DeepMind AI learns simple physics like a baby
nature.com
nature.com
I think that's likely, but also babies have ~9 months of development, at least some of that time (3-6 months) they're moving around/ have a brain.
My first thought is, what else are the hands and eyes supposed to do? Haha. They're evolved to do those exact things.
Basically the brain has to interpret each single image, then stitch together multiple images, understand that they are related, recognize objects within them, then start interpreting differences between them as movement etc.
I think this is part of the problem in computer vision. Everyone just takes cameras and stitches them into a computer, hoping to be able to figure it out. It basically doesn't work. I think there needs to be embodied vision. I'd love to see any research that explores this. I have an intuitive suspicion that our vision system and processing is tightly bound to our body, in particular our neck and hands and legs. (This is unfounded, since I'm not a biologist or cognitive scientist, and is just my naive hunch. I'd be surprised if it's far off though.)
Related to your hunch, I'm also speculating, but I'd think at the very least our eye muscles are an integral part of our vision system, it's not just the retina itself.
By the way, if anyone happens to know of the study I'm thinking of, I'd love to have a link. The last time I tried searching on Google / DDG, the results were drowned by unhelpful articles for dog articles.
A child learns that petting a cat is better than hitting the cat. Petting feels better, and the cat will stick around for longer. But how does AI learn the same thing without any pleasurable sense the petting provides?
Also will we see robots throwing tantrums? Children learn this tactic early, and it works to achieve desired outcomes. Later they progress to crafty manipulation tactics, such as carefully executed tone when asking for treats etc.
AI could use a pleasure function, but I'm not sure how it could justify seeking or getting more than it needs. The "pleasure" would need to be linked to a measurable benefit for its hardware or systems, and energy would need to be used to achieve that. I guess it could happen, but then we'll need greed functions to cap the pursuit of pleasure!
Sigh. Tantrum isn't a “learnt tactic” (it's the child's inability to control his frustration / other emotion) and it doesn't work to achieve desired outcome which is why they stop when the child is mature enough to have better emotional control (unless you're trying to raise a tyrannical child, in which case giving your child what he wants when he throws a tantrum is a good way to succeed…)
Is sighing a learnt tactic? How about a written sigh, what's that about? Inability to use normal words?
I'm quite sure some "tantrums" are last resort efforts to reverse whatever decision the parent made that upset the child.
In automated chip layout, for example, already one can be proposed odd NOP circuitry that reveals to be just relevant to timings - all tricks are possible.
"Sighs" could presumably be the reversal manouvers in decision trees walking.
Mine also really liked eating sand for a while :-)
They use some rudimentary logic even at that age. There is a thing that's in English called an "egg cup" (according to Google Translate :-)), there is one word for that thing in Dutch: "eierdop".
My daughter, just having heard the word "egg" for the first time that day looked at an egg cup and decided that it was "ei-bakje" in Dutch, which roughly translates to "small container for egg". Only took maybe 2 seconds.
I think that ability to find relationships between things and extrapolate very rapidly from them is what makes human kids so fast at learning. Which may have something to do with the fact that human memory is very associative, almost like what near 100% data locality would be in CPU design with many many cores.
I discovered actually 'dop' means something like 'shell'. So APPARENTLY eierdop all as one word means egg cup, but eier dop as two separate words means egg shell?
In Dutch "eier dop" does not exist. Could be "ei dop" and then you're taking about two separate objects, an egg and a cap (as in bottle cap).
And that's also immediately an example, I tried to type that as bottlecap and my phone corrected it with an extra space ;-)
As I age I am reminded by the same process you talk about but in reverse now. Watch the Robin Williams movie "Awakenings" [1]. Really portrays it. We come into this world absolutely unable to do much of anything. We fade in. Then ... trust me ... we fade out. But in between, if we're really lucky we get to shine.
I wonder if a neural network will ever do that?
