Course Announcement: Data Mining (36-462/662), Fall 2019

For the first time in ten years, I find myself teaching data mining in the fall. This means I need to figure out what data mining is in 2019. Naturally, my first stab at a syllabus is based on what I thought data mining was in 2009. Perhaps it's changed too little; nonetheless, I'm feeling OK with it at the moment*. I am sure the thoughtful and constructive suggestions of the Internet will only reinforce this satisfaction.

--- Seriously, suggestions are welcome, except for suggesting that I teach about neural networks, which I deliberately omitted because I am an out-of-date stick-in-the-mud reasons**.

*: Though I am not done selecting readings from the textbook, the recommended books, and sundry articles --- those will however come before the respective classes. I have been teaching long enough to realize that most students, particularly in a class like this, will read just enough of the most emphatically required material to think they know how to do the assignments, but there are exceptions, and anecdotally even some of thoe majority come back to the material later, and benefit from pointers. ^

**: On the one hand, CMU (now) has plenty of well-attended classes on neural networks and deep learning, so what would one more add? On the other, my admittedly cranky opinion is that we have no idea why the new crop works better than the 1990s version, and it's not always clear that they do work better than good old-fashioned machine learning, so there.

Corrupting the Young; Enigmas of Chance

Notes on "Intriguing Properties of Neural Networks", and two other papers (2014)

\[ \DeclareMathOperator*{\argmax}{argmax} \]

Attention conservation notice: Slides full of bullet points are never good reading; why would you force yourself to read painfully obsolete slides (including even more painfully dated jokes) about a rapidly moving subject?

These are basically the slides I presented at CMU's Statistical Machine Learning Reading Group on 13 November 2014, on the first paper on what have come to be called "adversarial examples". It includes some notes I made after the group meeting on the Q-and-A, but I may not have properly credited (or understood) everyone's contributions even at the time. It also includes some even rougher notes about two relevant papers that came out the next month. Presented now because I'm procrastinating preparing for my fall class in the interest of the historical record.

Paper I: "Intriguing properties of neural networks" (Szegedy et al.)

Background

• Nostalgia for the early 1990s: G. Hinton and company are poised to take over the world, NIPS is mad for neural networks, Clinton is running for President...
  • Learning about neural networks for the first time in cog. sci. 1
  • Apocrypha: a neural network supposed to distinguish tanks from trucks in aerial photographs actually learned about parking lots...
• The models
  • Multilayer perceptron \[ \phi(x) = (\phi_K \circ \phi_{K-1} ... \circ \phi_1)(x) \]
  • This paper not concerned with training protocol, just following what others have done
• The applications
  • MNIST digit-recognition
  • ImageNet
  • \(10^7\) images from YouTube
  • So we've got autoencoders, we've got convolutinal networks, we've got your favorite architecture and way of training it

Where Are the Semantics?

• Claim (the literature, passim): individual hidden units in the network encode high-level semantic features
• Support for the claim: look at the images which maximize the activation of units in some layer \[ \mathcal{X}_i = \argmax_{x}{\langle \phi(x), e_i \rangle} \] then do story-telling about what \(x \in \mathcal{X}_i\) have in common

• The critique: pick a random unit vector \(v\) and do similar story-telling about \[ \mathcal{X}_{v} = \argmax_{x}{\langle \phi(x), v \rangle} \]

• Real units
  • Comment on the top left: white flowers?!?

• Randomized pseudo-units

• My assessment of the critique:
  • Weakness of both claim and critique: people are very good at finding semantic features that link random objects (e.g., Zhu, Rogers and Gibson, "Human Rademacher Complexity", NIPS 2009 where random word lists come up with semantics like "related to motel service")
  • How much semantics would people be able to read into a random collection of training \(x\)'s of equal size to \(\mathcal{X}_i\) or \(\mathcal{X}_v\)?
• Implications of the critique: This is if anything a vindication for good old fashioned parallel distributed processing (as in the 1990s...)
  • Doesn't matter as engineering...
  • Also doesn't matter as a caricature of animal nervous systems: just as there are no grandmother cells in the brain, there is no "white flower" cell in the network, that's actually distributed across the network
• Suggestions from the audience:
  • Ryan Tibshirani: maybe the basis vectors are more "prototypical" than the random vectors? Referenced papers by Nina Baclan and by Robert Tibs. on prototype clustering
  • Yu-Xing Wang: What about images corresponding to weights of hidden-layer neurons in convolutional networks? Me: I need to see what's going on in that paper before I can comment...

The Learned Classifier Isn't Perceptually Continuous

• Claim: generalization based on semantics, or at least on features not local in the input space
• "Adversarial examples"
  • Find the smallest perturbation \(r\) we can apply to a given \(x\) to drive it to the desired class \(l\), i.e., smallest \(r\) s.t. \(f(x+r) = l\)
  • Robustness: use the same perturbation on a different network (different hyperparameters, different training set) and see whether \(f^{\prime}(x+r) = l\) as well
• Some adversarial images
  • "All images in the right column are predicted to be an"ostrich, Struthio camelus"
  • "The examples are strictly randomly chosen. There is not any postselection involved."

