Where is AI Headed?

I've spent a lot of effort on this blog arguing for the direction of higher expressiveness. Machine intelligence should be able to learn anything a human can learn, and in order for that to be possible, it should be able to conceive of any concept that a human can. I have proceeded with the belief that this is the direction to push in order for the field to make progress.

Yet, in some ways at least, the field is headed in the opposite direction.

I've often discussed the Chomsky hierarchy, and how most techniques at present fall very low on it. I've often discussed hierarchies "above" the Chomsky hierarchy; hierarchies of logic & truth, problems of uncomputability and undefinability. Reaching for the highest expression of form, the most general notion of pattern.

Machine learning has made artificial intelligence increasingly practical. Yet, the most practical techniques are often the least expressively powerful. Machine learning flourished once it abandoned the symbolic obsession of GOFAI. Fernando Pereira famously said: "The older I get, the further down the Chomsky Hierarchy I go."

There's a good reason for this, too. Highly structured techniques like logic induction and genetic programming (both of which would go high in the hierarchy) don't scale well. Commercial machine learning is large-scale, and increasingly so. I mentioned this in connection with word2vec last time: "Using very shallow learning makes the technique faster, allowing it to be trained on (much!) larger amounts of data. This gives a higher-quality result." 

The "structure" I'm referring to provides more prior bias, which means more generalization capability. This is very useful when we want to come to the correct conclusion using small amounts of data. However, with more data, we can cover more and more cases without needing to actually make the generalization. At some point, the generalization becomes irrelevant in practice.

Take XML data. You can't parse XML with regular expressions.¹ Regular expressions are too low on the Chomsky hierarchy to form a proper model of what's going on. However, for the Large Text Compression Benchmark, which requires us to compress XML data, the leading technique is the PAQ compressor. Compression is equivalent to prediction, so the task amounts to making a predictive model of XML data. PAQ works by constructing a probabilistic model of the sequence of bits, similar to a PPM model. This is not even capable of representing regular expressions. Learning regular expressions is like learning hidden markov models. PPM allows us to learn fully observable markov models. PAQ learns huge markov models that get the job done.

The structure of XML requires a recursive generalization, to understand the nested expressions. Yet, PAQ does acceptably well, because the depth of the recursion is usually quite low.

You can always push a problem lower down on the hierarchy if you're willing to provide more data (often exponentially more), and accept that it will learn the common cases and can't generalize the patterns to the uncommon ones. In practice, it's been an acceptable loss.

Part of the reason for this is that the data just keeps flowing. The simpler techniques require exponentially more data... and that's how much we're producing. It's only getting worse:

Has Big Data Made Anonymity Impossible? MIT Technology Review

At The New Yorker, Gary Marcus complains: Why Can't My Computer Understand Me? Reviewing the work of Hector Levesque, the article conveys a desire to "google-proof" AI, designing intelligence tests which are immune to the big-data approach. Using big data rather than common-sense logic to answer facts is seen as cheating. Levesque presents a series of problems which cannot (presently) be solved by such techniques, and calls others to "stop bluffing".

I can't help but agree. Yet, it seems the tide of history is against us. As the amount of data continues to increase, dumb techniques will achieve better and better results.

Will this trend turn around at some point?

Gary Marcus points out that some information just isn't available on the web. Yet, this is a diminishing reality. As more and more of our lives are online (and as the population rises), more and more will be available in the global brain.

Artificial intelligence is evolving into a specific role in that global brain: a role which requires only simple association-like intelligence, fueled by huge amounts of data. Humans provide the logically structured thoughts, the prior bias, the recursive generalizations; that's a niche which machines are not currently required to fill. At the present, this trend only seems to be increasing.

Should we give up structured AI?

I don't think so. We can forge a niche. We can climb the hierarchy. But it's not where the money is right now... and it may not be for some time.

1: Cthulhu will eat your face.

History of Distributed Representations

Commenting on the previous post, a friend pointed out that "distributed representations" are not so new. I thought I would take a look at the history to clarify the situation.

