# Locklin on science

## AI is not eliminating jobs

Posted in econo-blasphemy, machine learning by Scott Locklin on February 9, 2019

Midwits keep asserting that “AI” is going to eliminate jobs. They say things like “those jobs aren’t coming back because of AI” (or screws or whatever other dumb excuse: but they’re real sure about those jobs not coming back).  These are not statements of scientific or technological fact, or even a reasonable prediction based on present trends. These are ideological political statements. “AI” is a soundbite/fnord excuse for not doing anything about the policy problems of the present.

The ruling caste of American tech and FIRE lizard people continue to make these statements, not because they are inevitable, but because this is their desired future. Their preferred future is a population consisting of powerless, preferably drugged up serfs on the “universal basic income” dole, ruled over by our present ruling class of grifters, rentiers, pyramid scheme salesmen, watched over by a surveillance hellscape.  The lizard people would like to continue our present policy of de-industrializing the country, breaking what little labor negotiating power US citizens have, and atomizing people to their raw protoplasm. It’s almost like a Freudian slip. Don’t bother agitating for any rights, slave, we soon will have electric Golems and won’t need you!

The most murderous drug dealers who ever existed … Google’s AI dude thinks they’re great! https://twitter.com/JeffDean/status/1093953731756867584

Oh I am sure the Google doofs would like to develop and control some strong AI, and perhaps a robot maid to replace Juanita. I, too, would like to have a magic technology which gives me infinite power, and a robot maid to iron my shirts. If I were a Silicon Valley oligarch rather than a humble nerd, I might develop delusions the pile of C++, Javascript and tech drones which made me rich could become a Golem of infinite power. Personally,  I would build rockets. At least I could get away from lizard people who want to turn the world into a soy dystopia.

If these clowns really believed that “AI” were something actually like an “AI” which could replace humans in general tasks, they’d use it to replace computer programmers. At one point in history, people believed CASE tools would eliminate most programmer jobs. How’s that working out for the AI geniuses at Google? They can’t even automate devops jobs; devops being one of the most automatable roles in tech companies. Devops tasks don’t seem much different from a computer strategy game.

Google’s “AI” team can’t do this useful thing, which, even by my lights, actually seems  achievable. Yet somehow, google boobs think they’re going to violate Moravec’s paradox and replace drivers. Think about that for a minute. It’s becoming clear that autonomous vehicle “technology” as sold to people for the last 10 years is basically fraud, and is still stuck in the 1980s when Ernst Dickmanns was driving around the autobahn with Sun Workstations in his back seat. Demonstrations of this tech always have a human in the loop (remote or in vehicle), because moving automobiles without human control are death machines under most circumstances.

Inside of the UniBwM autonomous experimental vehicle VaMP, at the rear bench where the computing system was installed for easy access and monitoring. This was at the PROMETHEUS demonstration in Paris in October 1994 | Photo by Reinhold Behringer

Even assuming I’m wrong and the media hyperbole is right and full level 5 autonomous vehicles are “right around the corner” Google also has zero business interest in “disrupting” driving. Google is a tech driven advertising company with a  collection of loss leaders. Yet they go after this preposterously difficult, possibly impossible task. Why not disrupt a business they presumably know how to disrupt, like that of the lowly ops engineer? At least this would be good for their bottom line, and it would be a real step forward in “AI” rather than a parlour trick perpetuated by marketing nerds and started by obvious mountebanks.

From a semiotics point of view, this shows astounding hostility to the types of people who drive cars and trucks for a living. Drivers are … ordinary, usually uneducated, salt of the earth people who have a fairly independent lifestyle and make a decent living. Google overlords must really hate such people, since they’re dumping all this skrilla into ruining their lives for no sane business reason. They will almost certainly fail, but man, why would you try to blow up those people’s lives? If this country really wanted to get rid of driving, or considered it a serious problem that there are too many cars on the road, or thought that people now employed as drivers should do something else, we had a solution to this problem invented in the late 1800s.

The other professions  people “think” will be replaced always seem to be low caste irritations or lawyers (lol). You regularly hear “experts” talking about how presently common jobs won’t exist in 20 years because of “AI.”  I’ve said multiple times now that all estimates for delivery of something in 20 years are bullshit. A prediction that a technology will do X in 20 years means “we don’t know how to do this, but we want your money to fool around with anyway.” Controlled nuclear fusion researchers being the most amusing case of the perpetual 20 year rice bowl. 20 years is a magic number, as it’s plenty of time for a technological mountebank to retire; and it’s at least 2-3 generations of tenured academics, which is enough to turn a scam subject like “quantum computing” or “nanotech” into an actual field.

