tems I see in practically every "AI" article); if they need to be remade then the consumer is not only not AI -- it's not even a clever NN implementation.
Forgive my cynicism if you can, but in my eyes you guys just support what might make you money one day (or already does) and thus aren't objective. You're like the parents that are completely blind to their child's defects due to paternal / maternal hormones.
There's no AI on this planet. There are not even beginnings of an AI. Deep learning is practically a statistically biased classification algorithm and not much else.
To me the term AI is being abused. I want AI to exist, but I am seeing every indication that the area is falling victim to capitalistic interests and this won't change anytime soon.
> Your reply, while overall correct, still underestimates the problem of "how do we invent AI?"
I'm not talking about building a real AI.
I actually agree with you that we're nowhere near developing that. Not sure where you got any other idea from me.
I'm saying there are some machine learning problems that could be served by some simple data entry. This could save lives, including yours and mine, via advanced cancer detection [1]
You're right that since I studied data science, I'm incentivized to advertise its usefulness. But, I studied data science because I believe it is a growing part of our future.
You can try it yourself too. There are many tutorials online. Making use of machine learning gets easier every year.
There is pretty good reason to suppose faster than light travel is impossible. Aside from our current understanding of physics there is also the question of where all the aliens are.
If FTL is possible one could see even one intelligent species possessed of such technology spreading over the galaxy over thousands of years. It would also seem decidedly odd to suppose that if faster than light is possible only we are smart enough to invent it.
Removing faster than light travel doesn't remove the question of where are all the aliens but it sure does make it easier to swallow.
(My opinion on where are all the aliens is that collectively speaking, we're little more than ordinary jungle beasts with baseball caps (quote by George Carlin) and we're monitored and evaluated on when is a good time for a first contact. Let's just say we're easily at least a millennia away from that point.)
My point in my parent comment was that the overall schema of assertions like "breakthroughs are incoming" and "we just need to do X" are overly optimistic. So I gave an exaggerated example to demonstrate that point.
I agree with the fact that AI isn't stuck (flagged the article). I would say the research is ready, but there's still a lot of infrastructure work for integrating with data sets, learning and serving cheaply on high scale. Still, lots of people are making hard to productionize the research results.
One group ensures their works are a convoluted mess to maintain their dominance.Another group ensures the computational stack is a convoluted and resource hungry mess to maintain their dominance.
A match made in heaven. Two peas stuck in local minimum pot vacuuming up money and resources.
One pea says :
> there's still a lot of infrastructure work for integrating with data sets, learning and serving cheaply on high scale.
The other pea says :
> Lots of people are making it hard to productionize the research results.
Both peas agree : This is how I make my money and stay on-top.
You get what you get for reasons.
If neither of them wants to agree their stuck.
That's fine with those ushering in the new wave.
I'm taking my cues from Jeremy Howard, who believes we probably won't be a world full of data scientists, since data science requires less effort every year as the software and hardware improves. Rather, we'll still have domain experts, and everyone will just know how to apply data science to their field.
The tools get easier to use every year. While I have difficulty imagining my data science job disappearing, I tend to agree with Jeremy that more and more non-CS people are getting comfortable using computers. Programming, as a field of study, will stabilize at some point, and its usefulness will continue flowing into other fields.
Is the Author arguing there will be another AI Winter?
I have a simple theory ( I am not sure if there is a proper term for it ) that goes and solve all the above problem mentioned.
When ever a technology that is capable of producing some form of economical value, it will continue to improve and tackle what ever hurdle or barrier you think it has.
In all of the previous AI era, research were funded by government or large company like iBM. But none of those has ever made a impact or profits that value more then you have invested in. Expert System never caught on.
And this is why everyone is excited, for the first time ever we have AI ( Machine Learning ) producing useful results in a MUCH MORE cost effective way. And these saving will means companies are investing back into AI research for further improvement and benefits. The whole AI research has created a self sustained cycle that we know, at least for the next 5 - 10 years will not be lacking any fundings.