• Some more adversarial images: car vs. not-car

• Some adversarial digits
  • RMS distortion \(0.063\) on a 0-1 scale, accuracy 0

• Some non-adversarial digits
  • Gaussian white noise, RMS distortion \(0.1\), accuracy 51% (!)
  • \(\therefore\) it's not just the magnitude of the noise, it's something about the structure

• Assessing the adversarial examples
  • Obviously all adversial examples are mis-classified on the \(f\) they're built for
  • But much higher mis-classification rates on other networks as well
  • Much higher mis-classification rates than much bigger Gaussian noise
  • To be fair, what would adversarial examples look like for any other classifier? For any other classifier using the kernel trick? Someone should try this for a support vector machine
• Suggestions from the audience:
  • Ryan: stability somehow? Changing one data point (say one dog for the adversarial dog) must be really changing the function; me: but isn't stability about the stability of the risk (in R^1), not the stability of the classifier (as a point in function space)? Risk averages over the measure again, smoother than the function as such. Martin Azizyan: so even if you always get wrong a dense subset, it wouldn't affect your risk. Me: right, though good luck proving that there is a countable dense subset of adversarial dogs. Ryan: senses of stability ought to be related somehow? Me, after: maybe, see London et al.
  • David Choi: might this be some sort of generic phenomenon in very high dimensions? Since most of the volume of a closed body is very close to its surface, unless the geometric margin is very large, most points will be very close in the input space to crossing some decision boundary

How Can This Be?

• The networks aren't over-fitting in any obvious sense
  • CV shows they generalize to new instances from the same training pool
• The paper suggests looking at "instability", i.e., the Lipschitz constant of the \(\phi\) mapping
  • They can only upper-bound this
  • And frankly it's not very persuasive as an answer (Lipschitz constant is a global property not a local one)
• Speculative thought 1: what it is like to be an autoencoder?
  • Adversarial examples are perceptually indistinguishable for humans but not for the networks
  • \(\therefore\) human perceptual feature space is very different from the network's feature space
  • What do adversarial examples look like for humans? (Possible psych. experiment with Mechanical Turk)
• Speculative thoguht 2: "be careful what you wish for"
  • Good generalization to the distribution generating instances
  • This is \(n^{-1}\sum_{i=1}^{n}{\delta(x-x_i)}\), and \(n = 10^7 \ll\) dimension of the input space
  • Big gaps around every point in the support
  • \(\therefore\) training to generalize to this distribution doesn't care about small-scale continuity...
  • IOW, the network is doing exactly what it's designed to do
  • Do we need lots more than \(10^7\) images? How many does a baby see by the time it's a year old, anyway?
  • What if we added absolutely-continuous noise to every image every time it was used? (Not enough, they use Gaussians)
  • They do better when they fed in adversial examples into the training process (natch), but not clear whether that's not just shifting around the adversarial examples...

Paper II: "Deep Neural Networks are Easily Fooled: High Confidence Predictions for Unrecognizable Images" (Nguyen, Yosinski and Clune)

• Szegedy et al. looked for minimally-small perturbations which changed the predicted class
• Nguyen et al. look for the largest possible perturbations which leave the class alone: "it is easy to produce images that are completely unrecognizable to humans, but that state-of-the-art DNNs believe to be recognizable objects with 99.99% confidence (e.g. labeling with certainty that white noise static is a lion)."
  • Used evolutionary algorithms to find these examples, either "directly encoded" (genome = the pixels) or "indirectly encoded" (genome = a picture-producing network, with more structure in the image).
  • Also used gradient ascent

"Evolved images that are unrecognizable to humans, but that state-of-the-art DNNs trained on ImageNet believe with \(\geq 99.6\)% certainty to be a familiar object. ... Images are either directly (top) or indirectly (bottom) encoded."

• As you can tell from that figure, the evolved images look nothing at all like what human beings would think of as examples of those classes.

• Results for MNIST are similar: in each plot, we see images classified (with \(\geq 99.9\)% confidence) as the digits given by the columns; the rows are 5 different evolutionary runs, with either direct (white-noise-y panel) or indirect (abstract-art-y panel) encodings.

• Paper I says that imperceptible-to-humans perturbations make huge differences to neural networks; paper II says that huge-to-human perturbations can leave neural networks utterly indifferent. And this is, N.B., on networks that do well under cross-validation on the data.
• Conclusion: Clearly human iso-classification¹ contours are nothing at all like DNN iso-classification contours.
• It's entirely possible that any method which works with a high-dimensional feature space is vulnerable to similar attacks. I'm not sure how we'd do this with human beings as the classifiers (some sort of physiological measure of bogglement in place of the network's numerical confidence?), but it'd be in-principle straightforward enough to d with a common-or-garden support vector machine.
  • Desimone et al. (1984), in a now-classic² paper in neuroscience, demonstrated that macaques had neurons which selectively responded to faces. They went on to show that (at least some) of those cells still responded to quite creepy altered faces, with their features re-arranged, blanking bars drawn across them, and even faces of a different species (as in this portion of their Figure 6A):

It'd seem in-principle possible to use the technique of this paper to evolve images for maximal response from these cells, if you could get the monkey to tolerate the apparatus for long enough.

Paper III: "Visual Causal Feature Learning" (Chalupka, Perona and Eberhardt)

• The basic idea: defining what microscopic, low-level features cause an image to have macroscopic, high-level properties
  • This is a counterfactual / interventional definition, not just about probabilistic association
  • This gives an in-principle sound way of constraining a learning system to only pay attention to the causally relevant features
• With observational data, you'd need to do causal inference
  • They have a very clever way of doing that here, based on a coarsening theorem, which says that the correct causal model is (almost always) a coarsening of the optimally-predictive observational model
  • This in turn gets into issues of predictive states...

1. I guess a more purely Greek phrase would be "isotaxon", but that's just a guess.^

2. i.e., I learned about this paper from Shallice and Cooper's excellent The Organisation of Mind, but felt dumb for not knowing about it before.^

"Causal inference in social networks: A new hope?" (Friday at the Ann Arbor Statistics Seminar)

Attention conservation notice: Self-promoting notice of a very academic talk, at a university far from you, on a very recondite topic, solving a problem that doesn't concern you under a set of assumptions you don't understand, and wouldn't believe if I explained to you.

I seem to be giving talks again:

"Causal inference in social networks: A new hope?"