“AI” doesn’t exist. Machine learning is a force multiplier and productivity enhancer for statisticians. If you believe the “automation”=”no more jobs” ding dongs, machine learning should have at least automated away the job of statistician. Yet somehow, the  statistician (aka “data scientist”) jobs are among the best paid and most in-demand jobs out there at present.

The last job category I can think of which was automated away is Flight Engineer on airliners. It mostly went away because of automation of airliners, but it wasn’t even computer related; just normal improvements of systems monitoring and reliability; good old mechanical and systems engineering. Despite 1/3 fewer seats in airliner cockpits, there are now more people with airline flight officer jobs now than ever before. Planes got cheaper and there are more of them servicing vastly more people.

The example of Flight Engineer is how the world works. Technological advances increase human power over nature and makes more things possible. Actual “AI” advances, should any eventually materialize, will work exactly like this.

AI has eliminated exactly zero professions, and essentially no jobs. Since the best prediction tool for a market is generally a random walk, my forecast is, barring giant breakthroughs, this trend of “nothing important actually happened” regarding AI job destruction will continue. If you disagree with me and have an alternate prediction on a normal human (aka 5 or 10 year) timescale, I am happy to entertain any long bets on whatever platform you care to use.

## Machine learning & data science: what to worry about in the near future

Posted in machine learning by Scott Locklin on July 9, 2018

Henry Kissinger  recently opined about machine learning. OK, he used the ridiculously overblown phrase “AI” rather than “machine learning” but the latter is what he seemed to be talking about. I’m not a fan of the old reptile, but it is a reasonably thoughtful piece of gaseous bloviation from a politician. Hopefully whoever wrote it for him was well compensated.

There are obvious misapprehensions here; for example, noticing that chess programs are pretty good. You’d expect them to be good by now; we’ve been doing computer chess since 1950. To put this in perspective; steel belted radial tires and transistor radios were invented 3 years after computer chess -we’re pretty good at those as well. It is very much worth noting the first important computer chess paper (Shannon of course) had this sentence in it:

“Although of no practical importance, the question is of theoretical interest, and it is hoped that…this problem will act as a wedge in attacking other problems—of greater significance.”

The reality is, computer chess largely hasn’t been a useful wedge in attacking problems of greater significance.  Kissinger also mentioned Alpha Go; a recent achievement, but it is something which isn’t conceptually much different from TD-Gammon;  done in the 1990s.

Despite all the marketing hype coming out of Mountain View, there really hasn’t been much in the way of conceptual breakthroughs in machine learning since the 1990s.  Improvements in neural networks have caused excitement, and the ability of deep learning to work more efficiently on images is an improvement in capabilities. Stuff like gradient boost machines have also been a considerable technical improvement in usable machine learning. They don’t really count as big conceptual breakthroughs; just normal improvements for a field of engineering that has poor theoretical substructure. As for actual “AI” -almost nobody is really working on this.

None the less, there have been progress in machine learning and data science. I’m betting on some of the improvements having a significant impact on society, particularly now that the information on these techniques is out there and commodified in reasonably decent software packages. Most of these things have not been spoken about by government policy maker types like Kissinger, and are virtually never mentioned in dopey “news” articles on the subject, mostly because nobody bothers asking people who do this for a living.

I’d say most of these things haven’t quite reached the danger point for ordinary people who do not live in totalitarian societies, though national security agency type organizations and megacorps are already using these techniques or could be if they weren’t staffed with dimwits. There are also areas which we are still very bad at, which are to a certain extent keeping us safe.

The real dangers out there are pretty pedestrian looking, but people don’t think through the implications. I keep using the example, but numskull politicians were harping on the dangers of Nanotech about 15 years ago, and nothing came of that either. There were obvious dangerous trends happening in the corporeal world 15 years ago which had nothing to do with nanotech. The obesity rate was an obvious problem back then, whether from chemicals in the environment, the food supply, or the various cocktails of mind altering pharmies that fat people need to get through the day. The US was undergoing a completely uncommented upon and vast demographic, industrial and economic shift. Also, there was an enormous real estate bubble brewing. I almost think numskull politicians talk about bullshit like nanotech to avoid talking about real problems. Similarly politicians and marketers prefer talking about “AI” to issues in data science which may cause real problems in society.

