Linear Digressions

If you're anything like us, you a) always are curious to learn more about data science and machine learning and stuff, and b) are usually overwhelmed by how much content is out there (not all of it very digestible). We hope this podcast is a part of the solution for you, but if you're looking to go farther (who isn't?) then we have a few new resources that are presenting high-quality content in a fresh, accessible way. Boring old PDFs full of inscrutable math notation, your days are numbered!

As machine learning makes its way into more and more mobile devices, an interesting question presents itself: how can we have an algorithm learn from training data that's being supplied as users interact with the algorithm? In other words, how do we do machine learning when the training dataset is distributed across many devices, imbalanced, and the usage associated with any one user needs to be obscured somewhat to protect the privacy of that user? Enter Federated Learning, a set of related algorithms from Google that are designed to help out in exactly this scenario. If you've used keyboard shortcuts or autocomplete on an Android phone, chances are you've encountered Federated Learning even if you didn't know it.

Word2Vec is probably the go-to algorithm for vectorizing text data these days. Which makes sense, because it is wicked cool. Word2Vec has it all: neural networks, skip-grams and bag-of-words implementations, a multiclass classifier that gets swapped out for a binary classifier, made-up dummy words, and a model that isn't actually used to predict anything (usually). And all that's before we get to the part about how Word2Vec allows you to do algebra with text. Seriously, this stuff is cool.

It seems like every day there's more and more machine learning problems that involve learning on text data, but text itself makes for fairly lousy inputs to machine learning algorithms. That's why there are text vectorization algorithms, which re-format text data so it's ready for using for machine learning. In this episode, we'll go over some of the most common and useful ways to preprocess text data for machine learning.

This week we'll hop into the rapidly developing industry around predictive analytics for education. For many of the students who eventually drop out, data science is showing that there might be early warning signs that the student is in trouble--we'll talk about what some of those signs are, and then dig into the meatier questions around discrimination, who owns a student's data, and correlation vs. causation. Spoiler: we have more questions than we have answers on this one. Bonus appearance from Maeby the dog, who isn't a data scientist but does like to steal food off the counter.

In our follow-up episode to last week's introduction to the first self-driving car, we will be doing a technical deep dive this week and talking about the most important systems for getting a car to drive itself 140 miles across the desert. Lidar? You betcha! Drive-by-wire? Of course! Probabilistic terrain reconstruction? Absolutely! All this and more this week on Linear Digressions.

In October 2005, 23 cars lined up in the desert for a 140 mile race. Not one of those cars had a driver. This was the DARPA grand challenge to see if anyone could build an autonomous vehicle capable of navigating a desert route (and if so, whose car could do it the fastest); the winning car, Stanley, now sits in the Smithsonian Museum in Washington DC as arguably the world's first real self-driving car. In this episode (part one of a two-parter), we'll revisit the DARPA grand challenge from 2005 and the rules and constraints of what it took for Stanley to win the competition. Next week, we'll do a deep dive into Stanley's control systems and overall operation and what the key systems were that allowed Stanley to win the race.

Figuring out what features actually matter in a model is harder to figure out than you might first guess. When a human makes a decision, you can just ask them--why did you do that? But with machine learning models, not so much. That's why we wanted to talk a bit about both regularization (again) and also other ways that you can figure out which models have the biggest impact on the predictions of your model.

It's hard to get information to and from Mars. Mars is very far away, and expensive to get to, and the bandwidth for passing messages with Earth is not huge. The messages you do pass have to traverse millions of miles, which provides ample opportunity for the message to get corrupted or scrambled. How, then, can you encode messages so that errors can be detected and corrected? How does the decoding process allow you to actually find and correct the errors? In this episode, we'll talk about three pieces of the process (Reed-Solomon codes, convolutional codes, and Viterbi decoding) that allow the scientists at NASA to talk to our rovers on Mars.

You may be shocked to hear this, but sometimes, people on the internet can be mean. For some of us this is just a minor annoyance, but if you're a maintainer or contributor of a large project like Wikipedia, abusive users can be a huge problem. Fighting the problem starts with understanding it, and understanding it starts with measuring it; the thing is, for a huge website like Wikipedia, there can be millions of edits and comments where abuse might happen, so measurement isn't a simple task. That's where machine learning comes in: by building an "abuse classifier," and pointing it at the Wikipedia edit corpus, researchers at Jigsaw and the Wikimedia foundation are for the first time able to estimate abuse rates and curate a dataset of abusive incidents. Then those researchers, and others, can use that dataset to study the pathologies and effects of Wikipedia trolls.