In a very broad sense, I was discussing the technique of putting a potentially nonlinear problem into a linear vector space. This vague idea matches to many techniques in machine learning. A number of well-developed algorithms take advantage of linearity assumptions, including PCA, logistic regression, and SVM.¹ A common approach to machine learning is to find a number of features, which are just functions of your data, and use one of these techniques on the features (hoping they are close enough to linear). Another common technique, the kernel trick, projects features into a higher-dimensional space where the linearity assumption is more likely to get good results. Either way, a large part of the work to get good results is "feature engineering": choosing how to represent the data as a set of features to feed into the machine learning algorithm.

We could even argue that probability theory itself is an example: probabilities are always linear, no matter how nonlinear the underlying problem being described. (The probability of an event is the sum of the ways it could happen.) This gives us nice results; for example, there is always a Nash equilibrium for a game if we allow probabilistic strategies. This is not the case if we only consider "pure" strategies.

This theme is interesting to me, but, I was trying to be much more narrow in talking about recent developments in distributed representations. Like feature-based machine learning, a distributed representation will put data into a vector space to make it easier to work with. Unlike approaches relying on feature engineering, there is an emphasis on figuring out how to get the representations to "build themselves", often starting with randomly assigned vector representations.

The beginning of this kind of approach is probably latent semantic analysis (LSA), which is from 1988. LSA assigns 'semantic vectors' to words based on statistical analysis of the contexts those words occur in, based on the idea that words with similar meaning will have very similar statistics.

Given how old this technique is, the excitement around Google's release of the word2vec tool is striking. Reports spun it as deep learning for the masses. Deep learning is a much more recent wave of development. I think the term has lost much of its meaning in becoming a buzzword.² Calling word2vec "deep" takes this to farcical levels: the techniques of word2vec improve previous models by removing the hidden layer from the network. Using very shallow learning makes the technique faster, allowing it to be trained on (much!) larger amounts of data. This gives a higher-quality result.

One of the exciting things about word2vec is the good results with solving word analogies by vector math. The result of vector computations like France - Paris and Russia - Moscow are very similar, meaning we can approximately find the vector for a capital given the vector for the corresponding nation. The same trick works for a range of word relationships.

However, I've talked with people who had the incorrect impression that this is a new idea. I'm not sure exactly how old it is, but I've heard the idea mentioned before, and I did find a reference from 2004 which appears to use LSA to do the same basic thing. (I can't see the whole article on google books...)

One thing which I thought was really new was the emerging theme of combining vectors to form representations of compound entities. This, too, is quite old. I found a paper from 1994, which cites harder-to-find papers from 1993, 1990, and 1989 that also developed techniques to combine vectors to create representations of compound objects. Recent developments seem much more useful, but, the basic idea is present.

So, all told, it's a fairly long-standing area which has seen large improvements in the actual techniques employed, but, whose central ideas were laid out (in one form or another) over 20 years ago.

1: By the way, don't get too hung up about what makes one machine learning technique "linear" and another "nonlinear". This is a false dichotomy. What I really mean is that a technique works in a vector space (which more or less means a space where + is defined and behaves very much like we expect), and relies "largely" on linear operations in this space. What does "linear" actually mean? A function F is linear if and only if F(x+y) = F(x) + F(y) and for scalar a, F(ax) = aF(x). PCA, for example, is justified by minimizing a squared error (a common theme), where the error is based on euclidean distance, a linear operation. Notice that taking the square isn't linear, but PCA is still thought of as a linear approach.

2: Deep learning has come to mean almost any multi-layer neural network. The term caught on with the success related to Deep Belief Networks, which proposed specific new techniques. Things currently being called "deep learning" often have little in common with this. I feel the term has been watered down by people looking to associate their work with the success of others. This isn't all bad. The work on multi-layered networks seems to have produced real progress in reducing or eliminating the need for feature engineering.