The biggest issue we face has a real world example most people have seen by now. There exists various systems for road toll collection. To replace toll takers, people are encouraged to get radio tags for their car like “ezpass.” Not everyone will have one of these, so government choices are to continue to employ toll takers, removing most of the benefit of having such tools, or use an image recognition system to read license plates, and send people a bill. The technology which underlies this system is pretty much what we’re up against as a society. As should be obvious: not many workers were replaced. Arguably none were; though uneducated toll takers were somewhat replaced by software engineers. The real danger we face from this system isn’t job replacement; it is Orwellian dystopia.

Here is a list of  obvious dangers in “data science” I’m flagging over the next 10-20 years as worth worrying about as a society.

1) Face recognition software  (and to a lesser extent Voice Recognition) is getting quite good. Viola Jones  (a form of boosted machine) is great at picking out faces, and sticking them in classifiers which label them has become routine. Shitbirds like Facebook also have one of the greatest self-owned labeled data sets in the world, and are capable of much evil with it. Governments potentially have very good data sets also. It isn’t quite at the level where we can all be instantly recognized, like, say with those spooky automobile license plate readers, but it’s probably not far away either. Plate readers are a much simpler problem; one theoretically mostly solved in the 90s when Yann LeCun and Leon Bottou developed convolutional nets for ATM machines.

2) Machine learning  and statistics on large data is getting quite respectable. For quite a while I didn’t care that Facebook, google and the advertisers had all my data, because it was too expensive to process it down into something useful enough to say anything about me. That’s no longer true. Once you manage to beat the data cleaning problems, you can make sense of lots of disparate data. Even unsophisticated old school stuff like éclat is pretty helpful and various implementations of this sort of thing are efficient enough to be dangerous.

3) Community detection. This is an interesting bag of ideas that has grown  powerful over the years. Interestingly I’m not sure there is a good book on the subject, and it seems virtually unknown among practitioners who do not specialize in it. A lot of it is “just” graph theory or un/semi-supervised learning of various kinds.

4) Human/computer interfaces are getting better. Very often a machine learning algorithm is more like a filter that sends vastly smaller lists of problems for human analysts to solve. Palantir originated to do stuff like this, and while very little stuff on human computer interfaces is open source, the software is pretty good at this point.

5) Labels are becoming ubiquitous. Most people do supervised learning, which … requires labels for supervision. Unfortunately with various kinds of cookies out there, people using nerd dildos for everything, networked GPS, IOT, radio tags and so on; there are labels for all kinds of things which didn’t exist before. I’m guessing as of now or very soon, you won’t need to be a government agency to track individuals in truly Orwellian ways based on the trash data in your various devices; you’ll just need a few tens of millions of dollars worth of online ad company. Pretty soon this will be offered as a service.

Ignorance of these topics is keeping us safe

1) Database software is crap. Databases are … OK for some purposes; they’re nowhere near their theoretical capabilities in solving these kinds of problems. Database researchers are, oddly enough, generally not interested in solving real data problems. So you get mediocre crap like Postgres; bleeding edge designs from the 1980s. You have total horse shit like Spark, laughably insane things like Hive, and … sort of OK designs like bigtables… These will keep database engineers and administrators employed for decades to come, and prevent the solution of all kinds of important problems. There are people and companies out there that know what they’re doing. One to watch is 1010 data; people who understand basic computing facts, like “latency.” Hopefully they will be badly managed by their new owners. The engineering team is probably the best to beat this challenge. The problem with databases is multifold: getting at the data you need is important. Keeping it close to learning algorithms is also important. None of these things are done well by any existing publicly available database engines. Most of what exists in terms of database technology is suitable for billing systems, not data science. Usually people build custom tools to solve specific problems; like the high frequency trader guys who built custom data tee-offs and backtesting frameworks instead of buying a more general tool like Kx. This is fine by me; perpetual employment. Lots of companies do have big data storages, but most of them still can’t get at their data in any useful way. If you’ve ever seen these things, and actually did know what you were doing, even at the level of 1970s DBA, you would laugh hysterically. Still, enough spergs have built pieces of Kx type things that eventually someone will get it right.