I agree that AI is hyped, but I believe the problem is that we absolutely don't care about 50 years of neuroscience research. We know so little about the brain, but already much more than back when "artificial neurons" were modeled. The only company that I believe is on the right track is Numenta, they focus on reverse-engineering the neocortex. They have a living theory that is updated every time there is a research breakthrough.
AI definitely isn't stuck, unless you define it solely as creating artificial general intelligence. The problem there is that we don't understand general intelligence very well at all.
Of course the fixed graphical models we use have their own problems. For instance, we can't even effectively model a neural network with a variable number of inputs.
I'm not sure why you think neural networks can't handle variable numbers of inputs. Recurrent networks that ingest whole sequences have been around for a long time, and other structures have their own network topologies. Support for things other than classic RNNs is more limited, but e.g. TensorFlow Fold (https://github.com/tensorflow/fold) was specifically designed for that.
RNNs are a good model for things that are naturally sequential with limited state transfer. They are not so elegant for things with no defined ordering and a large amount of shared state.
Can you give an example of a problem where you have "no defined ordering and a large amount of shared state"? What kind of model is typically used in that domain?
One problem I've been considering is triangle surface meshes. The data is variable in size, with no defined start or end point, where points distant on the surface may share a high amount of mutual information (through symmetry, etc).
One approach I've thought about is applying kernel methods. You can compose kernels, so they scale up cleanly regardless of variations in the input dimension. The sum or product of kernels between each node in the input graph and some basis set is itself a kernel. If your kernels describe covariance between observations (i.e. Gaussian processes) then additional input dimensions have a constraining effect, rather than causing evidence inflation for larger inputs as a typical neural network might.
The article is calling us to teach AI to read words and phrases and look for the meaning, not just a statistical correlation.
I put dictionary data into Pingtype English to try to parse phrases instead of just words. e.g. "pick [something] up". The purpose is to do word-for-word translation to Chinese as an educational tool. It's not perfect, but the dictionary is editable. You can contact me if you want to discuss new ways of extending the features (e.g. data from UrbanDictionary, movie subtitles, etc).
I also want to correct the author that CERN does not have billions of dollars of funding. There's only about 5000 staff, and the other 10,000 people working there are funded by universities elsewhere who send them to CERN to do the research.
Do you know about Linguée? It's a database of human-translated texts with flexible search by phrase. One of its language pairs is English-Chinese. Reverso is the same idea, but doesn't have Chinese.
A billion dollars' research into AI would be more beneficial than another billion spent on CERN.
The Standard Model covers well over 99% of known physics already. Other moneys are being wasted paying students and professors to study string theory w/o any experiments possible.
Let's develop some true AI and let it close the gap. Two birds, one stone. [And maybe we can find out how we do analogies, at the same time!]
Most of the AI progress in last years is just tuning pattern recognition algorithms. We can not expect these algorithms to produce results like humans, because humans have a lot of information not just from percieving the world, but their patterns of thinking are also vastly dependend on the underlying structure of brain, that has developed over milions of years of evolution.
If there is a cliff, toddlers are scared of being nearby. They definetely don't have the ability to "imagine=simulate" the consequences of falling over the cliff. The fear is in the structure of neurons of brain.
If you feed classifier algorithm with images of black dogs and white swans and then want to classify black swan. Both classifying it as dog(because of color) or swan(because of shape) are right. The difference is only in bias, which features do you prefer.
You must watch the documentary "The secret life of babies". There is a scene where they put several babies on a "cliff" situation (with a glass in the same level to prevent them falling), they will confidently march into the glass - they do several other experiments that prove babies are fearless.
Babies are not born afraid of the cliff (i.e., it isn't hardwired into the infant brain); they develop fear of it at about 9 months' age, once their depth perception has developed.
> If there is a cliff, toddlers are scared of being nearby.
