Vicarious AI raises another $50M(venturebeat.com)
venturebeat.com
Vicarious AI raises another $50M
https://venturebeat.com/2017/07/25/khosla-ventures-leads-50-million-investment-in-vicarious-ai-tech/
50 comments
I am always skeptical of people trying to "emulate" the human brain in machine learning. We currently do not have the tools to accurately record and analyze the dynamics of networks of neurons in the brain, and any group that claims to advance ML with knowledge of the cortex is spouting bullshit. Modern advances in ML are driven by great engineering, not biological insight.
I recall a quote from some neuroscientist. That whenever he hears people say "we know nothing about the human brain", he wants to smack them with a 900 page neuroscience textbook.
I think the biggest issue is these domains are isolated and don't talk to each other. It's not that ML researchers couldn't be inspired by neuroscience research. They just don't know any.
I talked to a researcher outside of the mainstream who was obsessed with biologically plausible models. He got good results, but not SOTA.
However his main argument was that his methods were much faster and more data efficient than standard practice. E.g. they did online learning and didn't suffer from catastrophic forgetting. Didn't require supervision and labelled data.
Standard methods are optimized towards getting the most accuracy on benchmarks, and not necessarily under realistic conditions. Real brains don't get to save huge dataset and iterate over them later. They need to learn in real time and without forgetting previously learned knowledge. Given just a stream of unlabeled data. ANNs can't do this at all. Some biologically inspired models claim to be able to do this well.
I think the biggest issue is these domains are isolated and don't talk to each other. It's not that ML researchers couldn't be inspired by neuroscience research. They just don't know any.
I talked to a researcher outside of the mainstream who was obsessed with biologically plausible models. He got good results, but not SOTA.
However his main argument was that his methods were much faster and more data efficient than standard practice. E.g. they did online learning and didn't suffer from catastrophic forgetting. Didn't require supervision and labelled data.
Standard methods are optimized towards getting the most accuracy on benchmarks, and not necessarily under realistic conditions. Real brains don't get to save huge dataset and iterate over them later. They need to learn in real time and without forgetting previously learned knowledge. Given just a stream of unlabeled data. ANNs can't do this at all. Some biologically inspired models claim to be able to do this well.
> I talked to a researcher outside of the mainstream who was obsessed with biologically plausible models. He got good results, but not SOTA.
Can you expand on this? Do you have some resources that describe (something similar to) what he did?
Can you expand on this? Do you have some resources that describe (something similar to) what he did?
It was pretty similar to Numentas stuff and hierarchical temporal memory.
>Real brains don't get to save huge dataset and iterate over them later. They need to learn in real time and without forgetting previously learned knowledge. Given just a stream of unlabeled data.
It's more than just that. Real brains need to optimize their internal data for action. The ultimate test of whether you've represented the world correctly is: but can you do stuff? Can you control an inverted pendulum (to pick a task) while constrained to have your representation updates be Lipschitz functions, with a Lipschitz constant based on your actuators' state (ie: speed, angle, force, etc)?
Note the trick here! Your representation doesn't have to be reconstructive itself (allowing for you to conditionally simulate only Lipschitz transformations), but the updates you perform on that transformation from sensory reafferant signals do need to change only at a bounded rate, because the physics of the thing you're moving actually have that property.
It's more than just that. Real brains need to optimize their internal data for action. The ultimate test of whether you've represented the world correctly is: but can you do stuff? Can you control an inverted pendulum (to pick a task) while constrained to have your representation updates be Lipschitz functions, with a Lipschitz constant based on your actuators' state (ie: speed, angle, force, etc)?
Note the trick here! Your representation doesn't have to be reconstructive itself (allowing for you to conditionally simulate only Lipschitz transformations), but the updates you perform on that transformation from sensory reafferant signals do need to change only at a bounded rate, because the physics of the thing you're moving actually have that property.
Numenta has been working on this strategy since 2005. Has anyone identified any of their products in wide use?
https://en.wikipedia.org/wiki/Jeff_Hawkins#Numenta
https://en.wikipedia.org/wiki/Jeff_Hawkins#Numenta
Even DeepMind says that future advances will come from neuroscience.
https://www.theverge.com/2017/7/19/15998610/ai-neuroscience-...
