Advances in Machine Intelligence

[AaronK]

I am an undergraduate student pursuing computer science in the Southwestern United States (I just switched my major to comp. sci actually). Recently, I came across an individual who claimed to research AI professionally and who expressed to me the following view after some discussion of various technologies and the rate of advancement of such tech (this took place on a separate online forum). Here is what he said:

"I can sympathize with your point of view, but from the point of view of an actual AI researcher like myself, you have it exactly backwards. In reality, all the amazing new stuff like self-driving cars and Go AIs are things that are horribly old hat. The machine learning techniques they are based on date from the 1980s, for Turing's sake. While it may seem to the layperson that these technologies emerged from nowhere, to people in the field they have been long expected and in fact have been disappointingly slow.

It was really a big company with a lot of money like Google throwing real money behind the field that has allowed it to advance so quickly in the public eye, but from a theoretical perspective this isn't anything new. Only a company like Google has the budget to put together all the GPUs and CPUs AlphaGo is composed of and pay programmers familiar with Go to work on it for years just for a PR stunt -- but in reality the methods for AlphaGo are decades old and could have been done long ago. Same deal with all the resources needed for the self-driving car. So from my perspective, the exact opposite has been happening: slower and slower scientific progress, punctuated occasionally by amazing engineering stunts from big companies with a lot of money."

After reading this individual's response, I have to admit I am doubtful. Would you guys say there is any significant veracity to this person's view? Beyond the fact that apparently these "old" machine learning techniques were pioneered in the 80's, would you say this individual writes with accuracy?

Would you necessarily have a definitive response to such an individual, for or against this view?

I would very much love to read what anyone has to say regarding this, and would also greatly appreciate where you think machine intelligence will be in the next 10 year span (speculation is absolutely acceptable). I hope to resolve my thinking on this matter.

 

[Simon Bridge]

An opinion is an opinion: what makes you think he is not being sincere?

I suspect that scientific progress has always seemed like that to people in the field: that "current" progress is slow and incremental compared with progress in the past. It's a bit like how the "end of days" prophesies always seem to be just about to come true.
What evidence has been offered by this person to show that the observation made is special to "these days"?

Can you provide a link to the original discussion?

 

[256bits]

I hope to resolve my thinking on this matter.

How old is certain technology?
Sorting thoughts and techniques date back ages way before the 1980's. The personal computer only made it aware to the interested general public.
http://www.computerscijournal.org/pdf/vol7no3/vol7no3_369-376.pdf
Look at where the common Quick sort, or Merge sort had their debut.

The Turing test dates back to the 1950's.
https://en.wikipedia.org/wiki/Turing_test
( Descartes wrote about AI in his time, in regards to whether a machine can be made to think or just imitate. IIs that the middle 1600's)

One quibble I have with what the person wrote is that he says what he knows, all happened in the 1980's. To me that is well beyond belief.
Technology builds upon what is there before. It can progress in leaps and bounds, or crawl at a snail's pace awaiting the next major breakthrough which may never happen.

 

[AaronK]

An opinion is an opinion: what makes you think he is not being sincere?

I suspect that scientific progress has always seemed like that to people in the field: that "current" progress is slow and incremental compared with progress in the past. It's a bit like how the "end of days" prophesies always seem to be just about to come true.
What evidence has been offered by this person to show that the observation made is special to "these days"?

Can you provide a link to the original discussion?

Sure. Here is the link: https://forums.spacebattles.com/threads/automated-trucks-will-only-cost-1-8-m-jobs.410816/

It's a thread discussing the potential for the loss of truck driving jobs to autonomous vehicles, referencing a Vox article. The discussion starts around page 3 of the thread I think. The forum name is "Spacebattles.com".

 

[QuantumQuest]

I am an undergraduate student pursuing computer science in the Southwestern United States (I just switched my major to comp. sci actually). Recently, I came across an individual who claimed to research AI professionally and who expressed to me the following view after some discussion of various technologies and the rate of advancement of such tech (this took place on a separate online forum). Here is what he said:

"I can sympathize with your point of view, but from the point of view of an actual AI researcher like myself, you have it exactly backwards. In reality, all the amazing new stuff like self-driving cars and Go AIs are things that are horribly old hat. The machine learning techniques they are based on date from the 1980s, for Turing's sake. While it may seem to the layperson that these technologies emerged from nowhere, to people in the field they have been long expected and in fact have been disappointingly slow.

It was really a big company with a lot of money like Google throwing real money behind the field that has allowed it to advance so quickly in the public eye, but from a theoretical perspective this isn't anything new. Only a company like Google has the budget to put together all the GPUs and CPUs AlphaGo is composed of and pay programmers familiar with Go to work on it for years just for a PR stunt -- but in reality the methods for AlphaGo are decades old and could have been done long ago. Same deal with all the resources needed for the self-driving car. So from my perspective, the exact opposite has been happening: slower and slower scientific progress, punctuated occasionally by amazing engineering stunts from big companies with a lot of money."