I don't know about a study proving that, might be true. But from my own experience, toddlers are not afraid of anything until a. they hurt themselves, b. they develop more and understand the concepts like height, c. the parent repeats "no" to them and/or shows them what to do or not to do until they learn.
So there might be evolutionary pre-programming in the human brain, but toddlers brain still needs to develop until those became active. I think there should be more research into how toddlers learn to crawl, stand and walk, how they learn to speak, etc.
https://youtu.be/WanGt1G6ScA?t=92 (from that video, at a timepoint which shows a young a baby crawling over a glass surface with a drop beneath it, with no sign of discomfort).
I'd say this proves the point completely. It shows that babies aren't scared of heights, and they don't acquire the fear from experience. Instead it seems likely it is that they either develop a mental model of how the world works, or they learn from watching other's behavior.
It's great to read someone challenging the hype, but this analysis is a bit too negative. I'd point out that AlphaGo did something a little bit like the author's toddler learning to squeeze through the back of her chair. AG learned its new moves by playing against itself. I still wonder how they avoided over-fitting here. Imagine two twins who loved to play and learn about go, and played constantly against one another, and in the process discover new plays they both believe to be unbeatable. You'd expect some of their new moves to be weak when actually tested against other players. But AG's new moves really were strong.
Admittedly, games with simple scores are the only scenarios where this really kicks in. But then again, the stock market could fit this model.
This article is actually pretty accurate in that it identifies that neither academy nor industry is well suited to solving AGI.
Suppose that a real solution to AGI will actually take 10 years to solve with minimal milestone achievements along the way. In other words, until you have the complete system figured out, it'll be hard to see the results.
In academia, most people are ultimately focused on getting their paper published.
In industry, most people are ultimately focused on making a profit.
In both cases, people would get off track long before they reached the full solution.
Lastly, the principles behind which a real AGI operates are likely so abstract that everyone reading this will likely be long dead by the time humans stumble upon them.
The only way we can short cut this process is by looking at the solution (ie the way Numenta is doing it).
"Artificial intelligence will play a big role in the future of U.S. strategy in cyberspace, according to National Security Agency Director Adm. Michael Rogers, who told Congress Tuesday that relying primarily on human intelligence 'is a losing strategy.'”
The lineage of military intelligence systems using AI is (necessarily, historically) heavily biased toward language-based AGI ("old AI") rather than neural networks. The NN are there of course, but IMO the impressive work is in the AGI.
What do you do when this software discovers it's rewarding to compete with you for physical space?
The problem with AGI is it'll implicitly have to model the entities it interacts with and that may present two challanges:
1) Developing robust strategies for managing an AGI discovering a greater reward response from defecting than for cooperating with people / developing strategies for managing scenarios in which a quorum of AGIs discover it's rewarding to collude to the detriment of humans / cellular life.
2) The tractability of maintaining one language model per entity across channels.
The actual implementation could be done by plugging a handful of related techniques we've developed over the past couple of years together though.
Why does everyone assume human intelligence is computable? Seems we should be checking that assumption at this point since we've made so little progress, and a definitive answer is much more valuable than this ongoing speculation.
It's alright, but I don't recall the book articulating actions computers cannot do. It seemed to still leave open the possibility that we can automate all human work. What we need is a precise task that humans can do with ease but we can prove is impossible for any computational device whatsoever.
My understanding was that he thinks we can potentially automate all human work, just not with computers. A precise task isn't necessary: AI researchers are simply mistaken about the nature of intelligence.
I'm not sure about "stuck", or that a huge international affair like ITER is a good solution. We could have maybe AI with our maybe fusion for $70 billion in 40 years ;)
But watching my 1yo learn to toddle around and navigate does show just how limited current AI is. With tons of training and battery, we can coax a computer to barely do what my 1yo does on a belly of cherrios and a few hours of trial and error.
There's lots of great stuff and some terrifying stuff happening in AI and I don't doubt more to come, but watching kids learn puts it in perspective for me.
