As a father, the output feels really familiar. It's like a child learning to talk. At first, though the words they say are actual words (and mean something to you), they themselves have no idea what the meaning is. Eventually though they start understanding the meaning, which combined with the syntax creates a person who can communicate.
I wonder if all that's missing is just a few more layers, and another source of input. Maybe a list of requirements/output/input matched with the code so it understands why what was written was written. I wonder what would happen if you ran the program, took the output, and fed it back in as input.
I think you are right in that other inputs are needed to decipher meaning. Humans for example tend to have quite a lot of different sources of input -- as when we are children and learning new words we have the spelling (visual), how it sounds (auditory), and possibly another image that shows what the thing means ("cat"). Or maybe we have the auditory ("mommy") and the visual (the child's mother). If you were trained strictly on text, then the meaning of concepts is harder to decipher. It might be why abstract concepts like higher level math are hard for a lot of people to grasp -- their only exposure to the concepts is usually just in the form of text.
As an exercise, when I think of the word "circle", images of circles and spheres show up in my head. Also the equation of a circle. My quick definition of it would be "a perfectly round object" which leads to questions of what "round" and "perfect" mean. The more I think about it, all my knowledge seems quite circular in that there are no axiomatic concepts, everything is relative and it just builds on itself. I wonder if that's the key to decipher meaning, increase the connections of the web -- with strong enough references you can pinpoint which of the nodes in the web something refers to.
What about programming, for example? It's entirely abstract and not necessarily have any visual representation. Programming is best learned through examples. E.g. "here's a line of code, here is what it outputs. Now try to figure out what the rules of the language are."
In the case of this article, the NN isn't being asked to do any abstract task like "decipher meaning", but the very concrete task of "predict the next word". As the article shows NNs can do this fairly well.
> when I think of the word "circle", images of circles and spheres show up in my head
There seems some evidence that this stuff is fairly central to human intelligence and the ability to visualize in 3d is kind of hard wired. Deciphering meaning is approximately "seeing what it means" which can correspond to visualizing it in your head. For example "the cat sat on the mat" is a bunch of symbols but someone or some machine can convert that to an image of a cat sitting on a mat then I guess they've understood it.
This part of your comment, "...there are no axiomatic concepts, everything is relative and it just builds on itself", reminded me of this Marvin Minsky paper. If you haven't read him before - enjoy the ride!
As a father, and as someone interested in this discussion (about "child learning to talk"), I think you will love Prof Deb Roy's insights into how his infant son learned language.
As one of the other commenters pointed out - it is like a tree (words/concepts) branching out from one another. I would be fascinated by seeing if this research can be continued into adulthood, where the individual "concepts" aren't as important as the interplay between them.
Human children have the great benefit of interactively learning from their parents and other humans raising them. Could we expect a child to learn to speak if they only heard recordings of existing speech with otherwise no human interaction/feedback - correcting them or offering customized and contextual new bits of information? It would be interesting to add feedback path for human corrective input. i.e. because it's direct interaction, feed it back but somehow weight it a little more than just another corpus input.
> Could we expect a child to learn to speak if they only heard recordings of existing speech with otherwise no human interaction/feedback - correcting them or offering customized and contextual new bits of information?
I once asked a similar question on some online forum [1] where many linguists hung out. My question was if an English-only speaking household left a general interest Spanish language TV station on most of the time when they weren't actively using the TV to watch something, so that their child received a very large exposure to Spanish language programming (news, sports, soap operas, sitcoms, movies, etc) from birth onward, would the child naturally learn Spanish?
I don't recall for sure what the linguists who responded said, but I think they all said the child would not learn Spanish from this.
When I was little (5-7 years old) I had quite a few anime videos and magazines in italian sent to me by my parents who were abroad. Where I lived no one knew a single word of italian.
I often watched and rewatched those videos and read those magazines without any other external input. I can tell you that doing that I easily learned the language.
When I was 8 years old I also left for Italy and in two weeks of time I already started speaking fluently, albeit with a few mistakes.
If the child will actually watch the Spanish TV he will learn the language.
EDIT: Even now I often learn new japanese words (and remember them) just by watching animes. The difference is that now I have english subtitles but back then I had no subtitles, only the images to help me understand the meaning.
But this is a little different, in the AI we want the ability to form syntactically correct sentences, but also some intelligence behind the sentences too. You as a human had another foundation of intelligence to lean on, your native language, and an understanding of the world outside of learning the Italian language. If you had no other human interaction, would you have learned any language? That's the more the situation of these AI algorithms.
I was only arguing that the child could actually learn Spanish, nothing more.
Not knowing basically anything about AI state of the art what stops us from feeding a RNN image data and text data and make it correlate them automatically by context? Just like a child learns words by hearing them many times in similar contexts so could a RNN.