Advances in neural networks are moving fast enough that, even though it seems like we talk about them all the time around here, it also always seems like we're barely keeping up. So this week we have another installment in our "neural nets: they so smart!" series, talking about three topics. And all the topics this week were listener suggestions, too!

When you're estimating something about some object that's a member of a larger group of similar objects (say, the batting average of a baseball player, who belongs to a baseball team), how should you estimate it: use measurements of the individual, or get some extra information from the group? The James-Stein estimator tells you how to combine individual and group information make predictions that, taken over the whole group, are more accurate than if you treated each individual, well, individually.

Say you're looking to use some Bayesian methods to estimate parameters of a system. You've got the normalization figured out, and the likelihood, but the prior... what should you use for a prior? Empirical Bayes has an elegant answer: look to your previous experience, and use past measurements as a starting point in your prior. Scratching your head about some of those terms, and why they matter? Lucky for you, you're standing in front of a podcast episode that unpacks all of this.

Have you been out protesting lately, or watching the protests, and wondered how much effect they might have on lawmakers? It's a tricky question to answer, since usually we need randomly distributed treatments (e.g. big protests) to understand causality, but there's no reason to believe that big protests are actually randomly distributed. In other words, protest size is endogenous to legislative response, and understanding cause and effect is very challenging. So, what to do? Well, at least in the case of studying Tea Party protest effectiveness, researchers have used rainfall, of all things, to understand the impact of a big protest. In other words, rainfall is the instrumental variable in this analysis that cracks the scientific case open. What does rainfall have to do with protests? Do protests actually matter? What do we mean when we talk about endogenous and instrumental variables? We wouldn't be very good podcasters if we answered all those questions here--you gotta listen to this episode to find out.

Remember last week, when we were talking about how great the ROC curve is for evaluating models? How things change... This week, we're exploring calibrated risk models, because that's a kind of model that seems like it would benefit from some nice ROC analysis, but in fact the ROC AUC can steer you wrong there.

This week: everybody's favorite WWII-era classifier metric! But it's not just for winning wars, it's a fantastic go-to metric for all your classifier quality needs.

If one machine learning model is good, are two models better? In a lot of cases, the answer is yes. If you build many ok models, and then bring them all together and use them in combination to make your final predictions, you've just created an ensemble model. It feels a little bit like cheating, like you just got something for nothing, but the results don't like: algorithms like Random Forests and Gradient Boosting Trees (two types of ensemble algorithms) are some of the strongest out-of-the-box algorithms for classic supervised classification problems. What makes a Random Forest random, and what does it mean to gradient boost a tree? Have a listen and find out.

As anyone who's encountered a badly translated text could tell you, not all translations are created equal. Some translations are smooth, fluent and sound like a poet wrote them; some are jerky, non-grammatical and awkward. When a machine is doing the translating, it's awfully easy to end up with a robotic-sounding text; as the state of the art in machine translation improves, though, a natural question to ask is: according to what measure? How do we quantify a "good" translation? Enter the BLEU score, which is the standard metric for quantifying the quality of a machine translation. BLEU rewards translations that have large overlap with human translations of sentences, with some extra heuristics thrown in to guard against weird pathologies (like full sentences getting translated as one word, redundancies, and repetition). Nowadays, if there's a machine translation being evaluated or a new state-of-the-art system (like the Google neural machine translation we've discussed on this podcast before), chances are that there's a BLEU score going into that assessment.

Take Google-size data, the flexibility of a neural net, and all (well, most) of the languages of the world, and what you end up with is a pile of surprises. This episode is about some interesting features of Google's new neural machine translation system, namely that with minimal tweaking, it can accommodate many different languages in a single neural net, that it can do a half-decent job of translating between language pairs it's never been explicitly trained on, and that it seems to have its own internal representation of concepts that's independent of the language those concepts are being represented in. Intrigued? You should be...

Recently, Google swapped out the backend for Google Translate, moving from a statistical phrase-based method to a recurrent neural network. This marks a big change in methodology: the tried-and-true statistical translation methods that have been in use for decades are giving way to a neural net that, across the board, appears to be giving more fluent and natural-sounding translations. This episode recaps statistical phrase-based methods, digs into the RNN architecture a little bit, and recaps the impressive results that is making us all sound a little better in our non-native languages.

Today we are delighted to bring you an interview with Matt Might, computer scientist and medical researcher extraordinaire and architect of President Obama's Precision Medicine Initiative. As the Obama Administration winds down, we're talking with Matt about the goals and accomplishments of precision medicine (and related projects like the Cancer Moonshot) and what he foresees as the future marriage of data and medicine. Many thanks to Matt, our friends over at Partially Derivative (hi, Jonathon!) and the White House for arranging this opportunity to chat. Enjoy!