2) Database metadata is hard to deal with. One of the most difficult problems for any data scientist is the data preparation phase. There’s much to be said about preparation of data, but one of the most important tasks in preparing data for analysis is joining data gathered in different databases. The very simple example is the data from the ad server and the data from the sales database not talking to each other. So, when I click around Amazon and buy something, the imbecile ad-server will continue to serve me ads on the thing that Amazon knows it has already sold me. This is a trivial example: one that Amazon could solve in principle, but in practice it is difficult and hairy enough that it isn’t worth the money for Amazon to fix this (I have a hack which fixes the ad serving problem, but it doesn’t solve the general problem). This is a pervasive problem, and it’s a huge, huge thing preventing more data being used against the average individual. If “AI” were really a thing, this is where it would be applied. This is actually a place where machine learning potentially could be used, but I think there are several reasons it won’t be, and this will remain a big impediment to tracking and privacy invasions in 20 years. FWIIW back to my ezpass license plate photographer thing; sticking a billing system in with at least two government databases per state that something like ezpass works in -unless they all used the same system (possible), it was a clever thing which hits this bullet point.

3) Most commonly used forms of machine learning requires many examples. People have been concentrating on Deep Learning, which almost inherently requires many, many examples. This is good for the private minded; most data science teams are too dumb to use techniques which don’t require a lot of examples. These techniques exist; some of them have for a long time. For the sake of this discussion, I’ll call these “sort of like Bayesian” -which isn’t strictly true, but which will shut people up. I think it’s great the average sperglord is spending all his time on Deep Learning which is 0.2% more shiny, assuming you have Google’s data sets. If a company like google had techniques which required few examples, they’d actually be even more dangerous.

4) Most people can only do supervised learning. (For that matter, non-batch learning terrifies most “data scientists” -just like Kalman filters terrify statisticians even though it is the same damn thing as linear regression). There is some work on stuff like reinforcement learning being mentioned in the funny papers. I guess reinforcement learning is interesting, but it is not really all that useful for anything practical. The real interesting stuff is semi-supervised, unsupervised, online and weak learning. Of course, all of these things are actually hard, in that they mostly do not exist as prepackaged tools in R you can use in a simple recipe. So, the fact that most domain “experts” are actually kind of shit at machine learning is keeping us safe.

A shockingly sane exposition of what to expect from machine learning, which I even more shockingly found on a VC’s website:

## Predicting with confidence: the best machine learning idea you never heard of

Posted in machine learning by Scott Locklin on December 5, 2016

One of the disadvantages of machine learning as a discipline is the lack of reasonable confidence intervals on a given prediction. There are all kinds of reasons you might want such a thing, but I think machine learning and data science practitioners are so drunk with newfound powers, they forget where such a thing might be useful. If you’re really confident, for example, that someone will click on an ad, you probably want to serve one that pays a nice click through rate. If you have some kind of gambling engine, you want to bet more money on the predictions you are more confident of. Or if you’re diagnosing an illness in a patient, it would be awfully nice to be able to tell the patient how certain you are of the diagnosis and what the confidence in the prognosis is.

There are various ad hoc ways that people do this sort of thing.  The one you run into most often is some variation on cross validation, which produces an average confidence interval. I’ve always found this to be dissatisfying (as are PAC approaches). Some people fiddle with their learners and in hopes of making sure the prediction is normally distributed, then build confidence intervals from that (or for the classification version, Platt scaling using logistic regression).  There are a number of ad hoc ways of generating confidence intervals using resampling methods and generating a distribution of predictions. You’re kind of hosed though, if your prediction is in online mode.  Some people build learners that they hope will produce a sort of estimate of the conditional probability distribution of the forecast; aka quantile regression forests and friends. If you’re a Bayesian, or use a model with confidence intervals baked in, you may be in pretty good shape. But let’s face it, Bayesian techniques assume your prior is correct, and that new points are drawn from your prior. If your prior is wrong, so are your confidence intervals, and you have no way of knowing this.  Same story with heteroscedasticity. Wouldn’t it be nice to have some tool to tell you how uncertain your prediction when you’re not certain of your priors or your algorithm, for that matter?

Well, it turns out, humanity possesses such a tool, but you probably don’t know about it. I’ve known about this trick for a few years now, through my studies of online and compression based learning as a general subject. It is a good and useful bag of tricks, and it verifies many of the “seat of the pants” insights I’ve had in attempting to build ad-hoc confidence intervals in my own predictions for commercial projects.  I’ve been telling anyone who listens for years that this stuff is the future, and it seems like people are finally catching on. Ryan Tibshirani, who I assume is the son of the more famous Tibshirani, has published a neat R package on the topic along with colleagues at CMU. There is one other R package out there and one in python. There are several books published in the last two years. I’ll do my part in bringing this basket of ideas to a more general audience, presumably of practitioners, but academics not in the know should also pay attention.

