> A three-dimensional representation of objects also provides the basis for learning compositional structure, i.e. how objects are composed of other objects arranged in particular ways
I found this particularly interesting. I have a young daughter and I find it fascinating how quickly and reliably she can learn an object and then recognise others in a way a computer just can’t do at all. Eg see very rough line drawing of an object, then recognise that animal in real life.
It’s also interesting the way I can explain new objects or animals based on one she knows and then have her recognise them easily. Eg a tiger is like a lion but has no mane and does have stripes.
This comes very naturally when teaching and learning and feels very uniquely ‘human’ compared to the rigidity of computer vision I’ve seen so far.
I'm astounded. Zero-shot visual reasoning?? As an emergent capability?? Um. wat.
>"GPT-3 can be instructed to perform many kinds of tasks solely from a description and a cue to generate the answer supplied in its prompt, without any additional training. For example, when prompted with the phrase “here is the sentence ‘a person walking his dog in the park’ translated into French:”, GPT-3 answers “un homme qui promène son chien dans le parc.” This capability is called zero-shot reasoning. We find that DALL·E extends this capability to the visual domain, and is able to perform several kinds of image-to-image translation tasks when prompted in the right way."
It does appear pre-rendered and my guess is the blog might only showcase the use cases that worked the best. Or it might take a really long time to generate results.
It is also my understanding that it as a whole is not released.
However, iirc one of its two main components, has a smaller version of it released.
This component is called CLIP .
Given some text options and an image, CLIP will predict/guess which of the texts is a description of the image.
Very cool and interesting that the green pentagon clock had several images of non pentagons. It’s getting the general flat sides correct, but not counting the number of vertices.
I feel like that's true to how the average person thinks. My aunt told me she got octagonal tiles for her bathroom. I said you can't tile octagons in two dimensions. After a short argument she showed me a picture. They were hexagons.
That's insanely cool. I wonder, if similarly, it could be set up to pump out 3D models of things. We'd be pretty close to having an "AI storyteller" for games, movies, etc - and one that can improvise!
Well it's not that surprising when you consider that the training data for your daughter's neural network is being generated by a network of the same architecture which has been optimized to do so. We are not just learning machines, we are teaching machines. The rough drawing of a lion is your distillation of the various visual features which you consider to be hallmarks of something being a lion. Your description of the difference between lions and tigers is a summary of the features you apply the most weight to when you try to distinguish them. Even data in the wild like pictures on the internet or books in the library are overwhelmingly generated by humans for humans. Computers on the other hand are fundamentally alone - they have no ancestors with similar minds to pass understanding down to them, and must figure out everything on their own. Really machine learning is a misnomer, what they are doing is far more analogous to discovery than learning.
While Numenta doesn't have the amazing results of DeepMind and friends, I think they are doing really great stuff in the space of biologically plausible intelligence.
Seconded. Even if you aren’t ‘into’ this stuff, Jeff has some interesting ideas on how the brain works that map pretty well to my personal subjective experience. (Not saying that makes them true.)
His previous book (On Intelligence) is very stimulating and contains many ideas that have interesting connections with Deep Learning. I'm looking forward to read this new one.
Why have so many hypotheses failed when it comes to the mind? We understand quantum scale effects and measure and predict them to 17 decimals places of precision. How can we not know what is going on in the brain? We know what is going on in supercolliders when trillions of particles scatter (we only capture a tiny fraction of the data but still we know all the mechanisms available to the standard model).
Surely we can track any and every physical mechanism in the brain "accessible" to consciousness to build itself up from assuming it is physical. How can we still not take that set of physical mechanisms and build a lasting model of consciousness.
I have a personal bet that no progress will be made on consciousness in my lifetime. Maybe one of the theories we have is correct and we just can't test it correctly. I could easily be wrong. But it seems like we lack the imagination, not the technical ability, to find the answer. And not much has changed from my outside perspective in 30 years. Look, I'm genuinely puzzled by science's inability to have a working model of consciousness. I'm not asking rhetorically why we have no lasting models, I don't know what the reason could be.
To me these are even more "woo" than string theory which gets lambasted around the popular alt big-science channels.