We have the pleasure of bringing you a very special crossover episode this week: our friends at Partially Derivative (another great podcast about data science, you should check it out) recently interviewed White House Chief Data Scientist DJ Patil. We think DJ's message about the importance and impact of data science is worth spreading, so it's our pleasure to bring it to you today. A huge thanks to Jonathon Morgan and Partially Derivative for sharing this interview with us--enjoy! Relevant links: http://partiallyderivative.com/podcast/2016/12/13/dj-patil

Competing in a machine learning competition on Kaggle is a kind of rite of passage for data scientists. Losing unexpectedly at the very end of the contest is also something that a lot of us have experienced. It's not just bad luck: a very specific combination of overfitting on popular competitions can take someone who is in the top few spots in the final days of a contest and bump them down hundreds of slots in the final tally.

Imagine there's an important decision to be made about someone, like a bank deciding whether to extend a loan, or a school deciding to admit a student--unfortunately, we're all too aware that discrimination can sneak into these situations (even when everyone is acting with the best of intentions!). Now, these decisions are often made with the assistance of machine learning and statistical models, but unfortunately these algorithms pick up on the discrimination in the world (it sneaks in through the data, which can capture inequities, which the algorithms then learn) and reproduce it. This podcast covers some of the most common ways we can try to minimize discrimination, and why none of those ways is perfect at fixing the problem. Then we'll get to a new idea called "equality of opportunity," which came out of Google recently and takes a pretty practical and well-aimed approach to machine learning bias.

This week, we're doing a crash course in recurrent neural networks--what the structural pieces are that make a neural net recurrent, how that structure helps RNNs solve certain time series problems, and the importance of forgetfulness in RNNs. Relevant links: http://colah.github.io/posts/2015-08-Understanding-LSTMs/

Want another reason to be paranoid when using the free coffee shop wifi? Allow us to introduce WindTalker, a system that cleverly combines a dose of signal processing with a dash of machine learning to (potentially) steal the PIN from your phone transactions without ever having physical access to your phone. This episode has it all, folks--channel state information, ICMP echo requests, low-pass filtering, PCA, dynamic time warps, and the PIN for your phone.

Cryptography used to be the domain of information theorists and spies. There's a new player now: neural networks. Given the task of communicating securely, neural networks are inventing new encryption methods that, as best we can tell, are unlike anything humans have ever seen before. Relevant links: http://arstechnica.co.uk/information-technology/2016/10/google-ai-neural-network-cryptography/ https://arxiv.org/pdf/1610.06918v1.pdf

In 1997, Deep Blue was the IBM algorithm/computer that did what no one, at the time, though possible: it beat the world's best chess player. It turns out, though, that one of the most important moves in the matchup, where Deep Blue psyched out its opponent with a weird move, might not have been so inspired after all. It might have been nothing more than a bug in the program, and it changed computer science history. Relevant links: https://www.wired.com/2012/09/deep-blue-computer-bug/

If you're a data scientist, there's a good chance you're used to working with a lot of data. But there's a lot of data, and then there's Google-scale amounts of data. Keeping all that data organized is a Google-sized task, and as it happens, they've built a system for that organizational challenge. This episode is all about that system, called Goods, and in particular we'll dig into some of the details of what makes this so tough. Relevant links: http://static.googleusercontent.com/media/research.google.com/en//pubs/archive/45390.pdf

A few months ago, Katie started on a project for the Vice President's Cancer Moonshot surrounding how data can be used to better fight cancer. The project is all wrapped up now, so we wanted to tell you about how that work went and what changes to cancer data policy were suggested to the Vice President. See lineardigressions.com for links to the reports discussed on this episode.

Sick of the presidential election yet? We are too, but there's still almost a month to go, so let's just embrace it together. This week, we'll talk about one of the presidential polls, which has been kind of an outlier for quite a while. This week, the NY Times took a closer look at this poll, and was able to figure out the reason it's such an outlier. It all goes back to a 19-year-old African American man, living in Illinois, who really likes Donald Trump... Relevant Links: http://www.nytimes.com/2016/10/13/upshot/how-one-19-year-old-illinois-man-is-distorting-national-polling-averages.html followup article from LA Times, released after recording: http://www.latimes.com/politics/la-na-pol-daybreak-poll-questions-20161013-snap-story.html

What does it mean to steal a model? It means someone (the thief, presumably) can re-create the predictions of the model without having access to the algorithm itself, or the training data. Sound far-fetched? It isn't. If that person can ask for predictions from the model, and he (or she) asks just the right questions, the model can be reverse-engineered right out from under you. Relevant links: https://www.usenix.org/system/files/conference/usenixsecurity16/sec16_paper_tramer.pdf

Lots of data is usually seen as a good thing. And it is a good thing--except when it's not. In a lot of fields, a problem arises when you have many, many features, especially if there's a somewhat smaller number of cases to learn from; supervised machine learning algorithms break, or learn spurious or un-interpretable patterns. What to do? Regularization can be one of your best friends here--it's a method that penalizes overly complex models, which keeps the dimensionality of your model under control.