Distributed Representations

Distributed vector representations are a set of techniques which take a domain (usually, words) and embed it into a linear space (representing each word as a large vector of numbers). Useful tasks can then be represented as manipulations of these embedded representations. The embedding can be created in a variety of ways; often, it is learned by optimizing task performance. SENNA demonstrated that representations learned for one task are often useful for others.

There are so many interesting advances being made in distributed vector representations, it seems that a nice toolset is emerging which will soon be considered a basic part of machine intelligence.

Google's word2vec assigns distributed vector representations to individual words and a few short phrases. These representations have been shown to give intuitively reasonable results on analogy tasks with simple vector math: king - man + woman is approximately equal to the vector for queen, for example. This is despite not being explicitly optimized for that task, again showing that these representations tend to be useful for a wide range of tasks.

Similar approaches have aided machine translation tasks by turning word translation into a linear transform from one vector space to another.

One limitation of this approach is that we cannot do much to represent sentences. Sequences of words can be given somewhat useful representations by adding together the individual word representations, but this approach is limited.

Socher's RNN learns a matrix transform to compose two elements together and give them a score, which is then used for greedy parsing by composing together the highest-scoring items, with great success. This gives us useful vector representations for phrases and sentences.

Another approach which has been suggested is circular convolution. This combines vectors in a way which captures ordering information, unlike addition or multiplication. Impressively, the technique has solved Raven progressive matrix problems:

http://eblerim.net/?page_id=2383

Then there's a project, COMPOSES, which seeks to create a language representation in which nouns get vector representations and other parts of speech get matrix representations (and possibly tensor representations?).

http://clic.cimec.unitn.it/composes/

I haven't looked into the details fully, but conceptually it makes sense: the parts of speech which intuitively represent modifiers are linear functions, while the parts of speech which are intuitively static objects are getting operated on by these functions.

The following paper gives a related approach:

http://www.cs.utoronto.ca/~ilya/pubs/2008/mre.pdf

Here, everything is represented as a matrix of the same size. Representing the objects as functions is somewhat limiting, but the uniform representation makes it easy to jump to higher-level functions (modifiers on modifiers) without adding anything. This seems to have the potential to enable a surprisingly wide range of reasoning capabilities, given the narrow representation.

As the authors of that last paper mention, the approach can only support reasoning of a "memorized" sort. There is no mechanism which would allow chained logical inferences to answer questions. This seems like a good characterization of the general limitations of the broader set of techniques. The distributed representation of a word, phrase, image, or other object is a static encoding which represents, in some sense, a classification of the object into a fuzzy categorization system we've learned. How can we push the boundary here, allowing for a more complex reasoning? Can these vector representations be integrated into a more generally capable probabilistic logic system?

Progress in Logical Priors

It's been a while since I've posted here. I've been having a lot of ideas, but I suppose the phd student life makes me focus more on implementing than on speculating (which is a good thing).

I presented my first sole-authored paper (based on this blog post) at the AGI conference in December of last year, and it was cited by Marcus Hutter in an interesting paper about approximating universal intelligence (which was presented at this year's AGI conference, which was once again a summer conference, so already took place).

When I set out to write the paper, my main goal was to show that we gain something by representing beliefs as something like logical statements, rather than as something like programs. This allows our beliefs to decompose more easily, readily allows inference in "any direction" (whereas programs are naturally executed in one direction, producing specific results in a specific order), and also allowing incomputable hypotheses to be dealt with in a partial way (dealing somewhat more gracefully with the possibility by explicitly representing it in the hypothesis class, but incompletely).

My desire to put forward this thesis was partly out of an annoyance with people invoking the Curry-Howard isomorphism all-too-often, to claim that logic and computation are really one and the same. I still think this is misguided, and not what the curry-howard isomorphism really says when you get down to it. The "programs are proofs" motto is misleading. There is no consensus on how to deal with Turing-complete programs in this way; turing-complete programming languages seem to correspond to trivial logics where you can prove anything from anything!*

Annoyed, I wanted to show that there was a material difference between the two ways of representing knowledge.