Abstract: Latent homophily generally makes it impossible to identify contagion or influence effects from observations on social networks. Sometimes, however, homophily also makes it possible to accurately infer nodes' latent attributes from their position in the larger network. I will lay out some assumptions on the network-growth process under which such inferences are good enough that they enable consistent and asymptotically unbiased estimates of the strength of social influence. Time permitting, I will also discuss the prospects for tracing out the "identification possibility frontier" for social contagion.

Joint work with Edward McFowland III

Time and place: 11:30 am -- 12:30 pm on 8 February 2019, in 411 West Hall, Statistics Department, University of Michigan

--- The underlying paper grows out of an idea that was in my paper with Andrew Thomas on social contagion: latent homophily is the problem with causal inference in social networks, but latent homophily also leads to large-scale structure in networks, and allows us to infer latent attributes from the graph; we call this "community discovery". Some years later, my student Hannah Worrall, in her senior thesis, did an extensive series of simulations showing that controlling for estimated community membership lets us infer the strength of social inference, in regimes where community-discovery is consistent. Some years after that, Ed asked me what I was wanting to work on, but wasn't, so I explained about what seemed to me the difficulties in doing some proper theory about this. As I did so, the difficulties dissolved under Ed's questioning, and the paper followed very naturally. We're now revising in reply to referees (Ed, if you're reading this --- I really am working on it!), which is as pleasant as always. But I am very pleased to have finally made a positive contribution to a problem which has occupied me for many years.

Constant Conjunction Necessary Connexion; Enigmas of Chance; Networks; Self-Centered

On Godzilla and the Nature and Conditions of Cultural Success; or, Shedding the Skin

Attention conservation notice: 1100+ words of Deep Thoughts on a creature-feature monster and cultural selection, from someone with no qualifications to write on either subject. Expresses long-held semi-crank notions; composed while simultaneously reading Morin on diffusion chains and drinking sake; revived over a year after it was drafted because Henry was posting about similar themes, finally posted because I am procrasting finishing a grant proposal celebrating submitting a grant proposal on time.

Godzilla is an outstanding example of large-scale cultural success, and of how successful cultural items become detached from their original meanings.

Godzilla's origins are very much in a particular time and place, namely Japan, recently (if not quite immediately) post-WWII and the national trauma of the atomic bombings and their lingering effects. This is a very particular setting, on the world-historical scale. It is now seven decades in the past, and so increasingly gone from living memory, even for the very long-lived population of Japan.

Against this, Godzilla has been tremendously successful culturally all over the world, over basically the whole time since it appeared. I don't mean that it's made money (thought it has) --- I mean that it has been popular, that people have liked consuming stories (and images and toys and other representations) about it, that they have liked creating such representations, and that they have liked thinking about and with Godzilla.. (In contemporary America, for instance, Godzilla is so successful that the suffix "-zilla" is a morpheme, denoting something like "a destructive, mindlessly-enraged form of an entity".) Necessarily, the vast majority of this success and popularity has been distant in time, space, social structure and cultural context from 1950s Japan. How can these two observations --- the specificity of origins and the generality of success --- be reconciled?

To a disturbing extent, of course, any form of cultural success can be self-reinforcing (cf. Salganik et al.), but there is generally something to the representations which succeed (cf., again, Salganik et al.). But, again, Godzilla is endemic in many contexts remote in space, time and other cultural features from immediately-post-war Japan. So it would seem that whatever makes it successful in those contexts, including here and now as I write this, must be different from what made it successful at its point of origin.

It could be that Godzilla is successful in 1950s Japan and in 2010s USA because it happened to fit two very different but very specific cultural niches --- the trauma of defeat culminating in nuclear war, on the one hand; and (to make something up) a compulsive desire for re-enactments of 9/11 on the other hand. But explaining wide-spread success by a series of particular fits falters as we consider all the many other social contexts in which Godzilla has been popular. Maybe it happened, by chance, to appeal narrowly to one new context, but two? three? ten?

An alternative is that Godzilla has managed to spread because it appeals to tastes which are not very context-specific, but on the contrary very widely distributed, if not necessarily constant and universal. In the case of Godzilla, we have a monster who breaks big things and breathes fire: an object of thought, in other words, enduringly relevant to crude interests in predators, in destruction, and in fire. Since those interests are very common across all social contexts, something which appeals to them has a very good source of "pull".

This is not to say that Godzilla wasn't, originally, all about being the only country ever atom-bombed into submission. But it is to say that we can draw a useful distinction between the meanings successful cultural products had originally and those attached to them as they diffuse. It is analogous to the distinction the old philosophy of science used to draw between an idea's "context of discovery" and its "context of justification", though that had a normative force I am not aiming at. (For the record, I think that many of the criticisms of the discovery-justification distinction are weak, mis-conceived or just flat wrong, and that it's actually a pretty useful distinction. But that's another story for another time.)

For Godzilla, like many other successful cultural products, the "context of invention" was a very historically-specific confluence of issues, concerns and predecessors. But the "context of diffusion" was that it could appeal to vastly more generic tastes, and make use of vastly more generic opportunities. These are still somewhat historically-specific (e.g., no motion-picture technology, no Godzilla), but much less so. I am even tempted to formulate a generalization: the more diffused a cultural product is, in space or time or social position, the less its appeal owes to historically-specific contexts, and the more it owes to forces which are nearly a-historical and constant.

What holds me back from declaring cultural diffusion to be a low-pass filter is that it is, in fact, logically possible for a cultural product to succeed in many contexts because it seems to be narrowly tailored to them all. What's needed, as a kind of meta-ingredient, is for the cultural product to be suggestively ambiguous. It is ambiguity which allows very different people to find in the same artifact the divergent but specific meanings they seek; but it also has to somehow suggest to many people that there is a specific, compelling meaning to be found in it. When we consider cultural items which have endured for a very long time, like some sacred texts or other works of literature, then I suspect we are seeing representations which have been strongly selected for suggestive ambiguity.

