A Deep Learning Dissenter Thinks He Has a More Powerful AI Approach
technologyreview.com
technologyreview.com
"Marcus has a point - even is some of what is said about deep neural networks is incorrect (for instance, they can learn and generalize from very few example, one shot learning).
However, he got it wrong with the answer. The key for machines to reach the symbolic abstraction level is the way we train them. All training algorithms, supervised, unsupervised or reinforcement learning with LSTM rely on the assumption that there is an "utility function" imposed by some external entity. Problem is, by doing so, we are taking away their capacity [of the machines] to make questions and create meaning.
The most important algorithm for learning is "meaning maximization" not utility maximization. The hard part is that we cannot define what is meaning - maybe we can't, I'm not sure. That is something I will be glad to discuss."
This works in theory, but in practice you only have a limited amount of chances to try something out (because of the arrow of time). This makes learning a necessity. You need to keep a record of all trials you have performed so that you can reuse this information later when the same situation reoccurs. How to do this in an optimal way is described by Bayes' theorem.
The key to AI will be a certain set of priors, biases and fixed function units that make this computationally tractable; we'll likely need things like invariance of the learned information to various changes so that it can be reused in different settings, segmentation of data coming from the world into episodes (hippocampus), attention, control (basal ganglia), mental rotation (cortex) and path integration (hippocampus, grid cells).
However, I would argue the prior of Bayesian modeling can be just as nebulous and computationally intractable as an optimization objective. Like supervised learning, Bayesian modeling is just a tool.
I'm skeptical that we will reach AI through a deep understanding or modeling of the brain. Technology and computer science advances more quickly than the biological sciences, at least in recent times. You might argue a success in robotics like [0] is a motor control system. But they built this extending mathematical frameworks not being biologically inspired, and the big wins there didn't come from fixating on a learning framework or biological mimicry; just like humans learning to fly didn't come about from flapping wings like a bird. At some point we hacked an engine (invented for other purposes) onto a wing and came up with powered flight.
As an aside, only seeing input a limited number of times would likely improve your ability to find models that generalize as your model must be able to take these one off learnings and unify them in some way to achieve high training performance. With respect to human learning, a specific individual only has one chance, but nature has had many. We are only a selection of those chances that seemed to work well enough. There are many commonalities to existence that allow for this to work well in practice.
[0] http://groups.csail.mit.edu/rrg/papers/icra12_aggressive_fli...
A supervised machine learning framework can't tell the researcher what training instances it needs to see in order to improve its meaning. A supervised learning framework can't imagine where it might find that training instance, or describe what it may look like. A supervised learning framework never asks to go to the zoo.
I think you've correctly identified the million dollar question in AI, and that's why I believe the major breakthroughs in artificial general intelligence won't come from scientific researchers. They might even come from PG's dreaded literary theory types, ones who are also talented programmers (...Lit theory being maybe the only field that thinks about structuralism, deconstruction, how meaning is manufactured, etc. - things outright dismissed by the scientific community as useless, but which still hold potential to produce a 'meaning algorithm').
It's not just training that implies utility maximization, basic internal processes of a neural network, such as back propagation, are based on the utility function maximization. Back propagation is an approach that allows the network to do gradient descent on a utility function by propagating errors through the system.[1] Back propagation has been the way that deep networks have become tunable and effective. Without a utility function, it's hard to see how one would tune them.
And it's hard to think of what "mean maximization" could be but maximizing a utility function based on meaning.
And while you can construct a neural network that learns from a single example, the nn framework is fundamentally based on learning from multiple examples so such a construct is mostly meaningless.
But the point about machine learning being limited by the training methodology is good, it's just I'm pretty sure you'd different algorithms if you used a different methodology, the existing algorithms are used specifically because they fit the methodology.
It's just a simplification. We are consuming the algorithm's output, rather than it being for internal consumption in a larger mind. So we supply the goal, where in a larger more natural mind the goal comes from other parts of the mind - still computing ultimately against a fixed utility function. That function being our human instinctual basis and the "firmware" of culture it's primed to load.
Mostly because Marcus isn't willing to talk about what he's working on right now. Hand-wavy stuff about how children learn, a tired analogy that's been trotted out many times before. This article was too early, and it doesn't contain real news. Reproducible results, or it didn't happen...
Fwiw, the statement below is simply untrue.
> “If you want to get a robot to learn to walk, or an autonomous vehicle to learn to drive, you can’t present it with a data set of a million examples of it falling over and breaking or having accidents—that just doesn’t work.”
The entire premise of Deepmind's work combining DL and reinforcement learning shows that autonomous agents learn from millions of examples.
