Artificial Intelligence Is Stuck
nytimes.com
nytimes.com
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...
1. Computers can't do X (tic-tac-toe, learn XOR, chess, anything relating to images, Go, ...)
2. (4-5 years) researches decide we need to do this
3. proof delivered
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.
These things take time and money. The profit motive is a killer of invention.
Google has a self driving car going over 5k miles without a human. Stuck?
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.
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.
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).
http://people.csail.mit.edu/brooks/papers/representation.pdf
E.g. Task: grasp an egg without cracking it
Physical platform 1: actuators, no pressure sensors
Physical platform 2: actuators, pressure sensors where egg contacts robot
Inarguably, the simplest successful implementation of the task in driving code will be much more concise for platform 2 than platform 1.
... Now generalize the same idea to trying to teach disembodied AI to be human.
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 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.
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.
That means we have a model to start with. Chicken & egg.
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
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.
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).
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.
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.
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.
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.
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.
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?
Nost of human technology today is based on replicating nature.
I wonder if simulations are the right way to efficiently train artificial intelligences
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.
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.
>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"
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?
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.
- 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.
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.
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.
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
Edit: formatting
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.
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.
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..
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.
> 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.
>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.
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.
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.
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.
We can understand how individual chemical reactions in Einstein's brain work, but that doesn't make us smarter than him.
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...
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]
I'm no expert, but he seems like one to me.
[1] https://www.crunchbase.com/organization/geometric-intelligen... [2] https://www.linkedin.com/in/gary-marcus-b6384b4/
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.
The kind of people that attend AGI conferences do.
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.
> 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.
Check out Winograd Schema: https://en.wikipedia.org/wiki/Winograd_Schema_Challenge
Overview by an expert: http://www.cs.nyu.edu/faculty/davise/papers/WinogradSchemas/...
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%.
Here are the challenge problems used in the first round: http://www.cs.nyu.edu/faculty/davise/papers/WinogradSchemas/...
Human Subject Test Performance: http://www.cs.nyu.edu/faculty/davise/papers/WinogradSchemas/...
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.
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.
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.
Enjoy the party while it lasts.
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.)
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.
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?
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.
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.
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.
[1] http://money.cnn.com/2015/03/12/technology/enlitic-technolog...
Thank you for your kind answer.
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 honestly don't see the added value of having a "more efficient" human once the classifier is good enough.
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.
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.
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.
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.
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).
http://pingtype.github.io/english.html
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.
in 2016, 1127.2 million CHF or 1,163,675,662.22 USD
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!]
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.
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.
Not sure if you are a parent, but this isn't the case at all!
Not sure what to think about it.
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.
Admittedly, games with simple scores are the only scenarios where this really kicks in. But then again, the stock market could fit this model.
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).
http://www.insidesources.com/nsa-chief-without-ai-cyber-is-a...
From the 2016 article:
"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.
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 do you suppose this?
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.
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?
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.
Woah, you feel expectations are behind reality? I feel there's a lot of news lately predicting AGI.
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.
http://lesswrong.com/lw/kj/no_one_knows_what_science_doesnt_...
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 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'.
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.
Photonics are promising QC technologies - especially Phonons, but we are a long way off!
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.
Check also real computation https://en.m.wikipedia.org/wiki/Real_computation and work of Lenore Blum.
Im super pumped about the research into computational metamaterials http://users.ece.utexas.edu/~aalu/research%20-%20page%203.ht...
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:
https://blog.plan99.net/the-science-of-westworld-ec624585e47
Summarisation of stories is also something with good recent results:
https://research.googleblog.com/2016/08/text-summarization-w...
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.
Challenge accepted:
https://www.youtube.com/watch?v=gbYiKMisbME
And for the imagination component:
http://www.wired.co.uk/article/googles-deepmind-creates-an-a...
> 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:
https://deepmind.com/blog/cognitive-psychology/
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.
http://www.thoughtfulparent.com/2009/10/child-psychology-cla...
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.
- 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.
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.
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.
And so is my second.