It is a cliche of literary criticism that each generation gives its own interpretation of these great works. It is somewhat less of a cliche, though equally true, that every generation finds a reason to interpret them. Pace Derrida and his kin, I don't think that every text or artifact is equally amenable to this sort of re-interpretation and re-working. (Though that notion may have seemed more plausible to literary scholars who were most familiar with a canon of books inadvertently selected, in part, for just such ambiguity.) There are levels of ambiguity, and some things are just too straightforward to succeed this way¹. It is also plainly not enough just to be ambiguous, since ambiguous representations are very common, and usually dismal failures at propagating themselves. The text or artifact must also possess features which suggest that there is an important meaning to be found in it². What those features are, in terms of rhetorical or other sorts of design, is a nice question, though perhaps not beyond all conjecture. (I strongly suspect Gene Wolfe of deliberately aiming for such effects.) Something keeps the great works alive over time and space, saving them from being as dead as Gilgamesh, of merely historical interest. Because they are interpreted so variously, they can't be surviving because any one of their interpretations is the right one, conveying a compelling message that assures human interest. Rather, works outlast ages precisely because they simultaneously promise and lack such messages. This quality of suggestive ambiguity could, of course, also contribute to academic and intellectual success --- making it seem like you have something important to say, while leaving what that thing is open to debate, is one route to keeping people talking about you for a long time.

… or so I think in my more extreme moments. In another mood, I might try to poke holes in my own arguments. As for Godzilla, I suspect it's too early to tell whether it possesses this quality of suggestive ambiguity, but my hunch is that this dragon is not a shape-shifter.

The Collective Use and Evolution of Concepts; Scientifiction and Fantastica

Data Over Space and Time

Collecting posts related to this course (36-3467/36-667).

• Fall 2018:
  • Course Announcement
  • Lecture 1: Introduction to the Course
  • Lectures 2 and 3: Smoothing, Trends, Detrending
  • Lecture 4: Principal Components Analysis I
  • Lecture 5: Principal Components Analysis II
  • Lecture 6: Optimal Linear Prediction
  • Lecture 7: Linear Prediction for Time Series
  • Lecture 8: Linear Prediction for Spatial and Spatio-Temporal Random Fields
  • Lectures 9--13: Filtering, Fourier Analysis, African Population and Slavery, Linear Generative Models
  • Lectures 14 and 15: Inference for Dependent Data
  • Lecture 17: Simulation
  • Lectures 18 and 19: Simulation for Inference
  • Lecture 20: Markov Chains
  • Lectures 21--24: Compartment Models, Optimal Prediction, Inference for Markov Models, Markov Random Fields, Hidden Markov Models
  • Self-Evaluation and Lessons Learned
  • Books to Read While the Algae Grow in Your Fur, December 2018
  • Books to Read While the Algae Grow in Your Fur, January 2019

Books to Read While the Algae Grow in Your Fur, January 2019

Attention conservation notice: I have no taste. I also have no qualifications to discuss the history of millenarianism, or really even statistical graphics.

Bärbel Finkenstädt, Leonhard Held and Valerie Isham (eds.), Statistical Methods for Spatio-Temporal Systems

This is an edited volume arising from a conference, with all the virtues and vices that implies. (Several chapters have references to the papers which first published the work expounded in other chapters.) I will, accordingly, review the chapters in order.

Chapter 1: "Spatio-Temporal Point Processes: Methods and Applications" (Diggle). Mostly a precis of case studies from Diggle's (deservedly standard) books on the subject, which I will get around to finishing one of these years.

Chapter 2: "Spatio-Temporal Modelling --- with a View to Biological Growth" (Vedel Jensen, Jónsdóttir, Schmiegel, and Barndorff-Nielsen). This chapter divides into two parts. One is about "ambit stochastics". In a random field $Z(s,t)$, the "ambit" of the space-time point-instant $(s,t)$ is the set of point-instants $(q,u)$, $u < t$, where $Z(q,u)$ is (causally) relevant to $Z(r,t)$. (This is what, in my own work, I've called the "past cone" of $(s,t)$.) Having a regular geometry for the ambit imposes some tractable restrictions on random fields, which are explored here for models of growth-without-decay. The second part of this chapter will only make sense to hardened habituees of Levy processes, and perhaps not even to all of them.

Chapter 3: "Using Transforms to Analyze Space-Time Processes" (Fuentes, Guttorp, and Sampson): A very nice survey of Fourier transform, wavelet transform, and PCA approaches to decomposing spatio-temporal data. There's a good account of some tests for non-stationarity, based on the idea that (essentially) we should get the nearly same transforms for different parts of the data if things really are stationary. (I should think carefully about the assumptions and the implied asymptotic regime here, since the argument makes sense, but it also makes sense that sufficiently slow mean-reversion is indistinguishable from non-stationarity.)

Chapter 4: "Geostatistical Space-Time Models, Stationarity, Seperability, and Full Symmetry" (Gneiting, Genton, and Guttorp): "Geostatistics" here refers to "kriging", or using linear prediction on correlated data. As every schoolchild knows, this boils down to finding the covariance function, $\mathrm{Cov}[Z(s_1, t_1), Z(s_2, t_2)]$. This chapter considers three kinds of symmetry restrictions on the covariance functions: "separability", where $\mathrm{Cov}[Z(s_1, t_1), Z(s_2, t_2)] = C_S(s_1, s_2) C_T(t_1, t_2)$; the weaker notion of "full symmetry", where $\mathrm{Cov}[Z(s_1, t_1), Z(s_2, t_2)] = $\mathrm{Cov}[Z(s_1, t_2), Z(s_2, t_1)]$; and "stationarity", where $\mathrm{Cov}[Z(s_1, t_1), Z(s_2, t_2)] = $\mathrm{Cov}[Z(s_1+q, t_1+h), Z(s_2+q, t_2+h)]$. As the authors explain, while separable covariance functions are often used because of their mathematical tractability, they look really weird; "full symmetry" can do a lot of the same work, at less cost in implausibility.