I imagine the biggest problem is gathering and structuring the data. We humans receive lots of data and have lots of time to process it in our lives compared. And by lots I mean difference of a few orders of magnitude. It's amazing what this thing learns in just a few hours of processing.
I've seen something like this in action with young kids who are given a tablet and stumble upon cartoons they like on YouTube, but in a different language. After they watch a few cartoons, YouTube's recommender system keeps offering them more cartoons in that language. And it isn't long before they start spouting words and phrases in that language.
There are some recent examples where people have trained a collection of large nets which are then used to teach a smaller net. The smaller net can learn more quickly and finally achieves better performance than the large collection.
The methods involve providing more detailed feedback at each example. With most training data used now, we give a 0 or 1, does this example belong to this class. In the teacher networks, they were able to teach with more subtly: this is definitely not a car, it is very lizard like and a little snake like.
Geoffrey Hinton gave results to this effect in a talk about "Dark Knowledge" [1]. Haven't seen any of these results published, though. I think he mentions something in the talk about NIPS rejecting the paper.
I'm convinced that voice pitch, syllable meter, hand gestures, facial gestures, and other forms of non-word emphasis are also crucial to figuring out basic phrase chunking and word types, with minimal interaction, early on.
Seems it would be far harder to infer the basic initial structure from just plain text.
Well not exactly, if we had a cost function as intelligent as a human, the cost function to train the AI, would be an AI. Or maybe I'm completely off base here...
A machine will never get the meaning of a word, unlike a very small child. I am simply amazed by the fact that a child can learn a language, catch what a question is, offer an answer, say no (and how they like to say no), and all. As much as I wish it was possible, that much I believe it's not. The best we can do is put our knowledge of our ability to infer meaning of words into machine code.
So far scientists haven't found anything special about the human brain that can't be mimicked by a machine. Given enough neural connections, and a large enough data set, and a long enough training period there is no reason to think that a machine can't do everything a human brain can do.
Put another way there is nothing magical about a child learning about the world. A child's brain is just a large neural network being fed patterned data over the course of many years by a variety of extremely high resolution analog sensors. Eventually the child begins to respond to the patterns.
Not really, there are clearly epigenetic changes to neuron DNA w/r to memory formation and I don't think anyone has estimated what kind of computational firepower that represents.
Second, the 3D topology of a neuron is IMO more complex than reducing it to an FP32 activation threshold (all IMO of course).
Finally, I have to admit as a former biologist, I'm intrigued by microtubule activity and it seems like Dileep George and even Geoffrey Hinton are heading towards smarter but fewer neurons as opposed to just increasing the neuron count. Not surprisingly, the deep learning digerati are resisting this notion mightily just like the SVM peeps harped on neural networks until they kicked them in the keester.
TLDR: It's still early, and I'm biased that there are some interesting twists and turns yet to unfold here.
If you can computationally define how different common neurotransmitters affect the function of neurons at a broad, high level, then you can create your "psychoactive drug" by just writing a routine that excessively applies the function that those neurotransmitters represent.
An artificial serotonin reuptake inhibitor would just allow the serotonin-like activity to more active in the model.
The parameter Karpathy call 'temperature' seems not dissimilar in effect to a psychoactive drug, low temperature corresponding roughly to sober and high to being a bit, well, high.
> A child's brain is just a large neural network being fed patterned data over the course of many years by a variety of extremely high resolution analog sensors. Eventually the child begins to respond to the patterns
Seems a bit early to jump to the conclusion that we understand cognition. We don't. I agree that there is nothing exotic or metaphysical about brain meat, but really we're still feeling around in the dark with respect to how thinking occurs.
I'm confident that we'll get there eventually though.
My guess is that it's probably a bit like evolution in that fairly simple pressures and rules carried out by an astronomical number of times across a huge number of individuals interacting yields surprisingly complicated outcomes.
> So far scientists haven't found anything special about the human brain that can't be mimicked by a machine.
mimicking the brain's power-consumption-to-compute-power ratio is difficult, if not impossible, with today's technology.
an aside: since reading an article about the potential role of quantum mechanics in photosynthesis, i've wondered, as a lay person, whether quantum mechanics play a role in human cognition.
Theoretical physicist Roger Penrose is a proponent of this view, but theoretical computer scientist Scott Aaronson presents a rebuttal of his points [1]. Another article claims that the distance between synapses is two orders of magnitude too big for quantum mechanical effects to be effective, which seems like a plausible rebuttal to me [2].
There's regular quantum mechanics which underlies all chemistry and that you can use to calculate molecular properties and then the woo woo kind which Penrose seems to propose as behind consciousness on the basis that both are a bit mysterious so maybe one causes the other.
We can do the same thing with computers real soon now.
Haven't people been saying this for decades? AI has a long history of impressive results, but somehow none of them have actually produced "thought".