However we then humanly assemble stuff like cat = eyes+fur+pointy+cute+small+meowing. A neural net has no way yet to represent information at that level, but we are getting closer there than going beyond dumb numbers. A neural net has implicitly the concept, but is not yet doing the 3d representation inside itself, which makes it harder to define something like 'ears' and 'tail'
I'm sure humans have similar instincts. For example, the tendency of young kids to avoid plants until they are shown that they are edible - possibly leading to the general preference of artificial foods over anything that might be found in a garden. [1]
I've also noticed that toddlers don't gravitate towards knives the same way they do everything else. I certainly wouldn't rely on this behaviour, but I've noticed it in my own children and others.
[1] https://www.theatlantic.com/health/archive/2013/12/why-babie...
https://theconversation.com/researchers-trained-an-ai-model-...
> Typically, AI models start with a blank slate and are trained on data with many different examples, from which the model constructs knowledge. But research on infants suggests this is not what babies do. Instead of building knowledge from scratch, infants start with some principled expectations about objects [...] The exciting finding by Piloto and colleagues is that a deep-learning AI system modelled on what babies do, outperforms a system that begins with a blank slate and tries to learn based on experience alone
>> New research by Luis Piloto and colleagues at Princeton University – which I’m reviewing for an article in Nature Human Behaviour – takes a step towards filling this gap.
I don't reckon I've seen this kind of article before: "here's some new research I'm reviewing and I think it's rad". That just sounds so... dodgy. You're supposed to maintain at least the facace of being impartial and having some sort of integrity when making reviews, even when reviewing a DeepMind paper for Nature.
I recall Tenenbaum's lab had a similar paper a few years back.
in which animations are shown which reveal the close similarity to the DeepMind project.
I have seen some suggest that this is basically Google “allowing” the competition to catch up just to beat them a few weeks later but generally it just seems like they’re kind of all chatting with each other in the background instead.
You can figure out newtonian mechanics entirely on the ground and that clearly helps you understand flight. By analogy, getting more understanding about what limits exist in ANNs could plausibly help understand how the brain works (and vice versa).
Walking and flying may not have a lot in common to you and I, but to a thing learning to crawl, there is a lot to be understood.
Maybe we can teach an ML to whack-a-mole until we see ideas pop out which the ML doesn't think are repetitive correlation/causation/model-is-working effects
All theyre doing is feeding in pixel patterns and object information for specific scenes, and then calling the models ability to detect general correlations in those scenes "solidity", "inertia", and the like.
Any time where patterns in correlations are taken as an explantory model, we have pseudoscience.
Yes, inertia/solidity/etc. is the right answer for why physical objects show those visual correlations. But pixel patterns in some images are NOT models of inertia, solidity, etc.
Just as correlations in pixel patterns in the sky is not a model of gravity. The right model, eg., is F=GMm/r^2 which are variables not even present in the data. So no correlative model of a night sky will ever be a model of gravity.
If I make a machine that walks to the store, I have shown that walking is sufficient to get to the store. I have not ruled out the possibility that you bicycle to the store.
But if there has been a long running passionate debate on whether walking could ever get you to the store, showing sufficiency can settle that.
-- DeepMind AI learns physics by watching videos that don't make sense - An algorithm created by AI firm DeepMind can distinguish between videos in which objects obey the laws of physics and ones where they don't - https://www.newscientist.com/article/2327766-deepmind-ai-lea...
> Luis Piloto at DeepMind and his colleagues have created an AI called Physics Learning through Auto-encoding and Tracking Objects (PLATO) that is designed to understand that the physical world is composed of objects that follow basic physical laws. // The researchers trained PLATO to identify objects and their interactions by using simulated videos of objects moving as we would expect [...] They also gave PLATO data showing exactly which pixels in every frame belonged to each object. To test PLATO’s ability to understand five physical concepts such as persistence..., solidity and unchangingness..., the researchers used another series of simulated videos. Some showed objects obeying the laws of physics, while others depicted nonsensical actions [with the latter, correctly the AI returned wrong predictions, showing an acquired intuition of physics]
From the submitted one:
> [Jeff Clune, Uni British Columbia, Vancouver: ]«[Comparing AI with how human infants learn is] an important research direction. That said, the paper does hand-design much of the prior knowledge that gives these AI models their advantage». // Clune and other researchers are working on approaches in which the program develops its own algorithms for understanding the physical world
Right. "You have to tell yourself stories", as the late Prof. Patrick Winston said (you are intelligent because you can predict the unexperienced). Because you need concept development and critical thinking - an active process.