We can't even model most two-element chemical systems using QM as the state space just becomes too complex.
Having a model does not mean you can necessarily run the computations to sufficient accuracy to make good predictions. Even with something as well-understood as gravity, N-body problems are still quite difficult and calculations must be numerically approximated. This method works fine for most systems but diverges rapidly if a chaotic perturbation is present.
This is because the sense of smell bypasses the thalamus (router for all other senses) and is located deep in the lymbic system, very close to the hippocampus and amygdala which process memories.
Well, the OpenWorm project [1] estimates to be about 20-30% into simulating the 302 neurons and 95 muscle cells of Caenorhabditis elegans. The human brain has over 86000000000 neurons (about 300 million times more than the worm). I'd say we're about 0% into simulating (thus understanding) that. Mind you previous generation computers (32-bit) couldn't even count that high.
An idle thought... are there any professional ethics standards that control treatment of the Caenorhabditis elegans worm, like there are higher animals? Or (I expect) is it considered simple enough that concerns of animal cruelty don't apply?
I wonder, because if there were, such standards should apply to the simulated copy of the worm's mind, as well.
Looking this up - apparently in most jurisdictions animal welfare/ethics standards only apply to vertebrates and cephalopods. If it's an invertebrate and isn't in the smarter category of squids and octopodes and the like, it's open season.
It depends on whether the simulation itself is conscious, which is an open question in the philosophy of mind (is our wetware the only feasible substrate?) with much debate. If the answer is no, it's therefore not capable of suffering and not worthy of ethical concern, but as of yet we don't know the answer. We may of course wish to assume that it is conscious, just for the sake of risk aversion.
yeah, while i personally suspect that a high enough fidelity simulation would be comparably conscious, and i think we should certainly err on the side assuming that's the case from an ethical perspective, i do fully agree it's still an open question.
I've thought a lot about the morality of potential openworm simulations. How tiny it is and close to being working has made me wonder about some absurd thought experiments.
Imagine a worm hell webpage that had a full neural simulation of the worm with its pain receptors all firing constantly. Would it be immoral to open that page and leave it running? Imagine instead of getting rickrolled by a link, you get wormhelled by a link and cursed with the knowledge your computer just instantiated a worm spirit to hurt it. What if someone made a webpage that had that secretly embedded in the background?
(What about the opposite: what if you had a worm heaven simulation in a webpage? Are you doing anything good by leaving that running for a while and then closing it? Does it make a difference if there's persistence or not?)
Imagine if you placed the worm hell or heaven simulation in a smart contract on a blockchain and made transactions to execute it. The simulation would be executed by every node in the network that verifies the transaction, and it would be executed again by anyone that verifies the blockchain any time in the future.
Certain compressed file and video formats contain instructions for computing the decompressed results. Imagine encoding the worm hell or heaven simulation in a short video file, so that merely viewing the video file causes the computation to happen on your device.
---
Well this is pretty moot anyway while we're talking about worms specifically, but at some point it might be doable with higher animals like mice. Or humans. Substitute different animals for worm in any of this.
My current thinking is that computations that don't use their results to meaningfully affect the outside world are meaningless. Nobody who opens the webpage that has it secretly in the background without realizing it is committing any sin. Creating or deliberately accessing the worm hell webpage (without entangling the results meaningfully with the external world) is bad only to the degree that it normalizes the idea in your mind of hurting beings that are fully entangled in our world.
If you simulated a worm hell, and then physically instantiated the resulting worm in real life (say by writing its neural data into the neurons of a real live worm), then you should be considered as having done the moral equivalent of torturing a live worm. If you run worm hell and then give the resulting worm full internet access, then you've done the moral equivalent of torturing a live worm to the degree that worms can live, socialize, and empathize online (which is not much in the case of worms, but is significant for humans).
If you run worm hell and then persist the result, it's bad in the sense that there's continuous potential for the situation to become one of the above. Sharing the results of a worm hell simulation is more bad because it increases the potential for the above.
I don't think there's any moral difference between running two identical deterministic worm hell simulations and saving the results, vs running the worm hell simulation once and saving the results and then copying the results file once. (I don't think copying the results file makes much difference except as much as it increases the risk of sharing the results.)