After reading this individual's response, I have to admit I am doubtful. Would you guys say there is any significant veracity to this person's view? Beyond the fact that apparently these "old" machine learning techniques were pioneered in the 80's, would you say this individual writes with accuracy?

Would you necessarily have a definitive response to such an individual, for or against this view?

I would very much love to read what anyone has to say regarding this, and would also greatly appreciate where you think machine intelligence will be in the next 10 year span (speculation is absolutely acceptable). I hope to resolve my thinking on this matter.

I can see nothing more but stating the obvious (and of course true), here. It is really the case that a layperson, cannot see the stages of advancements in detail and maybe sometimes cannot see these stages at all, thinking that something came out of nothing. But it is not his/her job to do that. A scientist in whichever field and in this particular case in AI, has to know a lot of the details and nuances going on. But in my opinion, the important thing is to try to see why a big company like Google, as it is referred in the OP, invested on something the way it did and when it did it. And this inevitably leads to think about the advancements in IT and telecommunications industry, especially in the past 15 years or so, that is going really by leaps and bounds. The first thing is that computing machines became a commodity. Their operation became a very easy thing. New materials and scientific progress, led to very cheap and small hardware. Software development became an almost routine process. Great speeds on the net and tons of data, became accessible to everyone. The IT market became huge and with great opportunities, for individuals and companies alike. It was inevitable that tons of data and information will gather and the timing and well grounds for investing on technologies and commodities on this, had come. So, a scientific idea that had been kept at its infancy for many years or not particularly developed anyway, had its opportunity to make it to the market. And the rules are made by the market. Good investments are (reasonably enough) aiming at great revenues. Had it not be the case that some fundamental things became widely spread, the landscape would be totally different.

So, it is very reasonable that big companies made and make the investments they do, the way and with the timing they do. On the other hand, it is equally reasonable that "the time comes" for certain ideas and technologies from the past, to make it into the IT market. And as a whole, taken in a statistical manner, their rate of development is slow, but that dictates the market. Funding and investments by big commercial companies, cannot be made on things that won't create revenue. Of course, at national level there is good funding for many scientific endeavors in some countries, but this must be somewhat picky too, as funding is by and large given through taxes.

Although some predictions can be safely enough made for the foreseen future, there is a multitude of factors, that can influence the whole thing, so we can potentially see things in the future that we cannot foresee now. But again, in my opinion, the market will essentially make the rules. This thing although healthy, it is not always a flawless process or the best it could be.

 

[FactChecker]

My two cents:
I think he is overstating the capabilities of AI in the 1980's. The fundamental ideas of the 80s are still valid, but I think that research in neural networks, pattern recognition, distributed control, etc. was not very advanced. But it is hard to judge because, in my opinion, everyone and his brother was jumping on board and overselling their work. Then their results had to run on 1980's computers and were not very impressive. So much of AI depends on the efficiency of the algorithms. It's hard to separate improvements in algorithms from the massive increases in computer power. I don't know if they could have even considered some of the approaches that are realistic today.

PS. I should add that I am very impressed with what I am seeing and hearing now regarding the self-driving cars.

 

[DiracPool]

Beyond the fact that apparently these "old" machine learning techniques were pioneered in the 80's, would you say this individual writes with accuracy?

I'd say it's pretty accurate. I've been around for the whole thing although I didn't get into AI-neural networks (NN) until the 90's. Of course, just as with the Alpha Go thing, there were always new revolutionary advances in NN technology coming along all the time. There was backpropagation, simulating annealing, recursive networks, "Fuzzy" logic, holographic memory in crystals, chaotic K-sets, etc, etc. I actually discussed this in another thread and the point is that, after a while, one tends to become disillusioned in these things, but you keep trudging along anyway.

You know a revolution has come about when writers in the field start talking about the old technology in a certain way. It typically begins with, "In the old days researchers thought that things worked like this and that...but now we know that things work like that and this...etc." We haven't seen that with NN technology. When there's nothing new under the sun writers talk about "significant advances" which are, at best, "evolutionary" steps that really don't further the field much at all in the long run, although they may generate a lot of hype in the short run. I think that's what we're seeing here. For example, right now one of the big hypes is "deep learning" which may prove to go somewhere, but from what I can garner it's still fundamentally based on the tired-old backpropagation technique we have been using since the 80's..

 

[Aufbauwerk 2045]

For a look into at least one possible AI future, I recommend Ray Kurzweil's books.

 

[onoturtle]

I'm not familiar with the self-driving car applications, but AlphaGo is a combination of deep learning and reinforcement learning. I can't say much about the history of reinforcement learning, but deep learning has been around for a long time. It's artificial neural network research rebranded.

Quoting from the intro chapter of Deep Learning by prominent DL researchers Goodfellow, Bengio and Courville (free draft available here http://www.deeplearningbook.org/)

At this point in time, deep networks were generally believed to be very difficult
to train. We now know that algorithms that have existed since the 1980s work
quite well, but this was not apparent circa 2006. The issue is perhaps simply that
these algorithms were too computationally costly to allow much experimentation
with the hardware available at the time.