As I wrote the paper and got feedback from colleagues, it became clear that I was fighting a losing fight for that thesis: although the first-order prior represented a new mathematical object with interesting features along the lines I was advocating, it would be possible to write a somewhat program-like representation with the same features. I would still argue each of the advantages I mentioned, and still argue against naive invocations of Curry-Howard, but I was trying to make these arguments too strong, and it wasn't working. In any case, this was a point that didn't need to be made in order for the paper to be interesting, for two reasons:

1. If desired, you could re-do everything in a more computational way. It would still be a new, interesting distribution with features similar to but different from the Solomonoff distribution.
2. A universal distribution over logic, done right, is interesting even if it had turned out to be somehow equivalent to the Solomonoff distribution.

So, all told, I downplayed the "logic is different from computation" side of the paper, and tried to focus more on the prior itself.

After submitting the paper, I went back to working on other things. Although I still thought about logical priors every so often, I didn't make very much conceptual progress for a while.

At the July MIRI workshop, I got the opportunity to spend time on the topic again, with other smart folks. We spend roughly a day going over the paper, and then discussed how to take things further.

The main problem with the first-order prior is that the probability of a universal statement does not approach 1 as we see more and more examples. This is because all the examples in the world will still be consistent with the statement "there exists a counterexample"; so, if we are randomly sampling sentences to compose a logical theory, the probability that we add that sentences doesn't drop below a certain minimum.

So, for example, if we are observing facts about the natural numbers, we will not converge to arbitrarily high probability for generalizations of these facts. To make it more concrete, we cannot arrive at arbitrarily high probabilities for the Goldbach conjecture by observing more and more examples of even numbers being written as the sum of two primes.

This isn't a bad thing in all cases. Holding back some fixed probability for the existence of a counterexample matches with the semantics of first-order logic, which is not supposed to be able to rule out omega-inconsistent theories. (Omega inconsistency is the situation where we deny a universal statement while simultaneously believing all the examples.)

For some domains, though, we really do want to rule out omega-inconsistency; the natural numbers are one of these cases. The reason the first-order prior allows some probability for omega-inconsistent possibilities is that first-order logic is unable to express the fact that natural numbers correspond exactly to the finite ones. ("Finite" cannot be properly characterized in first-order logic.) More expressive logics, such as second-order logic, can make this kind of assertion; so, we might hope to specify reasonable probability distributions over those logics which have the desired behavior.

Unfortunately, it is not difficult to show that the desired behavior is not approximable. If the probability of universal statements approaches 1 as we observe increasingly many examples, then it must equal 1 if we believe all the examples. Let's take an example. If we believe all the axioms of peano arithmetic, then we may be able to prove all the examples of the Goldbach conjecture. In fact, we end up believing all true Pi_1 statements in the arithmetic hierarchy. But this implies that we believe all true Sigma_2 statements, if our beliefs are closed under implication. This in turn means that we believe all the examples of the Pi_3 universal statements, which means we must believe the true Pi_3 with probability 1, since we supposed that we believe universal statements if we believe their examples. And so on. This process can be used to argue that we must believe the true statements on every level of the hierarchy.

Since the hierarchy transcends every level of hypercomputation, there can be no hope of a convergent approximation for it. So, convergence of universal statements to probability 1 as we see more examples is (very) uncomputable. This may seem a bit surprising, given the naturalness of the idea.

Marcus Hutter has discussed distributions like this, and argues that it's OK: this kind of distribution doesn't try to capture our uncertainty about logically undecidable statements. Instead, his probability distribution represents the strong inductive power that we could have if we could infallibly arrive at correct mathematical beliefs.

Personally, though, I am much more interested in approximable distributions, and approaches which do try to represent the kind of uncertainty we have about undecidable mathematical statements.

My idea has been that we can get something interesting by requiring convergence on the Pi_1 statements only.