If you do it well, apparently, you win being published in Nature.
but also the idea and practice of finding new tricks that may help the whole field remains there. This research seems to have found a shortcut strategy - a faster ML.
[0]: https://www.soulmachines.com/resources/research/baby-x/
Intuitive physics learning in a deep-learning model inspired by developmental psychology
They evaluated their model by producing a test set with videos spliced so as to compose physically "impossible" sequences (that's the claim anyway - who knows how "impossible" those sequences are). Then they measured the trained network's prediction error on the spliced videos and they called that "surprise", implying that it is somehow similar to the "surprise" shown by human babies when they observe physically impossible sequences (to clarify, in the psych literature this "surprise" is measured as a function of the time spent looking at a scene). Then they measured "accuracy" as the difference in "surprise" between possible and impossible videos.
There are two problems with this. First, it's the old switcheroo: it just relabels prediction error as "surprise". Yeah, "surprise" is interpreted, in a Bayesian setting, as the difference between an expected event and an actually occuring event but here we're not in Bayesian world. There's just a classifier failing to classify correctly. By the same token, any classifier's classification error ever can be interpreted as "surprise", even though it's just failure to acquire the target concept.
Second, this measure of "surprise" has nothing to do with the measure of surprise used in experiments with baby humans, where "surprise" is a function of the time the baby spends looking at a scene. For a baby human it makes sense to intrepret longer time looking as a double-take: "what the hell just happened"? But for a neural net? How can a neural net do a double-take? First of all the time spent "looking" at any image is entirely controlled by the experimenters. This is just an apples-to-oranges comparison and the result is a lemon.
Third, their "flat" networks trained without hand-crafted object segmentation show lower "surprise" but since that's just the sum of squared errors, it only means that the flat networks have learned classifiers that perform better on the test set than the proposed approach. When "accuracy" is compared, the flat networks score lower but that should be expected: if "accuracy" is measured as the difference in classification error, the model that is better at classifying across all tasks should have lower "accuracy". But that's because it has higher classification accuracy.
(yeah, that's an off-by-one error, well spotted).
Then, to test whether their system generalises to unseen objects and events, they tested it on a separate dataset, called ADEPT, and found that it scores very well. The only problems is that they measured "accuracy" again as the difference in mean squared errors between "possible" and "impossible" scenes. All this is measuring is the _lack_ of generalisation from training to test images.
I'm sorry to be so negative, again, but all I see here is playing with language and rechristening error as "surprise" and "accuracy". There is nothing that shows acquisition of "intuitive physics" like the paper keeps claiming.
I wish ML researchers (EDIT: and engineers and journalists) stopped using anthropomorphizing language. This has decades of solid tradition, but that's no excuse. Any comparison of a machine to a human misleads the public. Machines aren't like babies, artificial neural networks aren't like actual neural networks or brains. Machines shouldn't be given human names (PLATO is a borderline case).
I know this is like talking to a wall -- money requires hype -- but still, please stop doing that.
"The name of AI": we call "intelligent" in this convention that which finds solutions - normally the natural intelligence of a professional, sometimes the artificial intelligence of a computerized system. As easy as that. It works, no intrinsic issue.
This experiment: children seem to rely on expectations in learning and this ANN based system tries to implement some form of "expectation based learning". No problem.
Just to zoom right in on neural networks:
People often say this, and I never see a solid argument.
I know very little about biological neural networks.
Clearly they are very different in some respects, for example, meat vs silicon.
But I never see a good argument that there's no perspective from which the computational structure is similar.