I believe this is what the person the OP talked to was referring to. It isn't AI theory that has made progress over the decades. It is computer hardware that has. It is only with today's hardware that we realize the ideas from the 80's are were actually viable. That and "big data" world we live in today is starting to give us enough sample size to build larger and larger networks. So it feels like the AI researchers in the 80's made huge strides but then had to take a break for a few decades for the rest of the world to catch up before it became practical. The practical neural nets these days are the same feedforward networks with parameters learned with the classic backprop. The biggest difference in recent years is that the preferred hidden unit activation function is now the rectifier rather than tanh, solving the vanishing and exploding gradient issues that can occur with backprop.

Though I don't want to discount the AI research in between the 80's wave and today's wave of deep learning research. The statistical learning theory and graphical modeling stuff created in the interim is very much an important part of a data scientist's toolbox and is used in a lot of applied research. AI research just goes in waves of fads. In 10 years, who knows, what the next buzzword will be. Maybe kernel methods will get rebranded as something else and be hip again haha

 

[AaronK]

I'm not familiar with the self-driving car applications, but AlphaGo is a combination of deep learning and reinforcement learning. I can't say much about the history of reinforcement learning, but deep learning has been around for a long time. It's artificial neural network research rebranded.

Quoting from the intro chapter of Deep Learning by prominent DL researchers Goodfellow, Bengio and Courville (free draft available here http://www.deeplearningbook.org/)
I believe this is what the person the OP talked to was referring to. It isn't AI theory that has made progress over the decades. It is computer hardware that has. It is only with today's hardware that we realize the ideas from the 80's are were actually viable. That and "big data" world we live in today is starting to give us enough sample size to build larger and larger networks. So it feels like the AI researchers in the 80's made huge strides but then had to take a break for a few decades for the rest of the world to catch up before it became practical. The practical neural nets these days are the same feedforward networks with parameters learned with the classic backprop. The biggest difference in recent years is that the preferred hidden unit activation function is now the rectifier rather than tanh, solving the vanishing and exploding gradient issues that can occur with backprop.

Though I don't want to discount the AI research in between the 80's wave and today's wave of deep learning research. The statistical learning theory and graphical modeling stuff created in the interim is very much an important part of a data scientist's toolbox and is used in a lot of applied research. AI research just goes in waves of fads. In 10 years, who knows, what the next buzzword will be. Maybe kernel methods will get rebranded as something else and be hip again haha

I would agree with this estimation of things. Since I posted this thread back in August of 2016, I've come a long way in understanding the history of AI and machine learning work through my own independent study. For technical studies, I bought the physical copy of the deep learning book by Goodfellow et al. (as well as the linear algebra book by Shilov and the probability theory text by Jaynes), and have been working through it alongside my python book.

I would say, however, that AlphaGo's success is particularly amazing and that the continuing work Deepmind is doing with that specific system is important. After they improved it, I think they let it play against the world's currently number 1 ranked player in a 3-game match and it managed to win all three (though I think I remember it nearly lost the first game? It took place very recently so I have to look into it more). Regardless, I wonder how long it will be before a general algorithm or set of algorithms will be developed. From what I know now, it seems hopelessly difficult (technically) but somehow still fairly close (like within the next 50+ years or so)

 

[stoomart]

I would say, however, that AlphaGo's success is particularly amazing and that the continuing work Deepmind is doing with that specific system is important. After they improved it, I think they let it play against the world's currently number 1 ranked player in a 3-game match and it managed to win all three (though I think I remember it nearly lost the first game? It took place very recently so I have to look into it more). Regardless, I wonder how long it will be before a general algorithm or set of algorithms will be developed. From what I know now, it seems hopelessly difficult (technically) but somehow still fairly close (like within the next 50+ years or so)

I've worked with supervised machine learning in IT security for proprietary document matching and event prediction, but it's always cumbersome to perform the initial training and tuning for each type of data; I still prefer frequency and standard deviation techniques for anomaly detection. Like you said, what they're doing here with unsupervised learning still seems a ways out before I could give it all my data and have it figure out what's important, but it's still super cool.

 

[Greg Bernhardt]

Just Google's image recognition is something to behold. The other day I was wondering what a type of plant was in my yard. I took a photo of it on m phone, uploaded it to Google's image recognition app and it told me what plant it was. The plant was a "Brunnera macrophylla 'Jack Frost'". That capability is stunning to me and I doubt possible in the 80s.

 

[Nugatory]

[quoting a third-party source] "While it may seem to the layperson that these technologies emerged from nowhere, to people in the field they have been long expected and in fact have been disappointingly slow."

There are two intertwined trends here. On the one hand we have the algorithmic and theoretical development of AI, and on the other we have the cost and capability of the computing hardware. For decades, the first trend was running ahead of the second; for example, people knew how to go about building a program that could in principle play grandmaster-level chess long before it was reasonable to actually do it.

Thus, as the capabilities of the computing platforms advance (and it's not just Moore's Law getting more out of each piece of silicon, but also much more effective distributed and parallel processing bringing many pieces of silicon together) more and more of the things that we "always knew how to do" suddenly start happening. That's how we can see huge advances out of seemingly nowhere, even while a specialist in the field can feel that all that's happening is consolidation and confirmation of old stuff.