Chapter 5: "Space-Time Modelling of Rainfall for Continuous Simulations" (Chandler, Isham, Belline, Yang and Northrop): A detailed exposition of two models for rainfall, at different spatio-temporal scales, and how they are both motivated by and connected to data. I appreciate their frankness about things that didn't work, and the difficulties of connecting the different models.

Chapter 6, "A Primer on Space-Time Modeling from a Bayesian Perspective" (Higdon): Here "space-time modeling" means "Gaussian Markov random fields". Does what it says on the label.

All the chapters combine theory with examples --- chapter 2 is perhaps the most mathematically sophisticated one, and also the one where the examples do the least work. The most useful, from my point of view, were Chapters 3 and 4, but that's because I was teaching a class where I did a lot of kriging ad PCA, and (with some regret) no point processes. If you have a professional interest in spatio-temporal statistics, and a fair degree of prior acquaintance, I can recommend this as a useful collection of examples, case studies, and expositions of some detailed topics.

Errata, of a sort: There are supposed to be color plates between pages 142 and 143. Unfortunately, in my copy these are printed in grey, not in color.

Disclaimer: The publisher sent me a copy of this book, but that was part of my fee for reviewing a (different) book proposal for them.

Kieran Healy, Data Visualization: A Practical Introduction

Anyone who has looked at my professional writings will have noticed that my data visualizations are neither fancy nor even attractive, and they never go beyond basic R graphics. This is because I have never learned any other system for statistical visualization. And I've not done that because I'm lazy, and have little visual sense anyway. This book is the best guide I've seen to (1) learning the widely-used, and generally handsome, ggplot library in R, (2) learning the "grammar of graphics" principles on which it is based, and (3) learning the underlying psychological principles which make some graphics better or worse visualizations than others. (This is not to be confused with learning the maxims or even the tacit taste of a particular designer, even one of genius.) The writing is great, the examples are interesting, well-chosen and complete, and the presumptions about how much R, or statistics, you know coming in are minimal. I wish something like this had existed long ago, and I'm tempted, after reading it, to totally re-do the figures in my book. (Aside to my editor: I am not going to totally re-do the figures in my book.) I strongly recommend it, and will be urging it on my graduate students for the foreseeable future.

ObLinkage: The book is online, pretty much.

ObDisclaimer: Kieran and I have been saying good things about each other's blogs since the High Bronze Age of the Internet. But I paid good cash money for my copy, and have no stake in the success of this book.

Anna Lee Huber, Mortal Arts

More historical-mystery mind candy, this time flavored by the (dismal) history of early 19th century psychiatry. (Huber is pretty good, though not perfect, at avoiding anachronistic language, so nobody says "psychiatry" in the novel.)

Norman Cohn, The Pursuit of the Millennium: Revolutionary Millenarians and Mystical Anarchists of the Middle Ages

I vividly remember finding a used copy of this in the UW-Madison student bookstore when I began graduate school, in the fall of 1993, and having my mind blown by reading it that fall*. Coming back to it now, I find it still fascinating and convincing, and does an excellent job of tracing millenarian movements among the poor in Latinate Europe from the fall of Rome through the Reformation. (There are a few bits where he gets a bit psychoanalytic, but the first edition was published in 1957.) If I no longer find it mind-blowing, that's in large part because reading it sparked an enduring interest in millenarianism, and so I've long since absorbed what then (you should forgive the expression) came as a revelation.

The most controversial part of the book, I think, is the conclusion, where Cohn makes it very clear that he thinks there is a great deal of similarity, if not actual continuity, between his "revolutionary millenarians and mystical anarchists" and 20th century political extremism, both of the Fascist and the Communist variety. He hesitates --- wisely, I think --- over whether this is just a similarity, or there is an actual thread of historical continuity; but I think his case for the similarity is sound.

*: I was supposed to be having my mind blown by Sakurai. In retrospect, this incident sums up both why I was not a very good graduate student, and why I will never be a great scientist.

Books to Read While the Algae Grow in Your Fur; Enigmas of Chance; Data over Space and Time; Pleasures of Detection, Portraits of Crime; Tales of Our Ancestors; Psychoceramica; Writing for Antiquity Commit a Social Science

Books to Read While the Algae Grow in Your Fur, December 2018

Attention conservation notice: I have no taste. I also have no qualifications to discuss poetry or leftist political theory. I do know something about spatiotemporal data analysis, but you don't care about that.

Gidon Eshel, Spatiotemporal Data Analysis

I assigned this as a textbook in my fall class on data over space and time, because I need something which covered spatiotemporal data analysis, especially principal components analysis, for students who could be taking linear regression at the same time, and was cheap. This met all my requirements.

The book is divided into two parts. Part I is a review or crash course in linear algebra, building up to decomposing square matrices in terms of their eigenvalues and eigenvectors, and then the singular value decomposition of arbitrary matrices. (Some prior acquaintance with linear algebra will help, but not very much is needed.) Part II is about data analysis, covering some basic notions of time series and autocorrelation, linear regression models estimated by least squares, and "empirical orthogonal functions", i.e., principal components analysis, i.e., eigendecomposition of covariance or correlation matrices. As for "cheap", while the list price is (currently) an outrageous \$105, it's on JSTOR, so The Kids had free access to the PDF through the university library.