The name of this basket of ideas is “conformal prediction.” The provenance of the ideas is quite interesting, and should induce people to pay attention. Vladimir Vovk is a former Kolmogorov student, who has had all kind of cool ideas over the years. Glenn Shafer is also well known for his co-development of Dempster-Shafer theory, which is a brewing alternative to standard measure-theoretic probability theory which is quite useful in sensor fusion, and I think some machine learning frameworks. Alexander Gammerman is a former physicist from Leningrad, who, like Shafer, has done quite a bit of work in the past with Bayesian belief networks. Just to reiterate who these guys are: Vovk and Shafer have also previously developed a probability theory based on game theory which has ended up being very influential in machine learning pertaining to sequence prediction. To invent one new form of probability theory is clever. Two is just showing off! The conformal prediction framework comes from deep results in probability theory and is inspired by Kolmogorov and Martin-Lof’s ideas on algorithmic complexity theory.

The advantages of conformal prediction are many fold. These ideas assume very little about the thing you are trying to forecast, the tool you’re using to forecast or how the world works, and they still produce a pretty good confidence interval. Even if you’re an unrepentant Bayesian, using some of the machinery of conformal prediction, you can tell when things have gone wrong with your prior. The learners work online, and with some modifications and considerations, with batch learning. One of the nice things about calculating confidence intervals as a part of your learning process is they can actually lower error rates or use in semi-supervised learning as well. Honestly, I think this is the best bag of tricks since boosting; everyone should know about and use these ideas.

The essential idea is that a “conformity function” exists. Effectively you are constructing a sort of multivariate cumulative distribution function for your machine learning gizmo using the conformity function. Such CDFs exist for classical stuff like ARIMA and linear regression under the correct circumstances; CP brings the idea to machine learning in general, and to models like ARIMA  when the standard parametric confidence intervals won’t work. Within the framework, the conformity function, whatever may be, when used correctly can be guaranteed to give confidence intervals to within a probabilistic tolerance. The original proofs and treatments of conformal prediction, defined for sequences, is extremely computationally inefficient. The conditions can be relaxed in many cases, and the conformity function is in principle arbitrary, though good ones will produce narrower confidence regions. Somewhat confusingly, these good conformity functions are referred to as “efficient” -though they may not be computationally efficient.

The original research and proofs were done on so-called “transductive conformal prediction.” I’ll sketch this out below.

Suppose you have a data set $Z:= z_1,...,z_N$, with $z_i:=(x_i,y_i)$ where $x_i$ has the usual meaning of a feature vector, and $y_i$ the variable to be predicted. If the $N!$ different possible orderings are equally likely, the data set $Z$ is exchangeable. For the purposes of this argument, most data sets are exchangeable or can be made so. Call the set of all bags of points from $Z$ with replacement a “bag” $B$.

The conformal predictor $\Gamma^{\epsilon}(Z,x) := \{y | y^{p} > \epsilon \}$ where $Z$ is the training set and $x$ is a test object and $\epsilon \in (0,1)$ is a defined probability of confidence in a prediction. If we have a function $A(B,z_i)$ which measures how different a point $z_i$ is the bag set of $B$.

Example: If we have a forecast technique which works on exchangeable data, $\phi(B)$, then a very simple function is the distance between the new point and the forecast based on the bag set. $A(B,z_i):=d(\phi(B), z_i)$.

Simplifying the notation a little bit, let’s call $A_i := A(B^{-i},z_i)$ where $B^{-i}$ is the bag set, missing $z_i$. Remembering that bag sets $B$ are sets of all the orderings of $Z$ we can see that our $p^y$ can be defined from the nonconformity measures; $p^{y} := \frac{\#\{i=1,...,n|A_i \geq A_n \} }{n}$ This can be proved in a fairly straightforward way. You can find it in any of the books and most of the tutorials.

Practically speaking, this kind of transductive prediction is computationally prohibitive and not how most practitioners confront the world. Practical people use inductive prediction, where we use training examples and then see how they do in a test set. I won’t go through the general framework for this, at least this time around; go read the book or one of the tutorials listed below. For one it is worth, one of the forms of Inductive Conformal Prediction is called Mondrian Conformal Prediction; a framework which allows for different error rates for different categories, hence all the Mondrian paintings I decorated this blog post with.

