Stanford AI Class thoughts, and a brief tour of AI history

"people who face a difficult question often answer an easier one instead, without realizing it"
— Daniel Kahneman,

Despite time management issues (which will only get worse this week) I managed to struggle through the first two weeks, or four units, of the online Stanford introduction to artificial intelligence course.

In the past I had already tried to struggle through Judea Pearl's Probabilistic Reasoning in Intelligent Systems. That was published well over 20 years ago, and yet this course uses many of the same examples. The course is much more about working out actual examples; it is practical, not so theoretical. We've covered Bayes networks and conditional probability, both concepts I had already learned because Numenta was using them. Pearl's book contains a lot of material about wrong directions to take; the Stanford course seems to be focussed on what actually works, at least for Google.

My impression is still that the Stanford AI paradigm, while very practical, is not going to provide the core methods for truly intelligent machines, which I characterize as machine understanding. I think this largely because I am ancient and have watched quite a few AI paradigms come and go over the decades.

When AI got started, let's say in the 1950's, there was an obsession with logic as the highest form of human intelligence (often equated with reasoning). That computers operated with logic circuits seemed a natural match. Math guys who were good at logic tended to deride other human activities as less difficult and requiring less intelligence. Certain problems, including games with limited event spaces (like checkers), could be solved more rapidly by computers (once a human had written an appropriate program) than by humans. By the sixties, at latest by the seventies, computers running AI programs would be smarter than humans. In retrospect, this was idiotic, but the brightest minds of those times believed it.

One paradigm that showed some utility was expert systems. To create one of these, experts (typical example: a doctor making a diagnosis) were consulted to find out how they made decisions. Then a flow chart was created to allow a computer program to make a decision in a similar manner. As a result a computer might appear intelligent, especially if provided by the then more difficult trick of an audio imitation voice output, but today no one would call such a system intelligent. That is no more intelligent than the early punch card sorters that once served as input and output for computers that ran on vacuum tubes.

In the 1980's there was a big push in artificial neural networks. This actually was a major step towards machines being able to imitate human brain functions. It is not a defunct field. Some practical devices today work with technologies developed in that era. But scaling, the problems grew faster than the solutions. No one could build an artificial human brain out of the old artificial neural networks. We know that if we can exactly model a human brain, down to the molecular (maybe even atomic) level, we should get true artificial intelligence. But simplistic systems of neurons and synapses are not easy to assemble into a funcioning human brain analog.

The Stanford model for AI has been widely applied to real world problems, with considerable success. This probabilistic model allows it to deal with more complex data than the old logic and expert system paradigms ever could. Machine systems really can learn new things about their environment and store and organize that information in a way that allows for practical decision making. Clearly that is one thing human brains can do, and it is a lot more difficult than playing in a set-piece world like tic-tac-toe or even chess.

Sad as the state of human reasoning can be at times, and as slow as we are to learn new lessons, and as proud as we are of our least bouts of creativity, (and as much as we may occasionally ignore the rule against run-on sentences), I think the Stanford model is not, by itself, going to lead to machine understanding. The human brain has a lot of very interesting structures at the gross level and at the synaptic level. Neurologists have not yet deciphered them. Their initial "programming," or hard-wiring is purely the result of human evolution.

When is imitated intelligence real intelligent? When does a machine (or a human, for that matter) understand something, as opposed to merely changing internal memory to reflect the external reality?

Then again, maybe a Stanford model computer/program/input/output system would have done better at the Stanford AI course than I have. I certainly have not been getting all the quizzes and homework problems right on the first try. On the other hand, I think it will be some good long time before a machine can read, say, a book on neurology and carry on an extended intelligent conversation about it.

Pointing Choices

Quote of the day:

"This convention is at variance with that used in many expert systems (e.g. MYCIN), where rules point from evidence to hypothesis (e.g., if symptom, then disease), thus denoting a flow of mental inference. By contrast, the arrows in Bayesian networks point from causes to effects, or from conditions to consequences, thus denoting a flow of constraints attributed to the physical world."

Judea Pearl, Probabilistic Reasoning in Intelligent Systems, p. 151

Understanding Probability and Probabilistic Reasoning

I think I mentioned earlier that I have had problems with the Numenta model as described in "Towards a Mathematical Theory of Cortical Micro-circuits" because of the use of probability-based mathematics. It seems to me that neurons are deterministic mechanisms. But I have noted in the past that I can be pretty dim-witted at times, and decided to study Numenta's HTM systems anyway. I reminded myself that quantum physics has two different formulations, one based on matrix algebra and the other on the Schrodinger equation. They both work, and some brilliant person showed that they are formally the same long, long ago. So when thinking about or solving problems you can use whichever is easiest or gives the best insights. The same way some physics and math problems are easier in polar coordinates than in rectangular coordinates.