In retrospect, there were strengths to the book, and some serious weaknesses --- some absolute, some just for my needs.

The most important strength is that Eshel writes like a human being, and not a bloodless textbook. His authorial persona is not (thankfully) much like mine, but it's a likeable and enthusiastic one. This is related to his trying really, really hard to explain everything as simply as possible, and with multitudes of very detailed worked examples. I will probably be assigning Part I of the book, on linear algebra, as refresher material to my undergrads for years.

He is also very good at constantly returning to physical insight to motivate data-analytic procedures. (The highlight of this, for me, was section 9.7 [pp. 185ff] on when and why an autonomous, linear, discrete-time AR(1) or VAR(1) model will arise from a forced, nonlinear, continuous-time dynamical system.) If this had existed when I was a physics undergrad, or starting grad school, I'd have loved it.

Turning to the weaknesses, some of them are, as I said, merely ways in which he didn't write the book to meet my needs. His implied reader is very familiar with physics, and not just the formal, mathematical parts but also the culture (e.g., the delight in complicated compound units of measurement, saying "ensemble" when other disciplines say "distribution" or "population"). In fact, the implied reader is familiar with, or at least learning, climatology. But that reader has basically no experience with statistics, and only a little probability (so that, e.g., they're not familiar with rules for algebra with expectations and covariances*). Since my audience was undergraduate and masters-level statistics students, most of whom had only the haziest memories of high school physics, this was a mis-match.

Others weaknesses are, to my mind, a bit more serious, because they reflect more on the intrinsic content.

• A trivial but real one: the book is printed in black and white, but many figures are (judging by the text) intended to be in color, and are scarcely comprehensible without it. (The first place this really struck me was p. 141 and Figure 9.4, but there were lots of others.) The electronic version is no better.
• The climax of the book (chapter 11) is principal components analysis. This is really, truly important, so it deserves a lot of treatment. But it's not a very satisfying stopping place: what do you do with the principal components once you have them? What about the difference between principal components / empirical orthogonal functions and factor models? (In the book's terms, the former does a low-rank approximation to the sample covariance matrix $\mathbf{v} \approx \mathbf{w}^T \mathbf{w}$, while the latter treats it as low-rank-plus-diagonal-noise $\mathbf{v} \approx \mathbf{w}^T\mathbf{w} + \mathbf{d}$, an importantly different thing.) What about nonlinear methods of dimensionality reduction? My issue isn't so much that the book didn't do everything, as that it didn't give readers even hints of where to look.
• There are places where the book's exposition is not very internally coherent. Chapter 8, on autocorrelation, introduces the topic with an example where $x(t) = s(t) + \epsilon(t)$, for a deterministic signal function $s(t)$ and white noise $\epsilon(t)$. Fair enough; this is a trend-plus-noise representation. But it then switches to modeling the autocorrelations as arising from processes where $x(t) = \int_{-\infty}^{t}{w(u) x(u) du} + \xi(t)$, where again $\xi(t)$ is white noise. (Linear autoregressions are the discrete-time analogs.) These are distinct classes of processes. (Readers will find it character-building to try to craft a memory kernel $w(u)$ which matches the book's running signal-plus-noise example, where $s(t) = e^{-t/120}\cos{\frac{2\pi t}{49}}$.)
• I am all in favor of physicists' heuristic mathematical sloppiness, especially in introductory works, but there are times when it turns into mere confusion. The book persistently conflates time or sample averages with expectation values. The latter are ensemble-level quantities, deterministic functionals of the probability distribution. The former are random variables. Under various laws of large numbers or ergodic theorems, the former converge on the latter, but they are not the same. Eshel knows they are not the same, and sometimes talks about how they are not the same, but the book's notation persistently writes them both as $\langle x \rangle$, and the text sometimes flat-out identifies them. (For one especially painful example among many, p. 185.) Relatedly, the book conflates parameters (again, ensemble-level quantities, functions of the data-generating process) and estimators of those parameters (random variables)
• The treatment of multiple regression is unfortunate. $R^2$ does not measure goodness of fit. (It's not even a measure of how well the regression predicts or explains.) At some level, Eshel knows this, since his recommendation for how to pick regressors is not "maximize $R^2$". On the other hand, his prescription for picking regressors (sec. 9.6.4, pp.180ff) is rather painful to read, and completely at odds with his stated rationale of using regression coefficients to compare alternative explanations (itself a bad, though common, idea). Very strikingly, the terms "cross-validation" and "bootstrap" do not appear in his index**. Now, to be clear, Eshel isn't worse in his treatment of regression that most non-statisticians, and he certainly understands the algebra backwards and forwards. But his advice on the craft of regression is, to be polite, weak and old-fashioned.

Summing up, the linear-algebra refresher/crash-course of Part I is great, and I even like the principal components chapters in Part II, as far as they go. But it's not ideal for my needs, and there are a bunch of ways I think it could be improved for anyone's needs. What to assign instead, I have no idea.