For many forms of inductive CP, the main trick is you must subdivide your training set into two pieces. One piece you use to train your model, the proper training set. The other piece you use to calculate your confidence region, the calibration set. You compute the non-conformity scores on the calibration set, and use them on the predictions generated by the proper training set. There are
other blended approaches. Whenever you use sampling or bootstrapping in  your prediction algorithm, you have the chance to build a conformal predictor using the parts of the data not used in the prediction in the base learner. So, favorites like Random Forest and Gradient Boosting Machines have computationally potentially efficient conformity measures. There are also flavors using a CV type process, though the proofs seem more weak for these. There are also reasonably computationally efficient Inductive CP measures for KNN, SVM and decision trees. The inductive “split conformal predictor” has an R package associated with it defined for general regression problems, so it is worth going over in a little bit of detail.
For coverage at $\epsilon$ confidence, using a prediction algorithm $\phi$ and training data set $Z_i,i=1,...,n$, randomly split the index $i=1,...,n$ into two subsets, which as above, we will call the proper training set and the calibration set $I_1,I_2$.

Train the learner using data on the proper training set $I_1$

$\phi_{trained}:=\phi(Z_i); i \in I_1$. Then, using the trained learner, find the residuals in the calibration set:

$R_i := |Y_i - \phi(X_i)|, i \in I_2$
$d :=$ the $k$th smallest value in $\{R_i :i \in I_2\}$ where
$k=(n/2 + 1)(1-\epsilon)$

The prediction interval for a new point $x$ is $\phi(x)-d,\phi(x)+d$

This type of thing may seem unsatisfying, as technically the bounds on it only exist for one predicted point. But there are workarounds using leave one out in the ranking. The leave one out version is a little difficult to follow in a lightweight blog, so I’ll leave it up as an exercise for those who are interested to read more about it in the R documentation for the package.

Conformal prediction is about 10 years old now: still in its infancy.  While forecasting with confidence intervals is inherently useful, the applications and extensions of the idea are what really tantalizes me about the subject. New forms of feature selection, new forms of loss function which integrate the confidence region, new forms of optimization to deal with conformal loss functions, completely new and different machine learning algorithms, new ways of thinking about data and probabilistic prediction in general. Specific problems which CP has had success with; face recognition, nuclear fusion research, design optimization, anomaly detection, network traffic classification and forecasting, medical diagnosis and prognosis, computer security, chemical properties/activities prediction and computational geometry. It’s probably only been used on a few thousand different data sets. Imagine being at the very beginning of Bayesian data analysis where things like the expectation maximization algorithm are just being invented, or neural nets before backpropagation: I think this is where the CP basket of ideas is at.  It’s an exciting field at an exciting time, and while it is quite useful now, all kinds of great new results will come of it.

There is a website and a book. Other papers and books can be found in the usual way. This paper goes with the R package mentioned above, and is particularly clearly written for the split and leave one out conformal prediction flavors. Here is a presentation with some open problems and research directions if you want to get to work on something interesting. Only 19 packages on github so far.

## Neglected machine learning ideas

Posted in machine learning, statistical tools, tools by Scott Locklin on July 22, 2014

This post is inspired by the “metacademy” suggestions for “leveling up your machine learning.” They make some halfway decent suggestions for beginners.  The problem is, these suggestions won’t give you a view of machine learning as a field; they’ll only teach you about the subjects of interest to authors of machine learning books, which is different. The level-3 and level-4 suggestions they make are not super useful either: they just reflect the tastes of the author.

The machine learning literature is vast, techniques are bewilderingly diverse, multidisciplinary and seemingly unrelated. It is extremely difficult to know what is important and useful. While “metacademy” has the horse sense to suggest reading some books, the problem is, there is no book which can even give you a survey of what is available, or make you aware of things which might be helpful. The best guide for the perplexed, in my not at all humble opinion, is Peter Flach’s introductory text, “Machine Learning: the Art and Science of Algorithms that Make Sense of Data” which at least mentions some of the more obscure techniques, and makes pointers to other resources. Most books are just a collection of the popular techniques. They all mention regression models, logistic regression, neural nets, trees, ensemble methods, graphical models and SVM type things. Most of the time, they don’t even bother telling you what each technique is actually good for, and when you should choose one over the other for an approach (Flach does; that’s one of many reasons you should read his book). Sometimes I am definitely just whining that people don’t pay enough attention to the things I find interesting, or that I don’t have a good book or review article on the topic. Sleep deprivation will do that to a man. Sometimes I am probably putting together things that have no clearly unifying feature, perhaps because they’re “not done yet.” I figure that’s OK, subjects such as “deep learning” are also a bunch of ideas that have no real unifying theme and aren’t done yet; this doesn’t stop people from writing good treatments of the subject. Perhaps my list is a “send me review articles and book suggestions” cry for help, but perhaps it is useful to others as an overview of neat things.