Months ago I ground to a halt in my reading of Judea Pearl's Probabilistic Reasoning in Intelligent Systems, which provides much of the background to the Numenta discussion. Yesterday I decided to tackle it again and commenced reading at page 143. I noticed that some notation was ambiguous, which is typical of expert writers who assume their readers are right up with them. So I decided to go back and make sure that P(A,B) really does mean the probability that both A and B are true. I thought I'd make sure I understood the Bayes interpretation of probability as well.

I ended up reading starting at page 29, Chapter 2, Bayesian Inference, 2.1, Basic Concepts, 2.1.1 Probabilistic Formulation and Bayesian Inversion. Note that I took two semesters of logic and one semester of probability in college, and as part of my profession deal with biostatistics, the kind reported from clinical trials, on a regular basis. Note also that I have studied philosphic issues of quantum physics and even the math involved.

Yet when I read this simple introduction this time, the scales fell from my eyes, or from my cortical networks.

With probabilistic reasoning, it is fair to say that we are not talking about rolling dice (even though Pearl uses the familar probabilities of two-die rolls to illustrate some points).

We are talking about the math of pobability theory. For most practical purposes, that is the math of fractions. Third or fourth grade stuff. (I had a fifth grade teacher I hated, Mrs. Lopez, who was all about memorizing things. We memorized the decimal equivalents of about 50 common fractions. I knew I could always get the decimal equivalent by dividing, so considered this a stupid exercise.)

When thinking about human memory, you can safely substitute "percentage of like situations" for probability.

Updating the "percentage of like situations" based on experience makes sense. Since we can test for novel situations, like "both A and B" or "A and not C, given B", by multiplying, adding, or subtracting fractions, these updates may effect a chain of knowledge or deductions across the brain (or mind, if you prefer).

Calling all background information and assumptions a person has K (I don't know why K, maybe it stands for Knowledge), I quote Pearl page 30: "However, when the background information undergoes changes, we need to identify specifically the assumptions that account for our beliefs and articulate explicitly K or some of its elements."

Many Philosophers, notably Ludwig Wittgenstein, have shown how reasoning goes awry when we use one word to mean multiple things, or one thing that is vague or complex. We think we are being clear using logic symbols or math equations or tech speak. But when something is amiss, it may not be a problem with our reasoning. It may be that we need to update our background assumptions.

See also Bayes' theorem

Bayesian Wasteland?

Despite being overloaded with other concerns, I continue to slog through two books that may (or may not) be helpful with the Machine Understanding project. On the information technology front we have Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference by Judea Pearl. I'm at page 116 as of last night. On the biological front I continue From Neuron to Brain by Stephen W. Kuffler; I'm on page 386.

Regarding Probabilistic Reasoning, so far I have seen a lot of interesting work on the problem of combining probability calculations with logic. I just finished the section on Markov Networks and am about to read up on Bayesian Networks. My problem so far is that I don't see any advantages to marrying anything I've seen of probabilistic reasoning to Jeff Hawkin's theory of predictive memory. Despite having read "Towards a Mathematical Theory of Cortical Micro-circuits," as reported in previous entries. Then again sometimes I have trouble taking up novel ideas. But my impression so far is that neural networks are not operating on a probability basis. The closest I can get, so far, to that kind of model is a signal-mixing basis, where analog functions might represent Pearl's probabilies. Nor do I see how the probability network models can cope with invariants. The word invariant is not found in the index to Probablistic Reasoning. I have higher hopes for a Tensor model, even though my own work on that is very preliminary. [I am reminded of the two seemingly totally different mathematical methods used in early quantum mechanics, which then were proven to be equivalants.]

In From Neuron, I have been reading astonishing details about how synapses and single nerve cells work, including how experiments were conducted, in mind-numbing detail. I am just getting to where how neurons and sets of neurons have been shown to operate, with the first example being neurons that sense the stretching of muscles. Again, I am pointed to tensors, which can be used to represent how multiple muscles representing various degrees of freedom of motion can lead to a coherent knowledge of where a body part is in three-dimensional, euclidean-modeled space.

More oddly, Neuron has basically nothing about Hebbian learning. True, the book is dated 1984. But did no one even try to find a physical basis for Hebbian learning as of that date? If you know of a definitive paper that appears to prove a biochemical mechanism for Hebbian learning, let me and my readers know.

I intended to explain what I was reading, in suitable chunks, in this blog. But it is easier to just keep reading at this point, rather than writing about details that I am not even sure are important yet. I keep finding I have to go back to basics. Today, worrying about tensors, muscles, feedback, and a neuron-level learning model, I am revisiting a two-neuron learning model that, oddly, I first worked on back in the 1980's. If anything comes of it, I'll let you know here.