In the field of AI, our "great engineering" is not even a worthy comparison to what nature has achieved. Maybe there are a few more things to learn from it.
Also we've been studying the neocortex for a long time and have learned a lot more about it than most people realize.
https://www.theverge.com/2017/7/19/15998610/ai-neuroscience-...
In the field of AI, our "great engineering" is not even a worthy comparison to what nature has achieved. Maybe there are a few more things to learn from it.
Also we've been studying the neocortex for a long time and have learned a lot more about it than most people realize.
While Demis Hassabis and probably some DeepMind researchers use some understandings from the neuroscience literature as an inspiration for their work, I am pretty sure that a majority of DeepMind researchers would rather use a combination of mathematics and trial and error experiments to build an intuition to guide the design of their next iteration of intelligent learning systems.
The point is that when mainstream ML begins to realize that there are limitations to the current cartoonish representations of neural networks, they then go back to the biology to see what they're missing.
However there are other companies, like Numenta, which realized decades ago that the current techniques will not be sufficient for general intelligence.
Numenta is not trying to emulate the brain like the Human Brain Project, they are aiming to learn the principles behind the neocortex and replicate it in software.
Again, I don't think most people know enough about the neocortex because if they did, we probably wouldn't be so quick to discard the only real example of intelligence we have.
However there are other companies, like Numenta, which realized decades ago that the current techniques will not be sufficient for general intelligence.
Numenta is not trying to emulate the brain like the Human Brain Project, they are aiming to learn the principles behind the neocortex and replicate it in software.
Again, I don't think most people know enough about the neocortex because if they did, we probably wouldn't be so quick to discard the only real example of intelligence we have.
>there are limitations to the current cartoonish representations of neural networks
How? on a broad level Deep Learning is the same as natural neural network. Signal in and then neuron decides to fire a signal out. The algorithms inside is what differentiates a human from a machine. As long as the algorithm can make intelligent decisions who cares how the human algorithm works. We are not trying to build a human brain, we are trying to build a better than human brain
How? on a broad level Deep Learning is the same as natural neural network. Signal in and then neuron decides to fire a signal out. The algorithms inside is what differentiates a human from a machine. As long as the algorithm can make intelligent decisions who cares how the human algorithm works. We are not trying to build a human brain, we are trying to build a better than human brain
To be more factual: a majority of papers published by researchers at DeepMind do not cite any result from the neuroscience literature in their bibliography. Instead they cite other papers from the Machine Learning community.
Hey we fully mapped a worm's ~300 neurons [1], now we just need to keep scaling up until we get to the billions that the human brain is made up of.
[1] http://www.openworm.org/
[1] http://www.openworm.org/
>Hey we fully mapped a worm's ~300 neurons [1]
And functionally, what did that tell us?
And functionally, what did that tell us?
They allowed the connectome to control a lego robot [1] and it showed some signs of working around obstacles (robot backs up when it encounters a wall). Sounds simple, but now we can say that only 300 neurons are needed to be able to react to the environment in this way, which is a huge stepping stone in understanding.
[1] Video: http://www.smithsonianmag.com/smart-news/weve-put-worms-mind...
[1] Video: http://www.smithsonianmag.com/smart-news/weve-put-worms-mind...
Ok, but what functional knowledge did we gain? Not "run this neural net, and an interesting thing happens". How are those 300 neurons implementing interesting motor behavior? What behavior are they really implementing? What's the algorithm, what's its goal?
The connectome model only addresses the third and lowest levels of Marr's analysis of a cognitive/biological system.
1. Computational: What does the system aim to do? What problem does it solve?
2. Algorithmic: How does the system solve or approximately solve that problem? How does it accomplish its purpose as a part of the organism?
3. Implementation: How are cells and/or organs put together to implement that solution?
You can have a very accurate picture of (3), and still lack any solid knowledge about (2) or (1). You can also sometimes have a clear picture of (2) and (3), but overgeneralize and wind up with a near-tautological theory for (1) -- that's the accusation being thrown at certain parts of theoretical neuroscience today.
The connectome model only addresses the third and lowest levels of Marr's analysis of a cognitive/biological system.
1. Computational: What does the system aim to do? What problem does it solve?
2. Algorithmic: How does the system solve or approximately solve that problem? How does it accomplish its purpose as a part of the organism?