One motivation for this is that Pi_1 convergence guarantees that a logical probability distribution will eventually recognize the consistency of any axiomatic system, which sort-of gets around the 2nd incompleteness theorem: an AI based on this kind of distribution would eventually recognize that any axioms you give it to start with are consistent, which would allow it to gradually increase its logical strength as it came to recognize more mathematical truth. This plausibly seems like a step in the direction of self-trusting AI, one of the goals of MIRI.

The immediate objection to this is that the system still won't trust itself, because it is not a set of axioms, but rather, is a convergent approximation of a probability distribution. Convergence facts are higher up in the arithmetic hierarchy, which suggests that the system won't be able to trust itself even if it does become able to (eventually) trust axiomatic systems.

This intuition turns out to be wrong! There is a weak sense in which Pi_1 convergence implies self-trust. Correctness for Pi_1 implies that we believe the true Sigma_2 statements, which are statements of the form "There exists x such that for all y, R(x,y)" where R is some primitive recursive relation. Take R to be "y is greater than x, and at time y in the approximation process, our probability of statement S is greater than c." (The arithmetic hierarchy can discuss the probability approximation process through a godel-encoding.) The relevant Sigma_2 statements place lower bounds on the limiting probabilities from our probability approximation. We can state upper bounds in a similar way.

This shows that a probability distribution which has Pi_1 convergence will obey something strikingly like the probabilistic reflection principle which came out of a previous MIRI workshop. If its probabilities fall within specific bounds, it will believe that (but the converse, that if it believes they fall within specific bounds, they do, does not hold). This gives such a probability distribution a significant amount of self-knowledge.

So, Pi_1 convergence looks like a nice thing to have. But is it?

During the MIRI workshop, Will Sawin proved that this leads to bad (possibly unacceptable) results: any logically coherent, approximable probability distribution over statements in arithmetic which assigns probability 1 to true pi_1 statements will assign probability 0 to some true pi_2 statements. This seems like a rather severe error; the whole purpose of using probabilities to represent uncertainty about mathematical truth would be to allow "soft failure", where we don't have complete mathematical knowledge, but can assign reasonable probabilities so as to be less than completely in the dark. This theorem shows that we get hard failures if we try for pi_1 convergence.

How concerned should we be? Some of the "hard failures" here correspond to the necessary failures in probabilistic reflection. These actually seem quite tolerable. There could be a lot more errors than that, though.

One fruitful idea might be to weaken the coherence requirement. The usual argument for coherence  is the dutch book argument; but this makes the assumption that bets will pay off, which does not apply here, since we may never face the truth or falsehood of certain mathematical statements. Intuitionistic probability comes out of a variation of the dutch book argument for the case when bets not paying off at all is a possible outcome. This does not require that probabilities sum to 1, which means we can have a gap between the probability of X and the probability of not-X.

An extreme version of this was proposed by Marcello Herreschoff at the MIRI workshop; he suggested that we can get Pi_1 convergence by only sampling Pi_1 statements. This gets what we want, but results in probability gaps at higher levels in the hierarchy; it's possible that a sampled theory will never prove or disprove some complicated statements. (This is similar to the intuitionistic probability idea, but doesn't actually satisfy the intuitionistic coherence requirements. I haven't worked this out, though, so take what I'm saying with a grain of salt.)

We may even preserve some degree of probabilistic reflection this way, since the true Pi_1 still imply the true Sigma_2.

That particular approach seems rather extreme; perhaps too limiting. The general idea, though, may be promising: we may be able to get the advantages of Pi_1 convergence without the disadvantages.

*(Source: Last paragraph of this section on wikipedia.)

Median Expectation & Robust Utility

I've been talking recently about robust statistics, and the consequences of replacing means with medians. However, I've only looked at this in a fairly limited way, asking about one particular distribution (the bell curve). Mean values are everywhere in statistics; perhaps to a greater degree than you realize, because we often refer to the mean value as the "expected value". It's a simple alias for the same thing, but that may be easy to forget when we are taking expectations everywhere.