Eventually however, problems that resist solution by the currently known techniques will start to appear; likely we're already aware of them but haven't yet recognized that they pose theoretical challenges that cannot be overcome by brute strength. When this happens the pendulum will swing back the other way. I find it interesting that the best computer programs for chess, bridge, and go use very different approaches; it seems unlikely that we've discovered all the promising approaches to machine problem solving.

 

[atyy]

The machine learning techniques they are based on date from the 1980s, for Turing's sake. While it may seem to the layperson that these technologies emerged from nowhere, to people in the field they have been long expected and in fact have been disappointingly slow.

It was really a big company with a lot of money like Google throwing real money behind the field that has allowed it to advance so quickly in the public eye, but from a theoretical perspective this isn't anything new.

Yes, most of the technology is from the 1980s. However, it required the vision and perseverance of those who understood its promise despite the "disappointingly slow" progress to stick it out and show that its promise could be realized, before Microsoft, Google, Facebook etc started their more recent investments.

A Brief Overview of Deep Learning
Ilya Sutskever
http://yyue.blogspot.sg/2015/01/a-brief-overview-of-deep-learning.html

 

[atyy]

I'm not familiar with the self-driving car applications,

Beware the advertising, of course, but here are some links.


NVIDIA Self-Driving Car Demo at CES 2017


NVIDIA Drive PX2 self-driving car platform visualized

http://selfdrivingcars.mit.edu/
6.S094: Deep Learning for Self-Driving Cars

 

[atyy]

Eventually however, problems that resist solution by the currently known techniques will start to appear; likely we're already aware of them but haven't yet recognized that they pose theoretical challenges that cannot be overcome by brute strength. When this happens the pendulum will swing back the other way. I find it interesting that the best computer programs for chess, bridge, and go use very different approaches; it seems unlikely that we've discovered all the promising approaches to machine problem solving.

Because the technology is old, many of the limitations have also been anticipated, eg. simple manipulations with intergers :)

https://stanford.edu/~jlmcc/Presentations/PDPMathCogLecture2015/PDPApproachMathCogCogSci.pdf

 

[PeroK]

Just Google's image recognition is something to behold. The other day I was wondering what a type of plant was in my yard. I took a photo of it on m phone, uploaded it to Google's image recognition app and it told me what plant it was. The plant was a "Brunnera macrophylla 'Jack Frost'". That capability is stunning to me and I doubt possible in the 80s.

In the 1980's you could have submitted your question to "Gardeners' Question Time", a much loved BBC radio programme.

 

[Ryan_m_b]

I'm not in the field at all but IMO if tools become available that allow you to test and develop old ideas in new and interesting ways then the field is progressing. It's not like new research is just repeating the same experiments as were done in the 80s, even if the fundamentals are the same.

 

[jerromyjon]

Google's new Tensor Processing Unit certainly takes machine learning to a new level... Apple is apparently working on a new AI chip for mobile devices as well. I can't wait to see what the future holds for this fascinating technology!

 

[Crass_Oscillator]

I actually agree with the argument that AI progress is far slower than people are giving it credit for. In particular, I doubt fully autonomous self driving cars are even remotely close to being deployed, and AlphaGo, while important, has been immensely overhyped.

Your perspective probably depends upon how impressed you are by these two mainstream, highly publicized applications. If you take a sober, conservative view of alphago and self driving, you probably perceive progress as far slower than laypeople do.

 

[Greg Bernhardt]

I doubt fully autonomous self driving cars are even remotely close to being deployed, and AlphaGo, while important, has been immensely overhyped.

I would be interested in an explanation with greater detail.

 

[Crass_Oscillator]

Sure, I'll just write some short points and you can ask questions about it. I'll stick to AlphaGo because less is known about what industry experts are doing with self driving cars, so for all I know they may possesses some magic I'm not aware of. The laziest thing I can do with SDC's is make an argument from authority, since a lot of academic experts have condemned the idea that we are anywhere near fully autonomous SDC's.

Regarding AlphaGo, the issues are:

DNN's are a very sloppy model, in the technical sense coined by Sethna (I can provide citations for the interested). In particular, it was found by Zhang et al (https://arxiv.org/pdf/1611.03530.pdf?from=timeline&isappinstalled=0) that DNN's, among other things, can achieve zero training error on randomly labeled or randomly generated data, pushing their generalization error arbitrarily high. To me this implies that DNN's have such enormous expressiveness that they can effectively memorize the dataset. With enough throughput and GPU toasters, you can span such an enormous portion of the Go gamespace that you can out muscle a human. Essentially it doesn't win via intelligence but via brutish Input/Output superiority that a human brain does not have access to. Consider the learning efficiency as a better measure (how many games must I win per game rank?). DeepMind is now moving on to the real time strategy computer game Starcraft which I think will illustrate this point very poignantly, since the data is much harder to handle. Moreover, they are much more carefully forcing I/O limitations on their "AI" algorithms so that I/O is properly normalized out.