We humans don't know how to learn. We don't know how learning works. We simply work work work work until we know whatever we set out to know, we don't learn how we learned it, but are happy that we simply know it and leave it at that.
Therefore teaching someone/something else how to learn will be almost inherently impossible, because we don't understand it ourselves (yet?)
And if we do learn how to learn, why would we need an AI to do it for us?
Take any NP-hard problem at its simplest form. Whatever half assed heuristic you use chances are the solution becomes optimal (if it is even half good.) The more complex the problem the more the heuristics fall apart. For the smaller ones we can simply brute force it. The same thing happens here. Compared to something like face recognition, recognizing simple shapes is much simpler. So even the not so great techniques (comparatively) work well enough. They fall apart when we use the same methods for the complex stuff.
A great example would be a greedy algorithm. It works for some problems but doesn't for some others. Take a simple enough problem and you get optimal solution. Push the algorithm to its limits and you don't even get a good solution. You don't have to understand how the best algorithms for a task work to come up with a greedy algorithm.
We (humanity) have made huge progress to understand images in terms of content and emotions of people. Imagenet is truly a gift to the world. However, that has brought us only a small but important step forward. Clearly expectation has to catch up to reality. However, all these solutions are becoming quickly more accessible to the laymen bringing another boost to operational efficiencies for companies worldwide.
> We (humanity) have made huge progress to understand images in terms of content and emotions of people.
We don't have systems that 'understand' anything, nor do we have anything remotely close. We have systems that can associate an arbitrary tag with an image based on a set of patterns that exist in the image. That's a pretty awesome accomplishment, but it is a tiny fraction of the problem of 'understanding'.
I actually do feel that expectations are behind reality, at least amongst those who are just barely too smart for their own good. I still see comments daily on HN or Reddit that promote the narrative that there is no AGI, people only work on ML, and all ML is a narrow party trick. And I think that is a terrible characterization of what, e.g., the computational neuroscientists are doing. Peruse some of the research happening at MIT and Stanford right now, and I don't see how anyone can cling to the "it's just ML" canned response.
Maybe because, gods forbid, people judge by real-world results and not by the words of a bunch of narrow specialists patting themselves on the back?
The author's points still stand. Robots do fall over trying to open doors and they don't invent new ways to climb a chair. This is a fact. The terrible characterization you speak of is well-founded in observable reality. That is a fact as well.
If this is your attitude towards long-term academic research, there's little hope I could expand your awareness of what the state of the art in AI actually is. I'm reminded of a point Yudkowsky made 10 years ago:
I am sure you misunderstand me but I gladly take the blame for it. I am all for people doing experiments just for the heck of it and being paid for it -- we as a race need a lot more leisure and discovery time. We're being robbed of leisure and discovery time more and more with each passing year, we always owe somebody money, there's always something else that is urgent to do, and in the end we never get to just slack for a year or two, especially after a burnout -- something that was deemed very normal even only 50 years ago. This is an awful period of human history and one I am sure will be remembered with great deal of shame one day. But let me not digress a lot...
I am 100% behind science, experimentation, and even silly / goofy discoveries whose usefulness might come centuries later (or never; I am okay with that). Please don't get me wrong. We need much more of that as a race.
I will also immediately agree that I am oblivious to what is happening in the AI area. But can you blame my cynicism? Everybody, their dog, and its butler are now claiming to do "AI innovation" and in the end 99% of them just swallow investment dollars, figure out a lucrative exit, and some even repeat that a year or two later. Naturally, people get worn out and start putting snarky remarks when they hear the now-meaningless term "AI" -- I am one of them, and I don't feel bad about it. I believe the sarcastic attitude is well justified.
Everybody keeps praising certain, very specifically tuned, NNs when they do certain very specific tasks. Fine. I will grant you that I can't code the algorithms needed to surpass human doctors in recognizing latent cancer or any kind of early signs of a dangerous disease. This is true. But the current way of doing things is like "input heckton of data, go to lunch, expect magic when you return". It definitely feels like it, even if I know that it's not factually true.