In some sense, the "expectation" seems to be a more basic concept than the "mean". We could think of the mean as simply one way of formalizing the intuitive notion of expected value. What happens if we choose a different formalization? What if we choose the median?

The post on altering the bell curve is (more or less) an exploration of what happens to some of classical statistics if we do this. What happens to Bayesian theory?

The foundations of Bayesian statistics are really not touched at all by this. A Bayesian does not rely as heavily on "statistics" in the way a frequentist statistician does. A statistic is a number derived from a dataset which gives some sort of partial summary. We can look at mean, variance, and higher moments; correlations; and so on. We distinguish between the sample statistic (the number derived from the data at hand) and the population statistic (the "true" statistic which we could compute if we had all the examples, ever, of the phenomenon we are looking at). We want to estimate the population statistics, so we talk about estimators; these are numbers derived from the data which are supposed to be similar to the true values. Unbiased estimators are an important concept: ways of estimating population statistics whose expected values are exactly the population statistics.

These concepts are not exactly discarded by Bayesians, since they may be useful approximations. However, to a Bayesian, a distribution is a more central object. A statistic may be a misleading partial summary. The mean (/mode/median) is sort of meaningless when a distribution is multimodal. Correlation does not imply... much of anything (because it assumes a linear model!). Bayesian statistics still has distribution parameters, which are directly related to population statistics, but frequentist "estimators" are not fundamental because they only provide point estimates. Fundamentally, it makes more sense to keep a distribution over the possibilities, assigning some probability to each option.

However, there is one area of Bayesian thought where expected value makes a great deal of difference: Bayesian utility theory. The basic law of utility theory is that we choose actions so as to maximize expected value. Changing the definition of "expected" would change everything! The current idea is that in order to judge between different actions (or plans, policies, designs, et cetera) we look at the average utility achieved with each option, according to our probability distribution over the possible results. What if we computed the median utility rather than the average? Let's call this "robust utility theory".

From the usual perspective, robust utility would perform worse: to the extent that we take different actions, we would get a lower average utility. This begs the question of whether we care about average utility or median utility, though. If we are happy to maximize median utility, then we can similarly say that the average-utility maximizers are performing poorly by our standards.

At first, it might not be obvious that the median is well-defined for this purpose. The median value coming from a probability distribution is defined to be the median in the limit of infinite independent samples from that distribution, though. Each case will contribute instances in proportion to its probability. What we end up doing is lining up all the possible consequences of our choice in order of utility, with a "width" determined by the probability of each, and taking the utility value of whatever consequence ends up in the middle. So long as we are willing to break ties somehow (as is usually needed with the median), it is actually well-defined more often than the mean! We avoid problems with infinite expected value. (Suppose I charge you to play a game where I start with a $1 pot, and start flipping a coin. I triple the pot every time I get heads. Tails ends the game, and I give you the pot. Money is all you care about. How much should you be willing to pay to play?)

Since the median is more robust than the mean, we also avoid problems dealing with small-probability but disproportionately high-utility events. The typical example is Pascal's Mugging. Pascal walks up to you and says that if you don't give him your wallet, God will torture you forever in hell. Before you object, he says: "Wait, wait. I know what you are thinking. My story doesn't sound very plausible. But I've just invented probability theory, and let me tell you something! You have to evaluate the expected value of an action by considering the average payoff. You multiply the probability of each case by its utility. If I'm right, then you could have an infinitely large negative payoff by ignoring me. That means that no matter how small the probability of my story, so long as it is above zero, you should give me your wallet just in case!"

A Robust Utility Theorist avoids this conclusion, because small-probability events have a correspondingly small effect on the end result, no matter how high a utility we assign.

Now, a lot of nice results (such as the representation theorem) have been derived for average utilities over the years. Naturally, taking a median utility might do all kinds of violence to these basic ideas in utility theory. I'm not sure how it would all play out. It's interesting to think about, though.

Philosopher's Carnival #148

Hi all! This month, I have the honor of hosting the monthly blog review, Philosopher's Carnival. This reviews some of the best philosophy postings of the previous month.