*: This is, I think, why he doesn't explain the calculation of the correlation time and effective sample size in sec. 8.2 (pp. 123--124), just giving a flat statement of the result, though it's really easy to prove with those tools. I do appreciate finally learning the origin of this beautiful and practical result --- G. I. Taylor, "Diffusion by Continuous Movements", Proceedings of the London Mathematical Society, series 2, volume 20 (1922), pp. 196--212 (though the book's citing it with the wrong year, confusing series number with an issue number, and no page numbers was annoying). ^

**: The absence of "ridge regression" and "Tikhonov regularization" from the index is all the more striking because they appear in section 9.3.3 as "a more general, weighted, dual minimization formalism", which, compared to ordinary least squares, is described as "sprinkling added power ... on the diagonal of an otherwise singular problem". This is, of course, a place where it would be really helpful to have a notion of cross-validation, to decide how much to sprinkle.^

Nick Srnicek and Alex Williams, Inventing the Future: Postcapitalism and a World Without Work

It's --- OK, I guess? They have some good points against what they call "folk politics", namely, that it has conspicuously failed to accomplish anything, so doubling down on more of it seems like a bad way to change the world. And they really want to change the world: the old twin goals of increasing human power over the world, and eliminating human power of other humans, are very much still there, though they might not quite adopt that formula. To get there, their basic idea is to push for a "post-work world", one where people don't have to work to survive, because they're entitled to a more-than-subsistence basic income as a matter of right. They realize that making that work will require lots of politics and pushes for certain kinds of technological progress rather than others. This is the future they want --- to finally enter (in Marx's words) "the kingdom of freedom", where we will be able to get on with all the other problems, and possibilities, confronting us.

As for getting there: like a long, long line of leftist intellectuals from the 1960s onwards, Srnicek and Williams are very taken with the idea, going back to Gramsci, that the key to achieving socialism is to first achieve ideological "hegemony". To put it crudely, this means trying to make your idea such broadly-diffused, widely-accepted, scarcely-noticed common notions that when madmen in authority channel voices from the air, they channel you. (In passing: Occupy may have done nothing to reduce economic inequality, but Gramsci's success as a strategist may be measured by the fact that he wrote in a Fascist prison.) Part of this drive for hegemony is pushing for new ideas in economics --- desirable in itself, but they are sure in advance of what inquiry should find *. Beyond this, and saying that many tactics will need to be tried out by a whole "ecology" of organizations and groups, they're pretty vague. There's some wisdom here --- who could propound a detailed plan to get to post-work post-capitalism? --- but also more ambiguity than they acknowledge. Even if a drive for a generous basic income (and all that would go with it) succeeds, the end result might not be anything like the sort of post-capitalism Srniceck and Williams envisage, if only because what we learn and experience along the way might change what seems feasible and desirable. (This is a Popperian point against Utopian plans, but it can be put in other language quite easily**.) I think Srnicek and Williams might be OK with the idea that their desired future won't be realized, so long as some better future is, and that the important point is to get people on the left not to prefigure better worlds in occasional carnivals of defiance, but to try to make them happen. Saying that doing this will require organization, concrete demands, and leadership is pretty sensible, though they do disclaim trying to revive the idea of a vanguard party.

Large portions of the book are, unfortunately, given over to insinuating, without ever quite saying, that post-work is not just desirable and possible, but a historical necessity to which we are impelled by the inexorable development of capitalism, as foreseen by the Prophet. (They also talk about how Marx's actual scenario for how capitalism would develop, and end, not only has not come to pass yet, but is pretty much certain to never come to pass.) Large portions of the book are given over to wide-ranging discussions of lots of important issues, all of which, apparently, they grasp through the medium of books and articles published by small, left-wing presses strongly influenced by post-structuralism --- as it were, the world viewed through the Verso Books catalog. (Perry Anderson had the important advantage, as a writer and thinker, of being formed outside the rather hermetic subculture/genre he helped create; these two are not so lucky.) Now, I recognize that good ideas usually emerge within a community that articulates its own distinctive tradition, so some insularity can be all to the good. In this case, I am not all that far from the authors' tradition, and sympathetic to it. But still, the effect of these two (overlapping) writerly defects is that once the book announced a topic, I often felt I could have written the subsequent passage myself; I was never surprised by what they had to say. Finishing this was a slog.

I came into the book a mere Left Popperian and market socialist, friendly to the idea of a basic income, and came out the same way. My mind was not blown, or even really changed, about anything. But it might encourage some leftist intellectuals to think constructively about the future, which would be good.

Shorter: Read Peter Frase's Four Futures instead.

*: They are quite confident that modern computing lets us have an efficient planned economy, a conclusion they support not be any technical knowledge of the issue but by citations to essays in literary magazines and collections of humanistic scholarship. As I have said before, I wish that were the case, if only because it would be insanely helpful for my own work, but I think that's just wrong. In any case, this is an important point for socialists, since it's very consequential for the kind of socialism we should pursue. It should be treated much more seriously, i.e., rigorously and knowledgeable, than they do. Fortunately, a basic income is entirely compatible with market socialism, as are other measures to ensure that people don't have to sell their labor power in order to live.

**: My own two-minute stab at making chapter 9 of The Open Society and Its Enemies sound suitable for New Left Review: "The aims of the progressive forces, always multifarious, develop dialectically in the course of the struggle to attain them. Those aims can never be limited by the horizon of any abstract, pre-conceived telos, even one designated 'socialism', but will always change and grow through praxis." (I admit "praxis" may be a bit dated.) ^

A. E. Stallings, Like: Poems

Beautiful stuff from one of my favorite contemporary poets. "Swallows" and "Epic Simile" give a fair impression of what you'll find. This also includes a lot of the poems discussed in Cynthia Haven's "Crossing Borders" essay.

Books to Read While the Algae Grow in Your Fur; Enigmas of Chance; Data over Space and Time; The Progressive Forces; The Commonwealth of Letters

Data over Space and Time: Self-Evaluation and Lessons Learned

Attention conservation notice: Academic navel-gazing, about a class you didn't take, in a subject you don't care about, at a university you don't attend.

Well, that went better than it could have, especially since it was the first time I've taught a new undergraduate course since 2011.