Nobody even understands how the brain "thinks" at a neural level, let alone how to model that. All we can do at this point is try different models (which way or may not actually match reality) and hope we find one that works. But there's no evidence that we'll find a working model "real soon now". Impressive results that we can kinda-sorta imagine being the product of an intelligent system haven't historically been enough.
A handful of years ago I put together a computer fully loaded that gave me 1 teraflop of commuting power.
Today I can put together a computer the same size that will give me 32 to 50 teraflops of programmable computing power.
Many of the "AI" advances since 2007 are just running old 1970s-1990s AI algorithms on faster and faster and more parallel hardware. If you have to train a model for a few hundred trillion instructions, but your CPU only does 20 operations per second (and you have to share it with 1,000 other people), you can't iterate your science fast enough to make progress. Now we can iterate our science almost too quickly.
> how the brain "thinks" at a neural level,
Planes don't fly like birds. Birds don't fly like bees. True AI doesn't have to replicate mammalian (or avian or reptilian) neural topology.
Upvote for "Birds don't fly like bees", I like the freshness of not merely pointing out that "humans can do better", but that in effect there are several paths that avoid different constraints to get to the same point!
I do not think that we have the capacity to create a brain capable of being conscious with our current technology. Storing zeroes and ones deterministically on pieces of silicon with crammed together transistors and doing computations by what are basically logical gates is kind of limiting and inefficient. And lets say that the teraflops we're talking about are meaningful. How many teraflops do we need anyway? Shouldn't Google's data-centers suffice already for reaching the potential of a piece of gray matter that fits in under 60 cm of circumference?
I also agree that AI will never be "human" (i.e. it will be different), however without understanding how the human brain works, what chances do we have to create AI?
And we have yet to crack that nut. We have yet to understand even high-level stuff in detail, like how information is flowing from short-term memory to long-term and how we forget and why we do that (i.e. forgetfulness is surely an evolutionary trait). A brain is also fascinating in how it recovers from serious strokes by re-purposing brain structures. We have yet to produce software that is that sophisticated. And we don't even understand the brain from a biological perspective yet.
Surely huge progresses have been made, but on the other hand we may still be hundreds of years away and there's a very real possibility that we lack the intellectual capability, or maybe the resources to do it (we have a history of settling for lesser solutions if we stop seeing financial benefits, like with space exploration).
Turing-complete platforms are universal simulators. There's nothing they can't represent.
> like how information is flowing from short-term memory to long-term
Sure, we know that. The little seahorse helps out.
> re-purposing brain structures
rudimentary artificial neural nets do the same thing. they also self-specialize automatically with no innate programming (line detectors, edge detectors, eye detectors, cat detectors, all the way up—automatically).
> we may still be hundreds of years away
lol. nope. gotta think exponentially.
> lesser solutions if we stop seeing financial benefits, like with space exploration
can't do space exploration without the approval of a nation-state. can do AI tomfoolery in your own basement with nobody else finding out until it's too late.
> Storing zeroes and ones deterministically on pieces of silicon with crammed together transistors and doing computations by what are basically logical gates is kind of limiting and inefficient.
And yet it is less limiting and more efficient than pretty much all analog computing devices we have built. I don't think the hardware is the issue anymore, I suspect that with the right models and training we can have thinking machines.
Jeff Hawkins' team of researchers and the people behind NuPIC and Numenta.org, at least, given how it was explained to me, believe that the human brain does compute digitally (ie the analog values don't matter, the presence or absence of the signals do). Geoff Hinton also appears to believe that the biological neural signals are interpreted in a binary way.
I could have misinterpreted their work, though, as I'm far from an expert, but that's what it sounded like to me.
I think what the parent is trying to say is not that it's easy (it's not) but that there is nothing, in principle, to stop us from writing a program that acts like a brain.
Yes, but computers are not whats important in calculation. Algorithms are. You could do exactly what a computer does with pen and paper (it would just take a much longer time). The physical properties of the computer don't matter in this regard.
Whether that's the case in human cognition remains to be shown (else we're taking for granted what we're trying to prove).
> Yes, but computers are not whats important in calculation. Algorithms are.
That's not correct.
> The physical properties of the computer don't matter in this regard.
That's not correct.
> You could do exactly what a computer does with pen and paper (it would just take a much longer time).
Yes, and that time matters greatly as it's the difference between practical and hypothetical. Beyond that, programs that can evolve their hardware have been shown to come up with optimizations no human could have created and thus the physical properties of the computer do matter.
>A simulated being in a simulated world is just as real in its world as we are in ours.
To be literally "as real in its world (as we in ours)" several things need to happen:
1) its world should be an 1-1 simulated mapping of our world. Perhaps not to its whole extend (e.g. not the whole universe), but to ANY extend that affects the final result.