Two Metaphysical Pictures, by Richard Yetter Chappell of Philosophy et cetera, outlines a broad classification of metaphysical theories into two different types. (For the record, I prefer the second type! However, as one comment rightly points out, we should avoid lumping views together to create false dichotomies in this way.)

Special relativity and the A-theory, at Alexander Pruss's Blog, discusses the relationship between the philosophical view of time and what we know from physics. (This is related to the two views of the previous post, at least to the extent that you buy the false dichotomy.)

A Question About Religious Experience and Safety Accounts of Knowledge, by ex-apologist, discusses the relationship between the safety condition for knowledge and Christian epistemology.

Metaphysical Skepticism a la Kriegel, by Eric Schwitzgebel of The Splintered Mind, reviews a paper by Uriah Kriegel which suggests that although there may be meaningful metaphysical questions about which there are true and false answers, we cannot arrive at knowledge of those answers by any means. In particular, there is no means by which we can come to know if sets of objects have a literal existence or are merely mental constructs.

A pair of posts by Jeffrey Ketland of M-Phi discuss the Quine-Putnam Indispensability Argument: The Quine-Putnam Indispensability Argument and Other Formulations of the Quine-Putnam Indispensability Argument. The indispensability argument also deals with the question of whether sets (and other mathematical constructions) have a literal existence. The idea is that our best scientific theories make use of sets of objects (and other mathematical constructs), so we must either accept their existence or reject our best science.

Grim Reapers vs. Uncaused Beginnings, by Joshua Rasmussen of Prosblogion, gives a discussion of some "grim reaper" arguments. The Grim Reaper argument is an argument which is supposed to show the implausibility of an infinite past. Joshua shows that a very similar argument would conclude that a finite past with an uncaused beginning is equally implausible.

A Modification to Lewis's Theory of Counterfactuals, by Tristan Haze of Sprachlogik, questions the role that "similarity" of possible worlds should play in our evaluation of counterfactuals.

Computational Metaphysics, by Tomkow, provides a metaphysical companion to computational physics. The idea is illustrated by giving a computational-metaphysics account of counterfactuals, including Lewis's "similarity".

Substitution and Models, Part 1: Bolzano, Quine, Tarski and Boolos, by Jason of Metaphysical Values, reviews the debate between substitution-based understanding of quantifiers and model-theoretic accounts (in preparation for a series of posts about the issue).

That's it for this month! Tune in for the next carnival at http://blog.kennypearce.net/, March 10. If you spy any interesting philosophy articles, submit them for the carnival!

Weird Curves

Last time, I mentioned a question:

Taleb and others have mentioned that the bell curve (or Gaussian) does not deal with outliers well; it gives them a very small probability, and the parameter estimates end up being highly dependent on them.

Yet, one of the justifications of the Gaussian is that it's the max-entropy curve for a given mean and standard deviation.

Entropy is supposed to be a measure of the uncertainty associated with a distribution; so, shouldn't we expect that the max-entropy distribution would give as high a probability to outliers as possible?

There are several answers.

First: a basic problem is that phrase "a given mean and standard deviation". In particular, to choose a standard deviation is to choose an acceptable range for outliers (in some sense). If we have uncertainty about the standard deviation, it turns out the resulting curve has a polynomial decay rather than an exponential one! (This means distant outliers are far more probable.) Essentially, estimating deviations from data (using maximum-likelihood) makes us extremely overconfident that the data will fall in the range we've experienced before. A little Bayesian uncertainty (which still estimates from data, but admits a range of possibilities) turns out to be much less problematic in this respect.

This is definitely helpful, but doesn't solve the riddle: it still feels strange, that the max-entropy distribution would have such a sharp (super-eponential!) decay rate. Why is that?

My derivation will be somewhat heuristic, but I felt that I understood "why" much better by working it out this way than by following other proofs I found (which tend to start with the normal and show that no other distribution has greater entropy, rather than starting with desired features and deriving the normal). [Also, sorry for the poor math notation...]