Some things that worked well:

1. The over-all choice of methods topics --- combining descriptive/exploratory techniques and generative models and their inference. Avoiding the ARIMA alphabet soup as much as possible both played to my prejudices and avoided interference with a spring course.
2. The over-all kind and range of examples (mostly environmental and social-historical) and the avoidance of finance. I could have done some more economics, and some more neuroscience.
3. The recurrence of linear algebra and eigen-analysis (in smoothing, principal components, linear dynamics, and Markov processes) seems to have helped some students, and at least not hurt the others.
4. The in-class exercises did wonders for attendance. Whether doing the exercises, or that attendance, improved learning is hard to say. Some students specifically praised them in their anonymous feedback, and nobody complained.

Some things did not work so well:

1. I was too often late in posting assignments, and too many of them had typos when first posted. (This was a real issue with the final. To any of the students reading this: my apologies once again.) I also had a lot of trouble calibrating how hard the assignments would be, so the opening problem sets were a lot more work than the later ones.
   (In my partial defense about late assignments, there were multiple problem sets which I never posted, after putting a lot of time into them, because my initial idea either proved much too complicated for this course when fully executed, or because I was, despite much effort, simply unable to reproduce published papers*. Maybe next time, if there is a next time, these efforts can see the light of day.)
2. I let the grading get really, really behind the assignments. (Again, my apologies.)
3. I gave less emphasis to spatial and spatio-temporal models in the second, generative half of the course than they really deserve. E.g., Markov random fields and cellular automata (and kin) probably deserve at least a lecture each, perhaps more.
4. I didn't build in enough time for review in my initial schedule, so I ended up making some painful cuts. (In particular, nonlinear autoregressive models.)
5. My attempt to teach Fourier analysis was a disaster. It needs much more time and preparation than I gave it.
6. We didn't get very much at all into how to think your way through building a new model, as opposed to estimating, simulating, predicting, checking, etc., a given model.
7. I have yet to figure out how to get the students to do the readings before class.

If I got to teach this again, I'd keep the same over-all structure, but re-work all the assignments, and re-think, very carefully, how much time I spent on which topics. Some of these issues would of course go away if there were a second semester to the course, but that's not going to happen.

*: I now somewhat suspect that one of the papers I tried to base an assignment on is just wrong, or at least could not have done the analysis the way it say it did. This is not the first time I've encountered something like this through teaching... ^

Data over Space and Time

Data over Space and Time, Lectures 21--24

Lecture 21: Compartment Models (.Rnw)

Optional Special Topics Lecture: Information Theory and Optimal Prediction

Lecture 22: Inference for Markov Chains (.Rmd, Handout)

Lecture 23: Markov Random Fields (.Rmd)

Lecture 24: Hidden Markov / State-Space Models (.Rmd)

Data over Space and Time

Books to Read While the Algae Grow in Your Fur, November 2018

Attention conservation notice: I have no taste. I also have no qualifications to discuss the history of photography, or of black Pittsburgh.

Cheryl Finley, Laurence Glasco and Joe W. Trotter, with an introduction by Deborah Willis, Teenie Harris, Photographer: Image, Memory, History

A terrific collection of Harris's photos of (primarily) Pittsburgh's black community from the 1930s to the 1970s, with good biographical and historical-contextual essays.

Disclaimer: Prof. Trotter is also on the faculty at CMU, but I don't believe we've ever actually met.

Ben Aaronovitch, Lies Sleeping

Mind candy: the latest installment in the long-running supernatural-procedural mystery series, where the Folly gets tangled up with the Matter of Britain.

Charles Stross, The Labyrinth Index

Mind candy; Latest installment in Stross's long-running Lovecraftian spy-fiction series. I imagine a novel about the US Presidency being taken over by a malevolent occult force seemed a lot more amusing before 2016, when this must have been mostly written. It's a good installment, but only suitable for those already immersed in the story.

Anna Lee Huber, The Anatomist's Wife and A Brush with Shadows

Mind-candy, historical mystery flavor. These are the first and sixth books in the series, because I couldn't lay hands on 2--5, but I will. (Update: More.)

Books to Read While the Algae Grow in Your Fur; Scientifiction and Fantastica; Pleasures of Detection, Portraits of Crime; Tales of Our Ancestors; Cthulhiana; Heard About Pittsburgh, PA

Data over Space and Time, Lecture 20: Markov Chains

(.Rmd)

Data over Space and Time

Course Announcement: Advanced Data Analysis (36-402/36-608), Spring 2019

Attention conservation notice: Announcement of an advanced undergraduate course at a school you don't attend in a subject you don't care about.

I will be teaching 36-402/36-608, Advanced Data Analysis, in the spring.

This will be the seventh time I'll have taught it, since I took it over and re-vamped it in 2011. The biggest change from previous iterations will be in how I'll be handling class-room time, by introducing in-class small-group exercises. I've been doing this in this semester's class, and it seems to at least not be hurting their understanding, so we'll see how well it scales to a class with four or five times as many students.

(The other change is that by the time the class begins in January, the textbook will, inshallah, be in the hands of the publisher. I've finished adding everything I'm going to add, and now it's a matter of cutting stuff, and fixing mistakes.)

Advanced Data Analysis from an Elementary Point of View

Data over Space and Time, Lectures 18 and 19: Simulation for Inference

Lecture 18: The Bootstrap (.Rmd)

Lecture 19: The Method of Simulated Moments and Indirect Inference (.Rmd)

Data over Space and Time

In Memoriam Joyce Fienberg

I met Joyce through her late husband Stephen, my admired and much-missed colleague. I won't pretend that she was a close friend, but she was a friend, and you could hardly hope to meet a kinder or more decent person. A massacre by a deluded bigot would be awful enough even if his victims had been prickly and unpleasant individuals. But that he murdered someone like Joyce --- five blocks from where I live --- makes it especially hard to take. I am too sad to have anything constructive to say, and too angry at living in a running morbid joke to remember her the way she deserves.

Data over Space and Time, Lecture 17: Simulation

Lecture 16 was canceled.

(.Rmd)

Data over Space and Time