Yes, the low level structure, and the optimization is different, but so? You can run quicksort on a computer made of water and wood, or vaccum tubes, or transistors, and it's still quicksort.
Are we sure there aren't similarities in terms of how the various neural networks process information? I would be interested in argument for this claim.
After all, the artificial neural networks are achieving useful high level functionality, like recognizing shapes.
One is the notion of time and connectivity loops - overwhelmingly, ANNs use a feed-forward architecture where the network is a directional graph without loop and some input is transformed to some output in a single pass - and weights can be adjusted in a single reverse pass, which is very practical for training. We do know that biological brains have some behavior that relies on signals "looping through" the neurons, and that is fundamentally different from, for example, running some network iteratively (like generating text word-by-word via GPT-3). We have artificial neural network simulations that do things like this, and also simulations of "spike-train" networks (which can model other time-related aspects which glorified perceptrons can't), but we don't use them in practice since the computational overhead means that for most common ML tasks we can get better performance by using an architecture that's easy to compute and allows to use a few orders of magnitude more parameters, as size matters more.
ML researchers don't write articles, journalists do.
Actual language used by the ML researchers: "Intuitive physics learning in a deep-learning model inspired by developmental psychology" [1]
Terms like "neural network" and "artificial intelligence" are frequently used by AI engineers and researchers despite the obvious image they evoke. Sometimes they even call their creations "brains". Also note the name DeepMind.
It's definitely not just the journalists.
In my opinion, this is still anthropomorphizing the algorithms. The term deep-learning is a poor representation of what actually goes on. Someone please correct me if I'm wrong, but all ML does is statistical regressions (in essence). It doesn't "learn" like a person learns. Neural networks are not actually like brains (as far as we understand how the brain works).
I feel like the whole industry is inundated with aphorisms that are kind of true, but not wholly true. Evolutionary algorithms, neural networks, deep learning, deep mind, this stuff all reeks of anthropomorphizing fundamentally mathematical processes. I get it, it's a lot easier to get the gist of "the computer is learning/training" than "the computer is refining the weights and biases to try to optimize the output".
deep learning isn't a bad name imo. it is learning, but nothing in the name suggests it's like a human brain.
Who cares? Why should you hold the idea that it would? They are systems moulded after data, 'learn' seemed to be a decent label. If it is not, it is because "learning" is _active_, by philological analysis, and happily consistently with an aim of AGI (intelligent entities learn actively).
A computer does not compute like a human would. Yet, no problem.
For that matter, you are using 'person' in a very individual way - not even "personal" (a "person" learns according to individual nature, while you are using it as a collective term).
As already expressed - nearby I wrote 'biomimicry' -, what you are calling "anthropomorphizing" is a wrong direction: "evolutionary algorithms" were born out of keys after the observation of the natural world, and the terms express that - it is not that you saw the algorithm and went "It looks like my uncle Oscar"¹ (this side is active - it "learns").
(¹Those anthropomorphizing Hollywood cultists and all that sculpture...)
> a "person" learns according to individual nature, while you are using it as a collective term
There's a very specific definition for learn:
>> to gain knowledge or understanding of or skill in by study, instruction, or experience[0]
There's a few more, but none of the definitions treat learn in a non-collective way. I guess meriam Websters dictionary doesn't like treating people as individuals or something lol.
Additionally, all the definitions there are speaking in human contexts. They talk about learning in the sense of being taught, or gaining experience, or gaining knowledge. Sure a computer kind of does this stuff, but it doesn't really. And that falls into the category of attributing human characteristics to an inanimate object.
I probably shouldn't have said that everything in the short list I wrote reeked of anthropomorphizing processes. But the evolutionary algorithm was more in line with what I mentioned immediately before. My whole comment read:
> I feel like the whole industry is inundated with aphorisms that are kind of true, but not wholly true. Evolutionary algorithms, neural networks, deep learning, deep mind, this stuff all reeks of anthropomorphizing fundamentally mathematical processes.