What does it mean to "understand something"? Why don't we consider "big lump of axons and dendrons" to be an understanding?
To me at least, understanding is quantified by the difference between the number of bits in your model and the number of bits in the thing you want to understand. "True understanding" is when you've gotten the model down to a small enough size that you can fit it in your head. But there is no a priori law that everything interesting in the universe should be able to fit in your head. We've just been lucky so far. Maybe we will keep making progress understanding how consciousness works, but never shrink it down far enough that we are satisfied with the "explanation".
I agree. This is the only definition that matters. People are confused because the language we use to talk about this stuff is full of assumptions and misunderstandings that leak into their logical reasoning. Imho, compression is a perfectly adequate description of intelligence. This is why GPT-3 does so well, intelligence is an emergent property of predictive power. Marcus Hutter really needs to rejig his compression challenge to allow much larger models, it could really be relevant to judging progress.
IANA neuroscientist but I suspect it's a measurement problem. The structures of the brain are microscopic and enormous in scale, so how do you capture all that data and make sense of it?
Not to mention, the minute you open up the brain the situation is entirely different than when things were functioning in state. It's like figuring out how a functional car works by analyzing a bag of seemingly random parts and loose scrap metal. Biology is difficult science.
It could just be that it requires scale.
Some science projects just can't be done by small research teams.
You mention our understanding of physics. That understanding only got pinned down and refined by spending billions on large research projects.
The only brain related research project I know off that has a budget even close to what we routinely spend on physics is the Human Brain Project [1], and that only started in 2013. It's not expected to be complete until 2023 and even then it is only laying the foundation for more ambitious projects to build on.
>Surely we can track any and every physical mechanism in the brain "accessible" to consciousness
To add to other replies, there's a gap in our ability to measure these mechanisms.
We can see the big picture with bad resolution (e.g. lesion studies, EEG, fMRI), or take tiny keyhole peeps with somewhat better resolution (e.g. single-neuron recording or MEAs). Moreover, each of these methods gives us just a narrow kind of indirect data (e.g. blood oxygenation, or electric activity at few synapses) out of a rich functional context.
They are amazing but perhaps still far too limited to inform truly detailed models and understanding.
Right, we would have to measure perhaps most of the neurons in a living brain at the same time to understand what is going on.
Building an artificial brain made of billions of artificial neurons and of synapses is a worthy goal but it could be like building a working computer. Having a computer does not tell us what is running in the computer. Software is the ghost in the machine. We don't want to simulate neurons and synapses, we want to simulate the programs that are running in actual brains. But what are they? And how could we find that out?
Because what those particles in a supercollider are doing is way simpler than what a neuron is doing, let alone a brain full of neurons. It's like the way you can simulate and predict billiard balls but if you take the billiard table apart and build a Turing machine out of the parts, suddenly all bets are off.
I believe we'll finally figure out some 'Maxwell's Equations of Consciousness', stare at them for a while, then collectively say 'ah crap' because seeing the maths doesn't equate to understanding what it means.
Yeah, dropout is conspicuously missing.
For those who haven't heard: Dropout zeros out random activations, creating a random subnetwork to make a prediction. Each training step updates a different random subnetwork. (...Keeping in mind that two random subnetworks can share lots of weights...) Then at inference time, you use the full network, which now acts as an ensemble of all of the subnetworks.
You could view it as 'thousand brains' with lots of weight sharing for efficiency+regularization.
Wait, why can't you? Backprop is basically adjusting weights to get the model closer to making correct predictions. Isn't that basically how the brain does it? (Though I'm sure the way the brain adjusts its weights uses a different approximation of differentiation)
Backprop requires global communication but biological brain works with local communication in neural assemblies, where signals spread in a way somewhat similar to diffusion.
There is no global “teaching signal” or “delta rule” error correction. Learning via reward and punishment is a wrong level of abstraction for fundamental cognitive tasks like visual object recognition or “parsing” auditory signal.
Sometime people mention dopamine as a kind of reinforcement signal but it operates on completely different time scale, orders of magnitude slower than any iterative optimization model would require.
And the energy and time spent on iterative optimization in ANNs is not available to living organisms with constrained resources.