And in fact, data is not the limiting parameter here. Trying to create an AI that doesn't need big data will be great for edge computing -- but we're actually swimming in data and using it to set new records every year -- http://deeplearning4j.org/accuracy.html -- so we might as well rely on it. Ignoring big data in the pursuit of AI is like ignoring sunlight in the pursuit of clean energy.
Deep learning is one of the best ways we have to use that data. In a sense, deep is the new normal.
I think this is a "looking under the lamppost for your keys because that's where the light is" situation. We hear about tons-of-data and deep learning because teams like Google/FB have problems that fit those situations well. This is not representative of the vast sea of learning applications most organizations face.
I work on learning applications day-in-day-out. In a typical year I work with over 50 different organizations/projects. Almost none of them have data anywhere near what deep learning requires. Here are just a few examples from the past year:
1) optimizing recruiting pipeline. Maybe you have ~1000s of data points per year
2) medical billing applications: typically hundreds of data points per year
3) novel but slow-moving financial instruments, maybe 1 data point per day
4) automated sensor calibration where each run of the hardware costs you a few hundred dollars: depends on your budget, but thousands of data points in a year are representative
I think Google/FB and the "web+mobile is everything hivemind" (I'm not saying that applies to you) have deeply distorted people's expectations for how much data is available to solve a typical problem.
Ignoring big data in the pursuit of AI is like ignoring sunlight in the pursuit of clean energy. It's everywhere. Why not use it?
My argument in response would be we are swimming in a lot of things
Data and sunlight are only two of them
This tendency leads to both helpful and dangerous heuristics and biases. From these biases humans build false beliefs that are damaging to themselves, their communities, and in the long run the species as a whole. If AI is about enabling humans to do more and better, should we not accept the failings of the current technologies in favor of assuring that they do not fall prey to the same biases and heuristics that lead humans to slaughter each other over the religious dogma of the past, destroy the environment with impunity, and accept ideas that 'feel' correct over being correct?
I don't know that anybody seriously proposes that deep learning is the be-all end-all of AI techniques. It's VERY powerful for a lot of things, but I think DL researchers are aware of things DL doesn't do / isn't good at. Look at the recent book The Master Algorithm which breaks down a lot of what it would take to create a truly general purpose learning algorithm: If you believe the author's thesis, Deep Learning (or something like that) is just one piece of a much larger picture.
And without trying to start a debate over the merits of ML versus "GOFAI" or symbolic computation, etc., I think it's fair to say that DL doesn't really add anything in terms of reasoning. It's great at saying "this picture has a cat in it" or "this wav file says 'Hello, my name is mindcrime'", but that's a pretty small part of what human intelligence can do.
It's hard to say. It's not a deeply technical book, I will say that. The main value I found in it, is that he covers a broad base of different techniques and then lays out some ideas on how they could all be combined to make a "general purpose learner". I found it worth reading, but YMMV.
He also pushes the same idea over and over again that there must be a master algorithm, when I got the feeling very early on that it wasn't a necessary thing or a practical concept to guide research. It's like arguing that you have to choose between electricity and magnetism, when they might just be aspects of the same underlying forces. Also, his reasons for claiming that certain algorithms are not the master algorithm are pretty weak and based on current progress, not theoretical limitations.
How does he know? He only sees the output and not how the brain works or how much information has been processed before hand. So it seems like the brain took an easy shortcut and generated a new rule, and you could just write a bit of code to do that as well.
But isn't this rule learning just a few more layers in a learning network. Basically a learning network that operates on concepts not on just raw pixels.
One thing I noticed with my kids is they have a ridiculously good memory. For a while they almost fooled me that they could read. They would remember a word in a book and then a month later point to it and pronounce the word. So I thought, oh wow, you could read that. But then realized they just remembered it. It seems a lot of how a brain learns is just be absorbing and processing massive amounts of data, not unlike a learning project Google or these other companies have.
Now on the subject at hand, haven't there been expert systems that do what he proposes. In a way IBM's Watson is a successor to many of those expert systems. So one can say this has already been implemented and tried, it works in some instances but not others.
So even though the idea is fairly simple, I think anyone who thinks about AI for any trivial amount of time will come up with the idea that AI's somehow have to automatically generate imperfect generalizations, the final result might be a truly innovative product.
Actually building an AI that does this on a scale that results in useful applications is something that probably has not been practical for a long time, and might or might not be now.
Maybe from the point of view of the investors themselves, in most cases, if they were expecting a big, immediate financial return. But the technology has been improving steadily (ish) throughout the years. As AI people sometimes say "AI is whatever computers can't to $today". That is, as soon as AI reaches a point where it can do something useful, people stop considering it AI.