All this said, clearly DNN's will have niche applications, it's just that they have been portrayed (largely by the media) in a highly misleading manner.

 

[stoomart]

I actually agree with the argument that AI progress is far slower than people are giving it credit for.

Would you agree that progress was slow due to limited technology resources?

Your perspective probably depends upon how impressed you are by these two mainstream, highly publicized applications. If you take a sober, conservative view of alphago and self driving, you probably perceive progress as far slower than laypeople do.

My perspective is from practical use of machine learning in IT security since around 2012 for DLP purposes. Just recently I've been researching non-signature behavior-based antivirus, and as a person who supported Norton/Symantec stuff since Windows 95, I can say this new breed of analytics-driven security tools just wasn't possible very long ago.

 

[Crass_Oscillator]

Would you agree that progress was slow due to limited technology resources?My perspective is from practical use of machine learning in IT security since around 2012 for DLP purposes. Just recently I've been researching non-signature behavior-based antivirus, and as a person who supported Norton/Symantec stuff since Windows 95, I can say this new breed of analytics-driven security tools just wasn't possible very long ago.

I do agree that progress was impacted by hardware, but I also don't consider the theoretical progress to be all that impressive. It's important, but it is not revolutionary.

ImageNet being conquered by DNN's was not AI's equivalent of the transistor being invented or general relativity. From a mathematical and theoretical point of view I consider the invention of the FFT to be a much greater achievement. We don't even have theoretical clarity regarding the algorithm.

That said I know nothing about IT security, although I would start by guessing that people have gotten a lot farther with simple Bayesian methods than DNN's, which can also require a lot of horsepower. Is this correct, or are DNN's a big part of modern security software?

 

[stoomart]

I do agree that progress was impacted by hardware, but I also don't consider the theoretical progress to be all that impressive. It's important, but it is not revolutionary.

ImageNet being conquered by DNN's was not AI's equivalent of the transistor being invented or general relativity. From a mathematical and theoretical point of view I consider the invention of the FFT to be a much greater achievement. We don't even have theoretical clarity regarding the algorithm.

That said I know nothing about IT security, although I would start by guessing that people have gotten a lot farther with simple Bayesian methods than DNN's, which can also require a lot of horsepower. Is this correct, or are DNN's a big part of modern security software?

Security has always been done with brute force tactics like building massive signature and intelligence databases to detect viruses, spam, and malicious network traffic; I see this as analogous to programming Deep Blue to play chess. When attackers upped their game with polymorphic and memory-only code, now we need something that can profile behavior rather collecting signatures, which has only recently started to mature due to machine learning and modern hardware.

My assumption is mastering the practical use of DNNs will drive innovation to bigger and better ideas, which may even come from the AI we build. Edit: I think a good start would be to train DeepMind how to do maths the same way it learned to play Go and Atari games, through observation/feedback instead of programming.

 

[jerromyjon]

I think a good start would be to train DeepMind how to do maths

This leads to the question of true intelligence... would DeepMind simply be able to mimic everything that is already known and understood by humans, or would it be able to "fill in the blanks" of what is yet undiscovered?

I'm firmly planted on @Crass_Oscillator 's side of the fence... there really isn't much more going on than brute force techniques with some simplification of data points to "compress" the data into workable volume. It still has great potential on it's own IMO, for example having an AI "supervisor" to learn common routines and point out human errors should be easily implemented, but I don't believe it would be able to "figure out" improvements to routines very well.

 

[stoomart]

This leads to the question of true intelligence... would DeepMind simply be able to mimic everything that is already known and understood by humans, or would it be able to "fill in the blanks" of what is yet undiscovered?

I suggest intelligence is the ability to learn, adapt, and improve, which I believe this thing is clearly demonstrating. Mimicking would be its first step, and optimization is where it would go.

 

[Crass_Oscillator]

I suggest intelligence is the ability to learn, adapt, and improve, which I believe this thing is clearly demonstrating. Mimicking would be its first step, and optimization is where it would go.

Suppose I have two students. One student is a typical A/B American high school student and scores in the 87th percentile on the SAT (an entrance exam) after working a single practice test.

The second student is an A/B student who memorizes the patterns of 100 million SAT practice/old tests and scores in the 99th percentile.

Who is exhibiting more intelligence?

 

[stoomart]

Suppose I have two students. One student is a typical A/B American high school student and scores in the 87th percentile on the SAT (an entrance exam) after working a single practice test.

The second student is an A/B student who memorizes the patterns of 100 million SAT practice/old tests and scores in the 99th percentile.

Who is exhibiting more intelligence?

I would say the student with a life. : )

 

[cosmik debris]

Suppose I have two students. One student is a typical A/B American high school student and scores in the 87th percentile on the SAT (an entrance exam) after working a single practice test.

The second student is an A/B student who memorizes the patterns of 100 million SAT practice/old tests and scores in the 99th percentile.

Who is exhibiting more intelligence?

Under most educational systems the latter is the most intelligent. It's the test that's at fault; I guess that's your point.

 

[GTOM]

For a look into at least one possible AI future, I recommend Ray Kurzweil's books.