NNs show bias. Seems like nobody cares, they're like "yeah we know it's a problem, we'll get to it" and yet there are NNs that very likely already deny black families loans due to the inherent bias in the datasets they've been fed with. The concept of implementing a truly explainable AI seems to be very new when it had to be there right from the start and shouldn't have ever been missing; what are you people even thinking?! A driverless car makes a strange decision and what, "the NN worked perfectly"?! Bah.
To me, "AI" advocates are very content to deny very real issues that exist RIGHT NOW and that makes me cynical about that branch of science since you guys always seem to try and sprint into the future while blindfolding yourself about things that need attention here and now.
I admit I got off on a tangent. In any case, these are my collective thoughts on the topic.
We are using the wrong computational paradigm. We have to abandon bits and go back to analog computing in the form of analog photonic computing that gives you fast Calculus. This is painfully obvious in the case of neural networks, which run faster on an analog computer and are also easier to program.
I think you are out of date. Rectification is pretty good at removing a lot of the vanishing gradient issues that nn's used to face, and the overwhelming power of modern digital computers (50k cores is common) make this all moot as far as I can see.
Nope, they aren't even in the same category. E.g. Notice that there is a cpu size limit because of heat. Do you know what doesn't overheat? A photonic computer. You can build a cpu the size of a house. Also how does the number of cores constitute "overwhelming power".
:o) well, due to the wavelength of light a photonic cpu has to be 50->100 times the size of a current gen electronic one.
When I were a young 'un we had one core, and it ran at 25Mhz and about 130 of us shared it. Now I have 50,000 cores that run at 2 Ghz and five people share it. Things aren't quite directly comparable but the speed up is at least 100,000x I am overwhelmed by this, things that would have taken 1000 days; approximately 3 years, can be achieved in ten or twenty minutes. In reality the use of these infrastructures has enabled (in neural net land) the development of techniques that improve performance by several more orders of magnitude - so things that would have taken several years are now done in a minute or so. I believe that there is plenty more headroom to be had.
How is an analog computer "easier to program" than a digital computer? Making neural networks do what you want is hard enough with the help of tons of libraries, decent scripting languages, the ability to dump the weights into a file and inspect them, etc. Programming with an analog computer, which I'm guessing would be something like programming with FPGAs, sounds like a nightmare in comparison.
Because neural networks are fundamentally dynamic systems that are much easier to model with continuous signals than discrete bits. A lot of the hardness comes from the fact that you are discretizing fundamentally continuous signals.
Well, yes, that's where some of the hardness comes from... but how do you build a predictive model capable of transfer learning without disentangling the latent factors? How do you perform one-shot learning without being able to lay down discrete episodic memories?
I'd say that rather than a continuous analog data stream needing an analog model, the real problem is that the causality (hence predictability - the goal) of this data stream is due to discrete actors and actions and therefore we need to discretize the stream into objects and spatio-temporal events.
Anyhow, we're making great strides with ANNs on the perceptual side to the point where it's almost a solved problem... What's lacking (outside of DeepMind) is more of a focus on intelligent embedded agents, complete with lifetime continuous learning, and adaptive behavior. IMO we're focusing too much on artificial isolated problems rather than the embedded systems/agents that are the real goal!
Just as ImageNet - and human competitiveness - drove vision research, what could accelerate AI research would be a similar annual competition for embedded agents (either in a simulated environment or maybe robots in a competition space), which would at least focus efforts on building systems and addressing the goal of AI rather than breaking it down into someone's (maybe incorrect) notions of the piece-parts necessary to get there.
Some people shy away from robotics as an unwelcome added complexity, but that never stopped the popular micromouse competitions, and these sorts of competition could go a very long way with simple robots/vehicles (e.g. based on Lego mindstorms or R/C vehicles) with remote compute.