3. Implementation: How are cells and/or organs put together to implement that solution?
You can have a very accurate picture of (3), and still lack any solid knowledge about (2) or (1). You can also sometimes have a clear picture of (2) and (3), but overgeneralize and wind up with a near-tautological theory for (1) -- that's the accusation being thrown at certain parts of theoretical neuroscience today.
You talk like the project is over, there is still a lot to learn and we aren't going to figure out everything immediately. A simulation provides an amazing test bed for ideas though; now you can see it working, you can pause it, change things, do experiments, and so on.
Never before has an AI startup done so little with so much. That puts their total funding with debt at around $130 million. It's an utter waste. Reminds me of the $15 billion IBM has spent on Watson. The people with the money are very poor judges of technologists in this space.
Can you expand what have Vicarious done?
They've published a few research papers. Few other researchers cite them, so it's hard to argue that their research has been influential, or even significant. In 2013 they announced they had solved/broken CAPTCHA, but refused to share their code, for which they were criticized:
https://www.vicarious.com/news-detail-02.html
More recently, they published a paper on what they call Schema Networks, which are an attempt to blend deep learning with concepts. They're winning Atari games with that, but DeepMind did it years ago.
https://arxiv.org/abs/1706.04317
They're seven years old, and until recently, they'd raised $70-$80 million. I would have expected more than a few research papers in that time, if I were their investors.
https://www.vicarious.com/news-detail-02.html
More recently, they published a paper on what they call Schema Networks, which are an attempt to blend deep learning with concepts. They're winning Atari games with that, but DeepMind did it years ago.
https://arxiv.org/abs/1706.04317
They're seven years old, and until recently, they'd raised $70-$80 million. I would have expected more than a few research papers in that time, if I were their investors.
The first time I heard about DeepMind (months before it was adquired by Google) I though something similar: "winning Atari games with reinforcement learning and neural networks doesn't sound like it is worth 50 millions of pounds in investment".
So I don't want to make the same mistake again and I will read carefully the paper.
Thanks!
Thanks!
Would be great if they showed some demonstrations / proof of any kind. DeepMind did great in that respect. I understand they're working on a different problem, but it's been a number of years; usually correct ideas don't take that long to show some promising progress.
I'm honestly surprised how willing investors are to throw money at these ML startups that actually show barely any progress.
"Show barely any progress" and "show barely any progress to the public" are very different. I don't have any inside information here, but a lack of public announcements is not proof of anything.
Their 2017 paper: https://www.vicarious.com/img/icml2017-schemas.pdf
That's it? That tells me the same thing I saw when I attended a Vicarious researcher's talk at MIT CSAIL: they're basically reinventing stuff the Bayesian generative modelers in cognitive science have been doing for years, throwing some video-game playing at it, and calling it AI instead of cognitive science.
Give me $50mil and I could do a good deal more than that, just because I bother to look stuff up!
Give me $50mil and I could do a good deal more than that, just because I bother to look stuff up!
There are a few papers up on the site already and some more coming. As a business, publications aren't really the KPIs we prioritize (vs an academic-style lab like DM or OAI).
Out of curiosity, what's the KPI?
p.s. I don't think papers qualify as a "demo". I'm with you re: not running it like an academic lab and chasing a minimal publishable unit or fashions of the time. Demo = Demonstrably do something with your technology no one could do before, ie beat people at Go in DeepMinds case, in your case you could set up your own benchmark / pass a Turing test for 3 year olds or something.
p.s. I don't think papers qualify as a "demo". I'm with you re: not running it like an academic lab and chasing a minimal publishable unit or fashions of the time. Demo = Demonstrably do something with your technology no one could do before, ie beat people at Go in DeepMinds case, in your case you could set up your own benchmark / pass a Turing test for 3 year olds or something.
KPI = amount of funds raised.
The economics for some of these labs is also pretty different from what you'd normally expect. There are numerous semi-private organizations in the defense space ( such as Boston Dynamics ) and similarly in the pharmaceutical space. Which may only produce commercially viable technology after a decade or more. For many of these entities, the funding dynamics are dependent on the overall scope of the problem along with the quality of the team rather than any direct foreseeable economic outcome.