First, let's pretend we have a discrete distribution over n points, x_n. The result will apply no matter how many points we have, which means it applies in the limit of a continuous distribution. Continuous entropy is not the limit of discrete entropy, so I won't actually be maximising discrete entropy here; I'll maximise the discrete version of the continuous entropy formula:

f(x) maximising sum_i: f(x_i)log(f(x_i))

Next, we constrain the distribution to sum to a constant, have a constant mean, and have constant variance (which also makes the standard deviation constant):

sum_i[f(x_i)] = C₁
sum_i[x_if(x_i)] = C₂
sum_i[x_i²f(x_i)] = C₃

To solve the constrained optimisation problem, we make lagrange multipliers for the constraints:

Lagrangian:
f(x_i)log(f(x_i))
- lambda₁*(sum_i[f(x_i)]-C₁)
- lambda₂*(sum_i[x_if(x_i)]-C₂)
- lambda₃*(sum_i[x_i²f(x_i)]-C₃)

Partial derivatives in the f(x_i):
1 - log(f(x_i))
- lambda₁
- lambda₂*x_i
- lambda₃*x_i²

Setting this equal to zero and solving for f(x_i):
f(x_i) = 2^{1 - lambda₁ - lambda₂*x_i - lambda₃*x_i2}

That's exactly the form of the Gaussian: a constant to the power of a 2nd-degree polynomial!

So, we can see where everything comes from: the exponential comes from our definition of entropy, and the function within the exponent comes from the Lagrange multipliers. The Gaussian is quadratic precisely because we chose a quadratic loss function! We can get basically any form we want by choosing a different loss function. If we use the kurtosis rather than the variance, we will get a fourth degree polynomial rather than a second degree one. If we choose an exponential function, we can get a doubly exponential probability distribution. And so on. There should be some limitations, but more or less, we can get any probability distribution we want, and claim that it is justified as the maximum-entropy distribution (fixing some measure of spread). We can even get rid of the exponential by putting a logarithm around our loss function.

Last time, I mentioned robust statistics, which attempts to make statistical techniques less sensitive to outliers. Rather than using the mean, robust statistics recommends using the median: whereas a sufficiently large outlier can shift the mean by an arbitrary amount, a single outlier has the same limited effect on the median no matter how extreme its value.

I also mentioned that it seems more intuitive to use the absolute deviation, rather than the squared deviation.

If we fix the absolute deviation and ask for the maximum-entropy function, we get something like e^-|x| as our distribution. This is an ugly little function, but the maximum-likelihood estimate of the center of the distribution is precisely the median! e^-|x| justifies the strategy of robust statistics, reducing sensitivity to outliers by making extreme outliers more probable. (The reason is: the max-likelihood estimate will be the point which minimizes the sum of the loss functions centred at each data point. The derivative at x is equal to the number of data points below x minus the number above x. Therefore the derivative is only zero when these two are equal. This is the minimum loss.)

What matters, of course, is not what nice theoretical properties a distribution may have, but how well it matches the true situation. Still, I find it very interesting that we can construct a distribution which justifies taking the median rather than the mean... and I think it's important to show how arbitrary the status of the bell curve as the maximum entropy distribution is.

Just to be clear: the Bayesian solution is not usually to think too much about what distributions might have the best properties. This is important, but when in doubt, we can simply take a mixture distribution over as large a class of hypotheses as we practically can. Bayesian updates give us nice convergence to the best option, while also avoiding overconfidence (for example, the asymptotic probability of outliers will be on the order of the most outlier-favouring distribution present).

Still, a complex machine learning algorithm may still need a simple one as a sub-algorithm to perform simple tasks; a genetic programming approximation of Bayes may need simpler statistical tools to make an estimation of distribution algorithm work. More generally, when humans build models, they tend to compose known distributions such as Gaussians to make them work. In such cases, it's interesting to ask whether classical or robust statistics is more appropriate.