Stuff I think is egregiously neglected in books, and in academia in unranked semi-clustered listing below:

Online learning: not the “Khan academy” kind, the “exposing your learners to data, one piece at a time, the way the human  brain works” kind. This is hugely important for “big data” and timeseries, but there are precious few ML texts which go beyond mentioning the existence of online learning in passing. Almost all textbooks concentrate on batch learning. Realistically, when you’re dealing with timeseries or very large data sets, you’re probably doing things online in some sense. If you’re not thinking about how you’re exposing your learners to sequentially generated data, you’re probably leaving information on the table, or overfitting to irrelevant data. I can think of zero books which are actually helpful here. Cesa-Bianchi and Lugosi wrote a very interesting book on some recent proofs for online learners and “universal prediction” which strike me as being of extreme importance, though this is a presentation of new ideas, rather than an exposition of established ones.  Vowpal Wabbit is a useful and interesting piece of software with OK documentation, but there should be a book which takes you from online versions of linear regression (they exist! I can show you one!)  to something like Vowpal Wabbit. Such a book does not exist. Hell, I am at a loss to think of a decent review article, and the subject is unfortunately un-googleable, thanks to the hype over the BFD of “watching lectures and taking tests over the freaking internets.” Please correct me if I am wrong: I’d love to have a good review article on the subject for my own purposes.

Reinforcement learning: a form of online learning which has become a field unto its own. One of the great triumphs of machine learning is teaching computers to win at Backgammon. This was done via a form of reinforcement learning known as TD-learning. Reinforcement learning is a large field, as it has been used with great success in control systems theory and robotics. The problem is, the guys who do reinforcement learning are generally in control systems theory and robotics, making the literature impenetrable to machine learning researchers and engineers. Something oriented towards non robotics problems would be nice (Sutton and Barto doesn’t suffice here; Norvig’s chapter is the best general treatment I have thus far seen). There are papers on applications of the idea to ideas which do not involve robots, but none which unify the ideas into something comprehensible and utile to a ML engineer.

“Compression” sequence prediction techniques: this is another form of online learning, though it can also be done in batch mode. We’re all familiar with this; when google tries to guess what you’re going to search for, it is using a primitive form of this called the Trie. Such ideas are related to standard compression techniques like LZW, and have deep roots in information theory and signal processing. Really, Claude Shannon wrote the first iterations of this idea. I can’t give you a good reference for this subject in general, though Ron Begleiter and friends wrote a very good paper on some classical compression learning implementations and their uses. I wrote an R wrapper for their Java lib if you want to fool around with their tool. Boris Ryabko and son have also written numerous interesting papers on the subject. Complearn is a presumably useful library which encapsulates some of these ideas, and is available everywhere Linux is sold. Some day I’ll expound on these ideas in more detail.

Time series oriented techniques in general: a large fraction of  industry applications have a time component. Even in marketing problems dealing with survival techniques, there is a time component, and you should know about it.In situations where there are non-linear relationships in the time series, classical regression and time-series techniques will fail. In situations where you must discover the underlying non-linear model yourself, well, you’re in deep shit if you don’t know some time-series oriented machine learning techniques.  There was much work done in the 80s and 90s on tools like recurrent ANNs and feedforward ANNs for starters, and there has been much work in this line since then. There are plenty of other useful tools and techniques.  Once in a while someone will mention dynamic time warping in a book, but nobody seems real happy about this technique.  Many books mention Hidden Markov Models, which are important, but they’re only useful when the data is at least semi-Markov, and you have some idea of how to characterize it as a sequence of well defined states. Even in this case, I daresay not even the natural language recognition textbooks are real helpful (though Rabiner and Juang is OK, it’s also over 20 years old). Similarly, there are no review papers  treating this as a general problem. I guess we TS guys are too busy racking in the lindens to write one.