He is now a dreamer hope to become immortal, which isn't likely to happen.
http://www.softmachines.org/wordpress/wp-content/uploads/2016/04/Against_Transhumanism_1.0_small.pdf

Personally i wonder, how serious hardware is needed to run Alphago and things like that?
Last time i read Watson that defeated champion in Jeopardy had the size of ten refrigirators.

 

[GTOM]

Under most educational systems the latter is the most intelligent. It's the test that's at fault; I guess that's your point.

Our education system is pretty bad IMHO... lots of cramming in useless things.
It should focus on basics, then creative problem solving together.

 

[mitchell porter]

Google is SkyNet. Kim Jong Un is John Connor, and will save the human race by EMPing Silicon Valley. Then Tencent from China will step in, and become the Matrix.

 

[StatGuy2000]

I do agree that progress was impacted by hardware, but I also don't consider the theoretical progress to be all that impressive. It's important, but it is not revolutionary.

In general, in the case of machine intelligence, how important do you think it is for the AI field to have a revolutionary progress, as opposed to incremental progress of the sort you are speaking of?

 

[1oldman2]

There's no end in sight.
http://www.sciencemag.org/news/2017/06/artificial-intelligence-learns-spot-pain-sheep

 

[GTOM]

Sure, I'll just write some short points and you can ask questions about it. I'll stick to AlphaGo because less is known about what industry experts are doing with self driving cars, so for all I know they may possesses some magic I'm not aware of. The laziest thing I can do with SDC's is make an argument from authority, since a lot of academic experts have condemned the idea that we are anywhere near fully autonomous SDC's.

Regarding AlphaGo, the issues are:

DNN's are a very sloppy model, in the technical sense coined by Sethna (I can provide citations for the interested). In particular, it was found by Zhang et al (https://arxiv.org/pdf/1611.03530.pdf?from=timeline&isappinstalled=0) that DNN's, among other things, can achieve zero training error on randomly labeled or randomly generated data, pushing their generalization error arbitrarily high. To me this implies that DNN's have such enormous expressiveness that they can effectively memorize the dataset. With enough throughput and GPU toasters, you can span such an enormous portion of the Go gamespace that you can out muscle a human. Essentially it doesn't win via intelligence but via brutish Input/Output superiority that a human brain does not have access to. Consider the learning efficiency as a better measure (how many games must I win per game rank?). DeepMind is now moving on to the real time strategy computer game Starcraft which I think will illustrate this point very poignantly, since the data is much harder to handle. Moreover, they are much more carefully forcing I/O limitations on their "AI" algorithms so that I/O is properly normalized out.

All this said, clearly DNN's will have niche applications, it's just that they have been portrayed (largely by the media) in a highly misleading manner.

As far as i know Go gamespace is practically infinite, so sheer brute force and memory isn't enough.

 

[Crass_Oscillator]

As far as i know Go gamespace is practically infinite, so sheer brute force and memory isn't enough.

DNN's are kind of like "brute force learning" which is distinct from a brute force search. The SAT student I used as an analogy who has studied 100 million SAT tests has a very high dimensional sloppy model of how to answer SAT questions, not a brute force search algorithm. To illustrate this, a common SAT problem might use some simple triangle geometry. If I know the basic theorems of triangle similarity, I can answer many questions with a low dimensional model. However there's no reason why I couldn't come up with some convoluted heuristic to answer such questions without knowing those theorems.

In general, in the case of machine intelligence, how important do you think it is for the AI field to have a revolutionary progress, as opposed to incremental progress of the sort you are speaking of?

I don't think AI is anywhere near achieving an acceptable solution to one of its outstanding problems, such as NLP. For that, we need to take more steps beyond, say, DNN's. But this will require something "paradigm shifting", to use a well abused phrase. For instance, if you examine the training process for any popular ML technique (or up and coming technique for that matter), the primary measure of success does not generally include parsimony of the model to my knowledge.

If I have a student who I am tutoring, "understanding" is defined in part by the fact that if I "rotate" the question to a new question that is effectively or approximately the same, the student still can answer it. In particular, there is some notion of explaining more with less. If I have two parameters in my model which answers 100 questions, and you have 100 parameters to answer 100 questions, we might suppose that you don't really understand anything about the subject; if they are 100 yes-no questions, and your parameters are all just the answers, then we would argue that you know basically nothing.

Something something symmetry/information theory something something.

 

[AaronK]

Here's a good overview paper (in pdf format) for the most interesting stuff (in my opinion) being worked on at present if anyone is interested: Deep Reinforcement Learning: An Overview

Here's the paper's abstract:

"We give an overview of recent exciting achievements of deep reinforcement learning (RL). We start with background of deep learning and reinforcement learning, as well as introduction of testbeds. Next we discuss Deep Q-Network (DQN) and its extensions, asynchronous methods, policy optimization, reward, and planning. After that, we talk about attention and memory, unsupervised learning, and learning to learn. Then we discuss various applications of RL, including games, in particular, AlphaGo, robotics, spoken dialogue systems (a.k.a. chatbot), machine translation, text sequence prediction, neural architecture design, personalized web services, healthcare, finance, and music generation. We mention topics/papers not reviewed yet. After listing a collection of RL resources, we close with discussions."