If you’re interested in authoritative opinion on what kind of learning is biologically plausible see e.g. prof. Edmund Rolls recent book called “Brain Computations”.
Jeff Hawkins has a great series of open workshops on computational cognitive models (with Marcus Lewis and the rest of Numenta team), check out Numenta on Twitter. This open seminar format is kind of unique in the industry.
They're proposing a theory of how the brain's cortex must work. Check out some interviews with Jeff Hawkins of Numenta, such as the forementioned interview with Lex Fridman. One essential insight of theirs is that the cortex is physiologically the same everywhere, so any differences between regions would be caused by what happens to be connected to them, rather than something that is inherent to them. This idea of 1000 brains is sort of along those lines.
What is meant by physiologically the same? Because there is a lot of heterogeneity among neurons in the human brain, and there are definitely some macro-level differences between cortical regions that are innate (i.e. happen even with sensory deprivation). So I'd honestly be very surprised if different cortical regions did not have some differences in gene expression patterns that are "hard coded". Which can affect neural dynamics in a subtle way. Determining how much that matters for this hypothesis would require a probably impossible experiment though.
FWIW there definitely was an old paper where the animal model (IIRC a ferret?) was deafened and had its visual cortex ablated, and then they wired the lower level visual regions to the auditory cortex, quite early in development. The animal could sort of see, and the auditory cortex had developed a bunch of visual cortex like properties, which is pretty awesome. But at the same time it was certainly not the same thing as using the visual cortex to do so, if you look at the actual behavior the animal had clear deficits in its visual capabilities. Of course that is far from a perfect version of the experiment being described though.
Anyway, I read that a long while ago but here is a link to a broader review written by the researcher that did that work, in case anyone is interested in digging further: https://pubmed.ncbi.nlm.nih.gov/16272112/
He's referring to the Columnar hypothesis [0] - that all so-called "cortical columns" in the neocortex have the same basic computational behavior. There's controversy about this, the main issue being, as you noted, heterogeneity between columns [1].
To me this still does not explain what is original or interesting about that idea. Artificial neural networks are "physiologically" the same everywhere too, and have been like that from the inception.
If you train a neural network you get some results. But presumably if you train a 1000 separate neural networks on different measurements made of the same target, and then let those thousand models vote, isn't that a new kind of learning mechanism, compared to having a single neural network that always takes in all the inputs?
Further each of the thousand neural nets could have a different learning algorithm.
Soon after you start thinking how voting can be implemented, you will quickly realize, that one of the best ways to do it is to connect each of 1000 networks to each of N outputs with some sort of weights (in trivial version - uniform), which effectively turns 1000 networks into 1 network with extra N unit layer, and no weight reuse (e.g. probably a disadvantage) in all but the last layer.
It is quite likely that the brain's relationships are not just 2- or 3-dimensional, but rather very highly multidimensionally interconnected.
When neural nets seriously hit the scene 30-odd years ago (I read a lot of the early papers, even having to mail off to get some in those pre-Internet days), no one seriously thought they were the way the brain works, just that they presented the opportunity to simulate a tiny slice of the way brains might work.
As Ted Nelson so perfectly puts it, "Everything is deeply intertwingled."
I wonder if this works for efferent behavior too. Perhaps when throwing a rock, we actually throw it a thousand times in our brain before letting go of it, as feedback from our muscles converges with our perception of ballistic forces. Like running a thousand simulations and picking the best one, as the event is happening.
I wish they would be working less like a company and would engage more with the research community on the topic. The topic of On Intelligence, Hierarchical Temporal Memory (HTM) was never open sourced, also didn’t have published math, and was thus treated by the community as weird and not actionable, and in the end ignored.
I found this particularly interesting. I have a young daughter and I find it fascinating how quickly and reliably she can learn an object and then recognise others in a way a computer just can’t do at all. Eg see very rough line drawing of an object, then recognise that animal in real life.
It’s also interesting the way I can explain new objects or animals based on one she knows and then have her recognise them easily. Eg a tiger is like a lion but has no mane and does have stripes.
This comes very naturally when teaching and learning and feels very uniquely ‘human’ compared to the rigidity of computer vision I’ve seen so far.