Quite the opposite, neural network research & experiments show that discreteness isn't a problem - in particular, there's no benefit on having a model with more fine-grained values and that even extremely discrete models (e.g 8 bits or less) work quite well.
So a three year old finding an unanticipated way to slip out of its chair is evidence that it is smart, but a neural net finding an unanticipated common pattern to all school bus images in its training set is evidence to the contrary?
I didn't read this article for the simple reason that NY Times is not the place to learn any insight about AI like this. They are article churning machine that serves political propaganda and useful local news and analysis. Anything science based, not really their thing.
The article is riddled with errors that undermine its own thesis.
It starts badly:
> Artificial Intelligence is colossally hyped these days, but the dirty little secret is that it still has a long, long way to go
This is not a secret, let alone a dirty one. Even 5 minutes casual research into the state of AI will reveal what it can do and what it can't.
It says:
> Such systems can neither comprehend what is going on in complex visual scenes (“Who is chasing whom and why?”) nor follow simple instructions (“Read this story and summarize what it means”).
In fact comprehension of (very) simple stories is now more or less a solved problem. I wrote about performance on the bAbI tests here:
Summarisation of arbitrary video is harder but given that object and path extraction already works well, it doesn't seem very implausible that we'll see some good research results in video summarisation systems within a few years. Extrapolation from what's happening to hypothesised explanations is a lot harder but not hard to imagine it being possible given the direction research is going.
> My daughter had never seen anyone else disembark in quite this way; she invented it on her own. Presumably, my daughter relied on an implicit theory of how her body moves, along with an implicit theory of physics — how one complex object travels through the aperture of another. I challenge any robot to do the same.
> To get computers to think like humans, we need a new A.I. paradigm
That's not clear at all, given recent research. It is an odd statement from someone who has worked in AI. But then as the author is not a computer scientist, perhaps not that odd.
Modern neural networks are so similar to how humans think that psychological techniques are being used to understand and "debug" them:
I'm not sure how "think like humans" can be easily defined, but using strategies developed to understand human thinking on robots seems like a good starting point. Making mistakes similar to what you'd expect humans to make is also a good sign.
> But it is no use when it comes to top-down knowledge. If my daughter sees her reflection in a bowl of water, she knows the image is illusory; she knows she is not actually in the bowl
She does now. But it takes time for babies to learn how to interpret mirrors.
Animals usually never learn this, though a few very intelligent species can.
I don't see any obvious theoretical reason why image recognition engines shouldn't be able to understand mirrors, given sufficient research.
> Corporate labs like those of Google and Facebook have the resources to tackle big questions, but in a world of quarterly reports and bottom lines, they tend to concentrate on narrow problems like optimizing advertisement placement or automatically screening videos for offensive content.
Another bizarre statement given the author's background. Google and Facebook have been investing massively in very long term AI research and building many things along the way of no direct commercial value, like AIs that play games. I don't see Google's public AI research focusing on the cited problems, although it would not surprise me if there are parallel efforts to apply research breakthroughs in these areas.
> An international A.I. mission focused on teaching machines to read could genuinely change the world for the better — the more so if it made A.I. a public good, rather than the property of a privileged few.
And here we have it ladies and gentlemen .... the reason the article is so filled with factually false and logically dubious statements. It is an advocacy piece for new social policy: a vast new government research investment in academia, in which presumably Mr Marcus would like to be employed (rather than at Uber).
Besides, even this last paragraph is disingenuous. There does not seem to be any risk of AI becoming "the property of the few". In fact the large corporate research labs are doing fantastically well at publishing research papers and making the results of their work publicly available and useful ... in fact given the relative quality of corporate vs academic open source releases I'd say they're doing better than academia is. It's hard to imagine universities producing something as robust and well documented as TensorFlow.
- Pursuing low-hanging applied-solutions fruit under the guise of a grand mission statement of (AGI) [which is now all the rage], results in one becoming (stuck). You maybe can fool investors and the lay with such madness, but you can't fool yourself nor the matter at hand.