Also: In particular, AlphaGo is discussed in section 12 of the paper--for anyone that just wants to read about that.

 

[AaronK]

Another great (hardware) advancement for deep learning that I've just learned about: a team of researchers at MIT have apparently managed to develop what they call a "programmable nanophotonic processor" as a more capable replacement for GPUs. Obviously, the dense matrix multiplications necessary for AI learning tasks are time consuming and use a lot of power. Using NPUs(?), the MIT team expect a computational speed enhancement of at least two orders of magnitude over the currently most powerful equivalent GPUs and an incredible three orders of magnitude in power efficiency. For demonstration, they implemented a basic neural network and just using their prototype system, they were able to achieve a 77% accuracy level, compared to approx. 90% for conventional systems--and it seems they are confident it won't be very technically difficult to scale for greater accuracy.

I can't wait until this is more accurate and can be made commercially available at the price of traditionally powerful GPUs. Something like this could make it a lot easier to do much more computationally heavy training and experiments with neural networks for individual researchers/machine learning enthusiasts (also, I don't want to pay so much money for an equivalent GPU setup that will only allow me to do so much).

Here is the abstract: Artificial Neural Networks are computational network models inspired by signal processing in the brain. These models have dramatically improved the performance of many learning tasks, including speech and object recognition. However, today's computing hardware is inefficient at implementing neural networks, in large part because much of it was designed for von Neumann computing schemes. Significant effort has been made to develop electronic architectures tuned to implement artificial neural networks that improve upon both computational speed and energy efficiency. Here, we propose a new architecture for a fully-optical neural network that, using unique advantages of optics, promises a computational speed enhancement of at least two orders of magnitude over the state-of-the-art and three orders of magnitude in power efficiency for conventional learning tasks. We experimentally demonstrate essential parts of our architecture using a programmable nanophotonic processor.

Link to the arXiv page where you can get access to the paper if you'd like: Deep Learning with Coherent Nanophotonic Circuits

 

[Buzz Bloom]

From my perspective from the 60s and 70s, progress seemed rapid then due to the quantity of low hanging fruit. Then for several decades the problems got harder, and the limitation of computer speed and capacity were handicaps. Also from that early era, there were attitudes that different approaches were in the wrong direction. For example, at MIT the heuristic programing approach was the way to go, and pattern recognition was looked at as not really AI.

A common metaphor at that time was that heuristic programing imitated the "left brain" intelligence where conscious decisions were made, while pattern recognition was trying to imitate "right brain" gestalt processing that was not performed consciously. I think this attitude lasted for quite a while discouraging researchers from combining different techniques. AlphaGo is a good example of doing that - combining the "left brain" look-ahead techniques with "right brain" pattern recognition techniques.

In a conversation I had recently with an AI professor at Johns Hopkins, he explained how combining techniques has become a lot more common in the last decade or so.

 

[DiracPool]

A common metaphor at that time was that heuristic programing imitated the "left brain" intelligence where conscious decisions were made, while pattern recognition was trying to imitate "right brain" gestalt processing that was not performed consciously. I think this attitude lasted for quite a while discouraging researchers from combining different techniques.

That sounds about right.

AlphaGo is a good example of doing that - combining the "left brain" look-ahead techniques with "right brain" pattern recognition techniques.

And how does it do that?

In a conversation I had recently with an AI professor at Johns Hopkins, he explained how combining techniques has become a lot more common in the last decade or so.

Can you elaborate a little on that conversation?

 

[Buzz Bloom]

And how does it do that?

Hi Dirac:

I may be mistaken about this, but my understanding is as follows.

The look-ahead component uses (1) a standard min-max tree-searching method with (2) position evaluation at nodes with no deeper moves to evaluate.

The pattern recognition component (1) accesses a very large database of positions each associated with good moves, and (2) a method for finding positions in the database that are similar to each position for which good plausible moves are to be used for deeper exploration.

Can you elaborate a little on that conversation?

I actually have some notes from the conversation for some suggestions for further reading, but I haven't yet used them, and I can't look for them now. I will see if I can find them tomorrow, and if I find them I will post them.

Regards,
Buzz

 

[Buzz Bloom]

Can you elaborate a little on that conversation?

Hi Dirac:

I found the notes. Unfortunately, the notes are about math rather than AI, but they did refresh my memory a bit about the conversation.

I had asked him if he know about any research related to an idea I have had for a long time which involves a specific combination of AI methods. I do not have an opportunity very often to discuss AI with someone with expertise, and when I do I ask this same question. Up until this conversation the answer has always been "no". On this occasion, however, the answer was "yes", and he told me he had recently read about such a research project, but he did not remember the details. I had planned to later email him to see if he could locate a source about this project, but it had slipped my mind completely until now.

If I can get this source information I will post it.

Regards,
Buzz

 

[jerromyjon]

This "multi-joint dynamics with contact" (www.mujoco.org) looks rather interesting! (from AaronK's post #38 link) Virtual reality for testing AI's... pretty cool stuff.