An evolutionary algorithm definitely falls into the category of kind of true but not wholly true. But it's not anthropomorphic.
> intelligent entities learn actively
Also, this is a very loaded statement. What is an intelligent entity? If you Google "is a computer intelligent" there are various papers, articles, and other pieces of media all claiming that we can't call a computer intelligent, and some claiming that we can consider certain algorithms somewhat intelligent. This is anything but an accepted standard today.
Give us an example of some relevant label that would be "«wholly true»" instead of "«just kind of true»". Because metaphors, and the whole system of fuzzy pattern relations, are based on fuzzy pattern relations.
> none of the definitions treat learn in a non-collective way
You have misunderstood my post. I would prefer that you read it again.
You are complaining about loose use of the language: I noted that you yourself used the term 'person' more than loosely, with a dubious jump. When one says '«like a person learns»', that is supposed to be "like a specific individual in his own individual characteristics will learn" - instead you used to say "like people in general learn". A "person" is a "definite form", not a general individual representing common features - it is the opposite.
> very loaded statement
Which you are taking out of context. I said that you have to call the moulding of your functions something, and that "learn" seems a very acceptable term, since it is bottom-up instead of top-down, it is automated instead of encoded: it is developed against data, it "learns". And that if the term is disliked, there could be a very good reason, because 'learn' was born as a sort of a hunting term¹ - it really means something like "investigate" -, which is a happy coincidence because what is largely missing in AI is critical thinking, part of the active process of learning ("learning" is active as investigation is). And the day John will have to check «accepted standard[s]» to see how things are, I will be willing to comply to his sad request for mercy.²
¹Irregardless of what the Merriam-Webster will write, because you get a none-the-wiser relative notion but not knowledge from a "dictionary of use", as at the paragraph for 'life' you will not find the meaning of life.
²John must be, tautologically, an "active learner". (He will check personally.)
Well it's not called person-learning-machine is it? Why would something have to learn "like a person" to be able to use the word "learn", those two concepts are not attached one to the other. If they were, saying "learn like a person" would be a pleonasm, yet it isn't. IMHO "learning" is a fine term, it conveys the idea of what is happening effectively and quickly.
Also, we don't know how a person learns anyway, it might very well be a similar process, just way more efficient and complex.
> Evolutionary algorithms
How would you propose calling them? You have a generation of agents, each with their own specificities, and from the agents most successful to accomplish the task at hand, we derive a new generation, slightly modified from their parent's.
It seems to me "evolution" is again the most suitable and efficient way of describing what is happening.
While I agree that there is definitely too much anthropomorphizing surrounding AI, I feel you are going way too far in the opposite direction. Not every word that can be composed with natural process/humans should be banned from being used anywhere else.
But at some point you need to name things for brevity. I understand why people say "activation function" instead of "elementwise monotonic nonlinear function".
Misuse is also rampant, like using "inference" to describe evaluating a neural network on an input, even when the NN isn't part of a probabilistic model.
Specifically, do scientists believe the next breakthroughs will come from humanity or are we waiting for AI at this point?
This leaves room for some 10 year old to do so in the near future, but places some kind of event horizon for AI achievement eclipsing human achievement closer than about 25 years into the future.
For AGI, you will have to wait.
This should be reserved to something that is truly intelligent. As long as this has not been achieved it's just simply a well trained model that results in high accuracy predictions for very narrowly defined tasks.
No. I had to write this so many times. It's pretty clear: at some point, we decided "Engineer, if you come up with a solution we will call you intelligent, and if your algorithm does we will call it intelligent as well".
There exists an interpretation of the term that makes the term sound.
It's pretty clear, really - and that "intelligence [tongue-in-cheek] [of specific solvers]" is _not_ "[intellectual] Intelligence", which by the way is not even the last and highest meaning of it.
It is the second time in this very page: «we call "intelligent" in this convention that which finds solutions - normally the natural intelligence of a professional, sometimes the artificial intelligence of a computerized system».