I saw an example of this yesterday while having a conversation with somebody much younger than me (I'm 42) when we got on this topic. I was saying that things like the voice recognition in that Google Now, Siri, Cortana (sp?), etc. use would be good examples of what would have been mind-blowing AI capabilities 20 or 30 years ago and he said "Oh, that's AI?"
No, it's not AI now, because now it's common-place. :-)
Of course, none of that is to say that we aren't still a long way away from a true "AGI". But AI research has been yielding useful techniques for quite a long time, even if we didn't get "there yet" for AGI.
In particular, this reminds me of some of the operational logic work done in OpenCog (http://opencog.org/) and, especially, Pei Wang's Non-Axiomatic Reasoning (http://cis-linux1.temple.edu/~pwang/papers.html). I've liked Non-Axiomatic Reasoning for a long time, and this sounds in the same broad area.
Pei Wang, incidentally, came up with the best operational definition of intelligence I've seen (and one I apply to other systems and approaches): Intelligence is the ability to act appropriately with limited knowledge and limited resources (including time and space).
For what it's worth, my (published, although very open to change) position is that a layered architecture (akin to old-school subsumption architectures) is probably going to be most effective here. Combining the 'symbol creation' abilities of a deep neural network with the evidential reasoning, generalisation, and planning capabilities of a cognitive layer will allow us to get the best of both worlds.
I have been going back and spending some time with the old "blackboard architecture" idea lately. I harbor a suspicion that that, or something like that, will turn out to be a useful way to integrate the capabilities of various different elements of cognition.
The definition from Wang makes sense. How does AIG differ from other methods? For example, I noticed this article on Bayesian Program Learning a few days ago which IMHO takes a layered approach and is based on human learning models. [https://news.ycombinator.com/item?id=10723485] Interestingly there were no comments on it.
Could you post refs to some the relevant publications your published opinions? I'm curious about your approach!
- It involves generative models (needed to make inferences from very few examples)
- It is still a connectionist approach (typical probabilistic programming is great if you have a lot of insight into the model, but not if you're trying to solve a general case... unless you're doing program induction but you need to represent programs...)
- It doesn't involve MCMC sampling for inference, because that's too slow or even intractable.
Some type of variational program induction where programs are represented as differentiable neural networks would be in that corner. Or it could be something totally different, but speculating is fun.
I've been trying (but failing) to do so by clustering similar generative rules of different complexity together then getting the algorithm to search the grammar space in a way that it recursively tries simpler models first then goes on to try generating (and learning) more complex rules that are known to have similar output to the best simpler ones.
My intuition is that you have to cluster your generative rules into an taxonomy that your algorithm can navigate from the top. It's exponentially inefficient to try to recognize "dog" until you have recognized that the simpler "animal" is a good approximation.
I also think that in the ultimate solution, the output leaf statements of the grammar will be parametrized like a normal programming language, so that for example, the algorithm can generate a color, a radius and then generate 100 circles referencing this single color and radius to represent repeated patterns without having to learn their size and color individually when they are clearly homogeneous or invariant across a bunch. The bayesian, occam's razor, solution to a bunch of similar things that you don't have a category for yet is shared parameters. These parameters enable learning with very few examples. The algorithm doesn't have to learn a full new category to make good predictions, it can simply notice a functional parametric pattern in a part of a scene and extrapolate immediately.
I haven't cracked it however. It's grueling to debug probabilistic algorithms compared to clear cut binary ones!
It's like a probabilistic binary search in the model taxonomy.
One thing that makes me hopeful that the learning process is possible is that the non-terminal grammar rules are generic and can always become anything. You don't tend to get stuck in local maxima even if you only search a small part of the program space when any node can always be swapped for a node of any another category.
My problem right now is that my MCMH doesn't mix well. Even though I tried to shorten the jumps between models of different complexity for a particular thing, once the vague model traces are well fitted to the training examples, there is too deep a score chasm to jump to the next level of complexity.
My simulations seem to get stuck in simpler modelisations of things. I want simpler models to recursively open the door to learning and recognizing slightly more complex models just under them in the hierarchy but this doesn't seem to be happening. The hierarchy is not really even forming properly.
I might be missing something stupid. I don't consider myself an expert in these things. I just dabble for fun. It's one of the reason I'm discussing it here. Maybe someone more knowledgeable will make me have an aha! moment.
There is still a bunch of things I could try but like I said, it's long and grueling debugging work.
Hence, the DNN's modelling a small part of the brain that doesn't even do our thinking isn't going to make thinking machines. The machines will be really stupid and hard to train. Humans are hard enough to train: takes two decades on average. The ability to automate abstraction, understanding, and feedback via example generation all on small amounts of data is critical.