I understand, I don't mean commercial progress either.
Interesting the bird analogy again - “airplanes don’t flap their wings” was in a recent article by Jeff Hawkins describing Numenta's approach (http://spectrum.ieee.org/computing/software/what-intelligent...). Vicarious co-founder Dileep George was also a co-founder at Numenta. Perhaps they have similar philosophies to approach but Numenta modeling the Neo-cortex while Vicarious modeling the entire brain?
Back in the Dileep George days Numenta sought mathematical support for their theory in a Bayesian approach à la Judea Pearl. But I don't think they ever had any illusions that the brain explicitly performs statistical calculations. Hawkins et al. at Numenta are now embracing the theory of sparse distributed representation, developing upon the work of Pentti Kanerva as a formal foundation. It's biologically much more plausible, but technologically less feasible (with the current hardware paradigm). I suspect Dileep George in the end was more eager to put the theory into use.
They agree on the bird analogy, except Numenta (IMHO rightfully) thinks you can't come up with airplane designs without having understood how birds actually fly.
Except, airplanes kinda do flap their "wings" - each motor blade is a wing. The motors then turn those blades - i.e. flap the blade wings, for the same purpose - push air and generate lift. Airplanes also have an additional set of wings for gliding.
Birds only have to carry themselves though, so it's an unfair comparison.
No, quite different approaches and goals.
Airplanes don't flap their their wings, but they are also orders of magnitude less energy efficient than birds.
If you want computational power on the scale of the brain, power consumption is a real concern.
If you want computational power on the scale of the brain, power consumption is a real concern.
Are planes less energy efficient than birds? Sure they consume more energy, but they can lift many orders of magnitude more weight. To carry that much weight, that much distance, you would need an unfathomable number of birds. Flying for a very long time.
The same with with brains. As far as I can tell, transistors are orders of magnitude more energy efficient than synapses. Synapses use tons of slow chemical reactions to send a signal. Transistors just send a few electrons near light speed.
The same with with brains. As far as I can tell, transistors are orders of magnitude more energy efficient than synapses. Synapses use tons of slow chemical reactions to send a signal. Transistors just send a few electrons near light speed.
Birds are more efficient in terms of energy expended to move the same mass the same distance as a plane.
Transistors are not more efficient than synapses. Neurons and transistors in subthreshold transport charge in the same way, through diffusion. Neurons just have a more efficient structure leading to less energy expenditure for an equivalent amount of information processing as the transistor.
If you operate your transistor above threshold as done in all digital circuits, you are orders of magnitude less efficient.
Transistors are not more efficient than synapses. Neurons and transistors in subthreshold transport charge in the same way, through diffusion. Neurons just have a more efficient structure leading to less energy expenditure for an equivalent amount of information processing as the transistor.
If you operate your transistor above threshold as done in all digital circuits, you are orders of magnitude less efficient.
Planes also travel 500-600 MPH while birds fly, maybe 30-50 MPH? I don't know what point you are trying to make with that comment. If it was more energy efficient for planes to flap their wings, they would be doing it right now.
I think you are conflating utility for humans and cost efficiency with energy efficiency.
There is no question that birds are more energy efficient than planes in terms of energy expended to move the same mass the same distance.
However, humans would prefer to spend large amounts of energy by burning fuel to cross the Atlantic in 6 hours. It is much more COST efficient, but that is because energy is cheap.
There is no question that birds are more energy efficient than planes in terms of energy expended to move the same mass the same distance.
However, humans would prefer to spend large amounts of energy by burning fuel to cross the Atlantic in 6 hours. It is much more COST efficient, but that is because energy is cheap.
Yet, no one tries to travel across the continent in an armchair tied to a flock of those energy efficient birds.
Good job on the hyperbole. They don't because fast airplanes have more utility to us as humans.
You of course have to consider what you are measuring. Birds have a very low speed, but they are more energy efficient in terms of mass times distance travelled per energy.
You of course have to consider what you are measuring. Birds have a very low speed, but they are more energy efficient in terms of mass times distance travelled per energy.
A rubber band driven toy plane wound by a child can fly for minutes. I'm not sure birds expend orders of magnitude less energy than that.
Did somebody say "Rocket AI"?
OpenAI & DeepMind > Vicarious