Conformal prediction: I will be surprised if anyone reading this has even heard of conformal prediction. There are no wikipedia entries. There is a website and a book. The concept is simple: it would be nice to well motivated put error bars on a machine learning prediction. If you read the basic books, stuff like k-fold cross validation and the jackknife  trick are the entire story. OK, WTF do I do when my training is online? What do I do in the presence of different kinds of noise? Conformal prediction is a step towards this, and hopefully a theory of machine learning confidence intervals in general. It seems to mostly be the work of a small group of researchers who were influenced by Kolomogorov, but others are catching on. I’m interested. Not interested enough to write one, as of yet, but I’d sure like to play with one.

ML in the presence of lots of noise: The closest thing to a book on it is the bizarro (and awesomely cool) “Pattern Theory: The Stochastic Analysis of Real World Signals” by Mumford and Desolneux, or perhaps something in the corpus of speech recognition and image processing books. This isn’t exactly a cookbook or exposition, mind you: more of a thematic manifesto with a few applications.  Obviously, signal processing has something to say about the subject, but what about learners which are designed to function usefully when we know that most of the data is noise?  Fields such as natural language processing and image processing are effectively ML in the presence of lots of noise and confounding signal, but the solutions you will find in their textbooks are specifically oriented to the problems at hand.  Once in a while something like vector quantization will be reused across fields, but it would be nice if we had an “elements of statistical learning in the presence of lots of noise” type book or review paper. Missing in action, and other than the specific subfields mentioned above, there are no research groups which study the problem as an engineering subject. New stuff is happening all the time; part of the success of “Deep Learning” is attributable to the Drop Out technique to prevent overfitting. Random forests could be seen as a technique which at genuflects at  “ML in the presence of noise” without worrying about it too much. Marketing guys are definitely thinking about this. I know for a fact that there are very powerful learners for picking signal out of shitloads of noise: I’ve written some. It would have been a lot easier if somebody wrote a  review paper on the topic. The available knowledge can certainly be systematized and popularized better than it has been.

Feature engineering: feature engineering is another topic which doesn’t seem to merit any review papers or books, or even chapters in books, but it is absolutely vital to ML success. Sometimes the features are obvious; sometimes not. Much of the success of machine learning is actually success in engineering features that a learner can understand. I daresay document classification would be awfully difficult without td-idf representation of document features. Latent Dirichlet allocation is a form of “graphical model” which works wonders on such data, but it wouldn’t do a thing without td-idf. [correction to this statement from Brendan below] Similarly, image processing has a bewildering variety of feature extraction algorithms which are of towering importance for that field; the SIFT descriptor, the GIST and HOG descriptors, the Hough transform, vector quantization, tangent distance [pdf link]. The Winner Take All hash [pdf link] is an extremely simple and related idea… it makes a man wonder if such ideas could be used in higher (or lower) dimensions. Most of these engineered features are histograms in some sense, but just saying “use a histogram” isn’t helpful. A review article or a book chapter on this sort of thing, thinking through the relationships of these ideas, and helping the practitioner to engineer new kinds of feature for broad problems would be great. Until then, it falls to the practitioner to figure all this crap out all by their lonesome.

Unsupervised and semi-supervised learning in general: almost all books, and even tools like R inherently assume that you are doing supervised learning, or else you’re doing something real simple, like hierarchical clustering, kmeans or PCA.  In the presence of a good set of features, or an interesting set of data, unsupervised techniques can be very helpful. Such techniques may be crucial. They may even help you to engineer new features, or at least reduce the dimensionality of your data. Many interesting data sets are only possible to analyze using semi-supervised techniques; recommendation engines being an obvious beneficiary of such tricks. “Deep learning” is also connected with unsupervised and semi-supervised approaches. I am pretty sure the genomics community does a lot of work with this sort of thing for dimensionality reduction. Supposedly Symbolic Regression (generalized additive models picked using genetic algorithms) is pretty cool too, and it’s in my org-emacs TODO lists to look at this more. Lots of good unsupervised techniques such as Kohonen Self Organizing Maps have fallen by the wayside. They’re still useful: I use them. I’d love a book or review article which concentrates on the topic, or just provides a bestiary of things which are broadly unsupervised. I suppose Oliver Chapelle’s book is an OK start for semi-supervised ideas, but again, not real unified or complete.

Images by one of my heroes, the Ukrainian-American artist Boris Artzybasheff. You can find more of it here.