 

[Buzz Bloom]

Hi @Dirac:

My Hopkins friend had been away at a conference, but he has recently returned and responded to my email. The following is from his email message.

The general setting that you are describing is called "feature selection" or sometimes (when the set of potential features is very large or infinite) "feature induction." Another term is "structure learning," e.g., learning the structure of a graphical model.​

Standard methods for doing this include

• selection of independent variables from a large but fixed feature space
  • stepwise selection (forward selection or backward selection), as in statistics
  • simultaneous feature selection and weight optimization
    • via sparsity-inducing regularization (possibly structured sparsity)
    • via an explicit prior distribution over the set of features (may require reversible-jump MCMC or other randomized search algorithms to find the feature set)
• decision trees, which gradually refine features -- they build up more and more complex conjunctive features as the depth of the tree increases
  • decision forests (a much more effective version of decision trees, which can be regarded as involving weights)
• split-merge EM, another way of gradually refining features by refining the domains of the latent variables that are used to generate feature templates
• inductive logic programming
  • more recent methods of learning programs, e.g., the neural Turing machine, or some work in probabilistic programming
• active set methods, which obtain new candidate features at each step by conjoining existing features that have nonzero weight. The seminal paper may have been https://arxiv.org/pdf/1212.2504.pdf .
• neural networks: recall that the hidden layer of a neural net can be regarded as a set of induced features, also known as "adaptive basis functions"

You may notice that the next to last bullet includes a link to a reference. My friend said he would give provide references for any of these topics in which you have an interest.

Regards,
Buzz

 

[DiracPool]

Regards,
Buzz

Hi Buzz, thanks for that. However, I'm not sure if or how informative that post is going to be for people who want to learn artificial intelligence. It looks to me like a bunch of jargon highlighted by bullet points. I think that this is a problem, no a solution. I'm not saying I have a solution right now either--I just can say that your post doesn't educate nor inspire me. Just saying.

I'm working under a department of defense contract in order to develop autonomous rovers. We don't know whether they'll be used for roving Mars or for some other task, but one of the things we demand is that we reduce jargon to a minimum. While the machine learning lingo sounds cool in the public market, we don't use this in the lab settings, at least in my lab setting. So, if you want to make a point, make it in a more colloquial manner.

 

[Buzz Bloom]

So, if you want to make a point, make it in a more colloquial manner.

Hi Dirac:

I am sorry my post was not helpful to you. I was not really trying to make any point, but just trying to share something I thought might be of interest. I am guessing that the language use in the bullets about the various technological approaches is common in academia even if it is avoided in non-academic labs.

I scanned the linked McCallum article, and it also has a lot of technical jargon unfamiliar to me, but I think it can be deciphered with some patience. It may well be easier for you to read since you work in the AI field, but I have not since graduate school half a century ago.

Regards,
Buzz

 

[DiracPool]

I am guessing that the language use in the bullets about the various technological approaches is common in academia even if it is avoided in non-academic labs.

Well, I'm definitely in an academic lab at a major university. It's a computer science lab and we are under a DARPA grant for 4 years, although we get evaluated quarterly and I'm guessing could have the funding pulled at any point. Or at least this is what the lab director likes us to think.

By virtue of the fact that I work in the CS lab I am constantly bombarded with invitations to watch PhD defenses in the main lecture hall, as well as lectures from visiting scholars. This happens once or twice a week and there's always a big food smorgasbord in the lobby for these events. So I typically partake in these events to get a free lunch. The dissertation defenses are typically boring, but the invited lectures are often interesting.

 

[Buzz Bloom]

And how does it do that?

Hi Dirac:

I found a January 2016 paper describing the techniques used in a go playing bot similar to that used in AlphaGo's March 2016 match with Lee Se-dol, and I found that I was partially mistaken in my description in post #42. The tree search method is much more complicated than what I had in mind.

The combination of the two techniques as described in this paper is too complex for me to try to summarize it adequately here. However, I will make an effort anyway by paraphrasing a few quotes.

The two techniques combined as described in this paper:
1, Deep Convolutional Neural Network (DCNN)
2. Monte Carlo Tree Search (MCTS)

(1) uses a pure pattern-matching approach that predicts the next move.

(2) does a tree search, but in a different way than the more primitive ones with which I am familiar. Each round of Monte Carlo tree search consists of four steps.
( From https://en.wikipedia.org/wiki/Monte_Carlo_tree_search )

• Selection: start from root R and select successive child nodes down to a leaf node L. The section below says more about a way of choosing child nodes that let's the game tree expand towards most promising moves, which is the essence of Monte Carlo tree search.
• Expansion: unless L ends the game with a win/loss for either player, create one (or more) child nodes and choose node C from one of them.
• Simulation: play a random playout from node C. This step is sometimes also called playout or rollout.
• Backpropagation: use the result of the playout to update information in the nodes on the path from C to R.

I hope this is helpful.

Regards,
Buzz

 

[jerromyjon]

pro.mising moves,

Should say "all possible" moves...