If that's not found, then my next question is, "Will DNN's deliver acceptable models for non-recognition tasks they're aiming at or will they experience The Second AI Winter when truth sets in?"
[Brain inspired] models are arguably closer to how the brain operates than artificial neural networks. "But they, too, are oversimplified," he says. "And so far I have not seen a knock-down argument that they yield better performance in any major challenge area."
There are a few of these brain inspired machine learning companies, yet I cant think of a single one acquired by Google.
Probabilistic modeling is great when you dont have much data and want to inject prior knowledge in a cohesive way. However Google's approach appears to be to create large and robust training sets to feed into somewhat conventional supervised learning frameworks.
[0] http://www.technologyreview.com/news/536326/ibm-tests-mobile...
Cofounder Gary Marcus has a publication list at his academic webiste:
http://www.psych.nyu.edu/gary/marcus_pubs.html
Cofounder Zoubin Ghahramani has his research available at his academic website:
http://www.automaticstatistician.com/about/#
"The current version of the Automatic Statistician is a system which explores an open-ended space of possible statistical models to discover a good explanation of the data, and then produces a detailed report with figures and natural-language text. While at Cambridge, James Lloyd, David Duvenaud and Zoubin Ghahramani, in collaboration with Roger Grosse and Joshua Tenenbaum at MIT, developed an early version of this system which not only automatically produces a 10-15 page report describing patterns discovered in data, but returns a statistical model with state-of-the-art extrapolation performance evaluated over real time series data sets from various domains. The system is based on reasoning over an open-ended language of nonparametric models using Bayesian inference."
"How humans do it" isn't necessarily the only way to solve a problem. We don't know how we add 2 numbers in our minds; but the computer can add 2 numbers orders of magnitude faster than us.
We actually have a pretty good idea of the mechanisms for basic arithmetic. We do things about as inefficiently as possible.
To put it one way, what is the _rate_ of learning? How many _concepts_ do children learn per unit time?
Concepts: Our computer systems take whole images and train on them, but before a pixel-based image gets to the learning part of a child's brain, it has probably been conceptualized in some way already.
Rates: And then, from this feature vector—maybe 1 feature vector per 10 seconds, for a child—how often is it incorporated in a learned concept? Maybe one concept per hour? Per day? Think about how challenging it is for children to learn vocabulary. A young child doesn't know that much.
The important thing is that a child is a self-learning machine. It chooses what's important and has a way to shift its attention, a critical enhancement of new (2015) AI approaches, deep learning and otherwise. It can explore its own world and somehow choose the "training instances" that matter.
By comparison, a neural network training for days looks at millions of images. A very high rate. But the most effective approaches (the massively-multilayer networks, the deep in deep learning) are making concrete what a child's brain must do already: separate perceptual components from learning components, in a conclusively more structured and hierarchical way.
In fact I was working with Dr. Frank Guerin [1] back in 2008 about this exact approach. He wrote an interesting paper that approached ML from a pedagogy perspective titled A Piagetian Model of Early Sensorimotor Development
[1] http://homepages.abdn.ac.uk/f.guerin/pages/ [2] http://homepages.abdn.ac.uk/f.guerin/pages/EpiRob2008.pdf
Does this actually happen, or is it baseless paranoia?
He spoke about how he was determined to prove that non-deep learning methods could perform as well as deep learning at tasks like ImageNet. The interview was in March 2015, and AFAIK they haven't published anything yet.
OTOH, his work on the Automatic Statistician sounded very interested.
[1] http://www.thetalkingmachines.com/blog/2015/3/26/3mixrq61fb0...
"The idea for a search engine that maps associations came to Franks by way of his three young children. He noticed how each child processed information by taking two pieces of knowledge, combining them, and coming up with something new. Franks wondered whether he could get a computer to do the same thing" From:
A Search Engine that Thinks, Almost http://newscenter.lbl.gov/2005/03/31/a-search-engine-that-th...
With wearable, if you allow the eyes of machines to read the world with you, hear your conversation, with enough time, your machine should be capable of learning enough to become like you. With millions of people doing so, you essentially create an AI that is like human able to answer complex questions.
"A deep-learning system can be trained to recognize particular species of birds, but it would need millions of sample images and wouldn’t know anything about why a bird isn’t able to fly."
Deep-learning ≠ only image recognition
Of course if you train it only on images it will only know how a bird looks like, but if you train it on the properties of the birds and their ability to fly, then it will learn why a bird isn't able to fly.
Shut up and show us the code!