2) its world should have randomness equivalent to the quality of randomness (not sure if it's perfect) that our world has.
As for "Anything from our world can be simulated" -- that's a bold claim, provided that we haven't simulated ANYTHING at all yet, to the degree of interactions and complexity that exist in our world.
When we simulate the behavior of water in a fluids physics simulation, or the behavior of planets etc, it's amazing how much stuff we leave out. Our simulations are to a full-blown simulation what South Park cut-outs are to a photograph.
Besides, this notion reminds me of the naive 19th century ideas, that they could predict the course of the universe if only they had the details (motion, momentum, weight, etc) of all objects and the capacity to calculate their interactions. QM put a hole in that.
Regarding 1) There is no requirement for the simulated world to be a 1:1 mapping of our world. It can be completely different, a simplified subset or whatever it likes; this does not change the premise that to the inhabitants of that world, it is real. Its not our world, but that has no relevance to anything, there is no rule that says it has to be a 1:1 mapping.
As for 2) likewise, randomness isn't a requirement, you're arbitrarily picking one quality and saying that quality has to be identical for it to be real. why? I don't believe that for a second.
I'm fully aware of the simplifications of simulation... being simplified compared to an external universe does not change the premise of it being real to its inhabitants.
Quantum Mechanics does not say that the universe is not mechanistic, just that there is a random element (that in itself may ultimately be modelled).
you're mixing your frames of reference....
The simulated 'moisture' would exist within the simulated universe using simulated water and simulated fabric (for example).
Within the simulation, that fabric would be just as moist as a piece of cloth left out in the rain in our universe.
you can't mix your frames of reference.
I often wonder if the gap is that we're so caught up on training our networks on vision and text that we're ignoring the fact that living beings have a sense of time and reward as part of their input.
A child knows that if it says "Mama food," it is likely to get attention, and if it gets attention, it is likely to minimize its hunger. Right now, a neural network can be trained to know that "Mama" occurs often in human dialogue, what words occur around it, even its dictionary definition and images of mothers. But it's not making the deeper connection to a strategy that minimizes hunger.
When I think about this, I wonder if insights from the world of gaming "AI" would be useful in developing the training datasets for real AI. Because you can't be a mother to a billion virtual babies, but you might be able to program a set of heuristics to be a mother to a billion virtual babies. Then you have some system that trains on their life experiences...? All speculation, but very interesting stuff.
There is a huge amount of research which is combining the power of deep learning for automated feature extraction with reinforcement learning for learning "natural reward signals" without label information.
See any of the recent papers from Google DeepMind, such as [1] or their most recent work which is startlingly good [2]
I'd forgotten about that! The "game" needs to get much more complex to simulate life, of course. Now I wonder if they could throw that infrastructure at Minecraft survival mode...
Physical properties? What if those kind of properties of physical materials are needed in cognition?
The problem is simulations of the brain are not "machines", they are algorithms, e.g. they assume everything is happening at the information processing level.
To use your own example, we can design an algorithm to simulate making coffee. But the algorithm can never make coffee -- unless it's fitted and connected to a coffee making apparatus.
Or take something being "wet" for example. We can emulate the motions and powers in play in liquids, but not "wetness" in the sense of the physical property (moisture etc). If something depends on it, e.g. the emulation actually watering some actual flowers, then it will fail. An emulation can only water emulated flowers.
> The problem is simulations of the brain are not "machines", they are algorithms,
Simulations are executed on concrete machines that exist in the real world. Algorithms are abstract concepts.
> e.g. they assume everything is happening at the information processing level.
Everything does happen at the information processing level. Any kind of physical process can be seen as a type of information processing. Information processing is not an abstract concept like an algorithm, for it to occur requires the time-evolution of concrete physical processes.
> We can emulate the motions and powers in play in liquids, but not "wetness" in the sense of the physical property (moisture etc).
The physical property is experienced as sensory input. Machines can have sensory input.
> An emulation can only water emulated flowers.
You are asserting that virtual reality is different from reality, which is true. That's not the GP's question. The question is whether there is a fundamental difference between machines in the real world (with sensors and arms and so on) and the human body and brain.
> The question is whether there is a fundamental difference between machines in the real world (with sensors and arms and so on) and the human body and brain.
This is pure philosophy, as no one yet knows the answers, but what if brain-like intelligence is an emergent property of non-deterministic processes? Wouldn't it then follow that a classical computer could not be able to compute the "think function" before the heat death of the universe?
personally my intuition says that strong AI cannot be encoded in silicon, or that it is a victim of the halting problem. I think we need a different substrate on which to model cognition. Or maybe not. Who knows?
I wonder if all that's missing is just a few more layers, and another source of input. Maybe a list of requirements/output/input matched with the code so it understands why what was written was written. I wonder what would happen if you ran the program, took the output, and fed it back in as input.
Really cool stuff here.