On Bayesian Belief Propagation

Towards a Mathematical Understanding of Cortical Micro-circuits [Dileep George & Jeff Hawkins, PLoS Computational Biology, October 2009, Volume 5, Issue 10] assumes some knowledge of Bayesian Belief Propagation. I was hoping the article would explain this topic as it went along, but apparently, to AI specialists, this is a well-known area. I found the Wikipedia entry, belief propagation, jumps into mathematical formulism rather quickly. Some words of explanation without math symbols would probably help many people who are interested in machine understanding; that is my goal here. If some of you Bayesian belief propagation experts think I have it wrong, or want to add anything, feel free to add comments. I have ordered a copy of the classic work on the subject, Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference, and so hope to be more fluent in this area at a later date.

Bayesian Belief Propagation (BBP) was formulated by Judea Pearl, who later authored Probabilistic Reasoning in Intelligent Systems. It appears to be a system for determining unknown probabilities within certain types of complex systems. Let's start with a review of a very simple probability model.

Much of early probability work was formulated in terms of a pair of ordinary gambling dice, six-sided cubes with faces numbered one through six. Thrown many times, they make a nice distribution pattern of numbers two through twelve. It can be shown with a little math and logic that there are percentages that are expected for each of these numbers. If you make a large number of dice throws, say more than 100, and you get results that differ by very much from the expected distribution of results, you may be looking at one of two things. It is possible that you have a distribution skewed purely by chance. On the other hand, you might want to investigate whether your dice and throwing systems are physically constructed to be truly random in the results they give.

Now imagine a black box, say the size of a shoe box. If you shake the box and then stop shaking, it gives you a number. In fact, imagine a room with a variety of such boxes. You shake them, keeping tracks of the resulting numbers for each. If one box consistently gives numbers between 2 and 12 inclusive, and the frequencies for each number match the pattern that we believe two dice would give, then we can conclude that the internal mechanism of the box, whatever it is, has a probability distribution that is created by the same logic that creates the dice probability distribution. There might even be a tiny pair of nano-dice inside.

In computer programming we tend to think in purely causal mechanics. If something does not happen as expected, we have a bug in our program. But in the case of living groups of biological neurons we cannot (so far) look at the inner mechanics of each neuron while watching how the whole group operates. However, if we can control the inputs and read the outputs of a system, then, if we can do the math, we might figure out the probabilities of individual neurons firing given known inputs from the other neurons in the system.

Formally, BBP allows us to calculate the probability characteristics for nodes we can't observe directly, if we can observe enough of the rest of the system. A node does not need to be a neuron. In an AI system, a node is a processing unit with inputs, outputs, or both, that connect it to the system of nodes in question. In our case the nodes are part of a Bayesian network, also known as a belief network. Communication (or messaging) within the network is in terms of probabilities. (See also Bayesian probability)



Think of a network of three people, where each of the people represents a node. Let there be a piece of paper we will call the message (it is blank; the paper is the message, in this case). Let each person have a single die and the rule: if you have the message, throw the die. If it comes up a one, a two, or a three, pass the message to the left. Otherwise, pass it to the right. If we watch the game progress, the message moves around or back and forth in a random pattern. However, it is not a Bayesian network, because it is cyclical. Bayesian networks are not cyclic. Messages more only in one direction.

Now start a series of games in which the message is always handed to person 1 (node 1). Person 1 throws the die and passes the message to person 2 if the result is 3 or less, and to person 3 if the result is 3 or more. Person 2 also throws a die and simply holds the message if the result is 3 or less, but passes if to person 3 if the result is 3 or more. When person 3 gets a message the game is over.

It is a simple game and a tad boring, but if we focus on the propagation time, the number of die throws from the beginning of a game until the end of a game, we can see a set of variations. Over the long run they should form a probability distribution of the times.

Now change the rules slightly. Say person 2 forwards the message if the die comes out 2 or higher. The range of possible results does not change, but the probability distribution will.

Now suppose you know about this type of game, but you can't see the game because it is in a tent. You see a message handed in, and can time how long it is before the message is handed back out by person 3.

By keeping track of the times, you can get the probability distribution. That should tell you the exact rules of the game being played, including the probability that person 1 will give the message directly to person 3, and how long person 2 might keep the message before passing it.

In more complex networks, you need to pin down the probabilities at more of the nodes, if you want to be able to characterize the probability rules for an unknown node.

Which is what the BBP algorithm is supposed to do. It calculates the probabilities for an unobservable node based on the observed nodes.