- Not staffing or structuring like you understand or respect what General Intelligence is results in a narrow and specialized mindset among your employee base that produces narrow and specialized solutions.
It's called a local minimum. It's where you land when you don't focus on the bigger picture.
> How to move forward?
There's several techniques for that.
I don't see them being used. Which either means they don't understand they're stuck or they know they're stuck and don't care.
Why would the latter mindset be willfully chosen?
Current models and methods for training AI require huge data sets and computational power to be effective. Who currently maintains such resources? Whose fueling and molding the perception and direction of current efforts? See the conflict of interest?
Furthermore, given how convoluted the approaches/math are, it lends itself to specialized individuals...PhDs. A match made in heaven that allows the market to be narrowed and segmented to a specialized group of people. The problem with this is : It results in narrow and Weak AI.
> Not knowing you're stuck
Enough people have made sound arguments. You either grasp them and change or, given how comfortable you are, stay the course. It could be, even with a PhD and clout, that you're just not that intelligent enough to grasp the sound arguments... But, if this is the case, do you really think you're going to solve general intelligence?
Those that (truly) seek to move forward have been moving forward with (AGI).
> Those attempting to preserve old business models with a fresh top layer coating of the new.
> Specialist who refuse to respect anything beyond their group's chosen methods and thus respect the scope of AGI.
> Well-funded groups who exclusively hire from a narrow scope and narrow specialized focus
> VC groups that only invest in low hanging fruit applied engineering ventures
> VC groups that don't give those in (true) pursuit of this funding
Will just get let behind with the new wave.
It's the same as it's always been. You maybe can fool yourself and others. However, you can't fool the laws of nature and the universe.
Enough people have spoken. Enough hints have been given. Enough people have taken and borrowed concepts of the small fry and called it their own only to find themselves lost in what it meant.
Enough time has elapsed. If you're not acting and steering your resources in accordance with the new, you just get left behind grasping the old.
Same as its always been...
(True) disruption.
It's on the horizon.
It's coming.
So, keep your eyes peeled.
Some of the best image-recognition systems, for example, can successfully distinguish dog breeds, yet remain capable of major blunders, like mistaking a simple pattern of yellow and black stripes for a school bus.
That's exactly the problem. Robots lack sanity checks because they lack real understanding. If you cannot recognize an object that is far away, you are instantly aware of your inability to identify this object. A computer just runs its code over it and outputs complete garbage, and this nonsense then enters the system and does who knows what damage.
Plausibility checks are incredibly complex! If you are in central Europe and you are not in a zoo and you see a leopard fur pattern, it's probably not the living animal! And so on.
Image recognition system just recognize images. They essentially do the first pass of what your brain can do.
You too can mistake yellow and black stripes for a school bus or see an actual leopard in Poland. That's when you put what you've seen in context that you rule out the idea. And if you really want to see something in a picture, you will, especially with faces.
It is no different with computers. You train your algorithm so see school buses exclusively and it will see school buses everywhere. Conversely can also teach it context, for example by taking account of the webpage hosting the image.
Computers algorithms usually have a confidence rating too. They can tell "definitely a school bus (99%)" or "looks vaguely like a school bus (30%), but it may also be a wasp (10%)", so they can be aware of their own flaws. In fact, confidence intervals are often a key part of machine learning.
Research seems to be cracking along with AlphaGo, self driving cars and the like. Recently DeepMind have been doing interesting stuff with dreams[1], imagination[2] and body movement[3], the last one being a little reminiscent of his daughter inventing a way to exit her chair mentioned in the article.
Re government intervention it's not something like CERN where you need billions in capital and it's not an area where a big government project is likely to be the best use of capital.
[1] https://www.bloomberg.com/news/articles/2016-11-17/google-de... [2] http://www.wired.co.uk/article/googles-deepmind-creates-an-a... [3] https://www.theverge.com/tldr/2017/7/10/15946542/deepmind-pa...