You might sometimes find that it's hard to get started doing something, but once you're going, it gets easier. Turns out machine learning algorithms, and especially recommendation engines, feel the same way. The more they "know" about a user, like what movies they watch and how they rate them, the better they do at suggesting new movies, which is great until you realize that you have to start somewhere. The "cold start" problem will be our focus in this episode, both the heuristic solutions that help deal with it and a bit of realism about the importance of skepticism when someone claims a great solution to cold starts. Relevant links: http://repository.upenn.edu/cgi/viewcontent.cgi?article=1141&context=cis_papers

If you work in tech, software or data science, there's an excellent chance you use tools that are built upon open source software. This is software that's built and distributed not for a profit, but because everyone benefits when we work together and share tools. Tim Head of scikit-optimize chats with us further about what it's like to maintain an open source library, how to get involved in open source, and why people like him need people like you to make it all work.

We're excited to welcome a guest, Tim Head, who is one of the maintainers of the scikit-optimize package. With all the talk about optimization lately, it felt appropriate to get in a few words with someone who's out there making it happen for python. Relevant links: https://scikit-optimize.github.io/ http://www.wildtreetech.com/

It's a funny thing to realize, but data science modeling is usually about either explainability, interpretation and understanding, or it's about predictive accuracy. But usually not both--optimizing for one tends to compromise the other. Leo Breiman was one of the titans of both kinds of modeling, a statistician who helped bring machine learning into statistics and vice versa. In this episode, we unpack one of his seminal papers from 2001, when machine learning was just beginning to take root, and talk about how he made clear what machine learning could do for statistics and why it's so important. Relevant links: http://www.math.snu.ac.kr/~hichoi/machinelearning/(Breiman)%20Statistical%20Modeling--The%20Two%20Cultures.pdf

You've got an optimization problem to solve, and a less-than-forever amount of time in which to solve it. What do? Use a heuristic optimization algorithm, like a hill climber or simulated annealing--we cover both in this episode! Relevant link: http://www.lizsander.com/programming/2015/08/04/Heuristic-Search-Algorithms.html

If modeling is about predicting the unknown, optimization tries to answer the question of what to do, what decision to make, to get the best results out of a given situation. Sometimes that's straightforward, but sometimes... not so much. What makes an optimization problem easy or hard, and what are some of the methods for finding optimal solutions to problems? Glad you asked! May we recommend our latest podcast episode to you?

Ok, this episode is only sort of about DEADLY RADIOACTIVE GAS. It's mostly about multilevel modeling, which is a way of building models with data that has distinct, related subgroups within it. What are multilevel models used for? Elections (we can't get enough of 'em these days), understanding the effect that a good teacher can have on their students, and DEADLY RADIOACTIVE GAS. Relevant links: http://www.stat.columbia.edu/~gelman/research/published/multi2.pdf

Continuing the discussion of how polls do (and sometimes don't) tell us what to expect in upcoming elections--let's take a concrete example from the recent past, shall we? The Brexit referendum was, by and large, expected to shake out for "remain", but when the votes were counted, "leave" came out ahead. Everyone was shocked (SHOCKED!) but maybe the polls weren't as wrong as the pundits like to claim. Relevant links: http://www.slate.com/articles/news_and_politics/moneybox/2016/07/why_political_betting_markets_are_failing.html http://andrewgelman.com/2016/06/24/brexit-polling-what-went-wrong/

Not sure if you heard, but there's an election going on right now. Polls, surveys, and projections about, as far as the eye can see. How to make sense of it all? How are the projections made? Which are some good ones to follow? We'll be your trusty guides through a crash course in election forecasting. Relevant links: http://www.wired.com/2016/06/civis-election-polling-clinton-sanders-trump/ http://election.princeton.edu/ http://projects.fivethirtyeight.com/2016-election-forecast/ http://www.nytimes.com/interactive/2016/upshot/presidential-polls-forecast.html?rref=collection%2Fsectioncollection%2Fupshot&action=click&contentCollection=upshot&region=rank&module=package&version=highlights&contentPlacement=5&pgtype=sectionfront

Genomics data is some of the biggest #bigdata, and doing machine learning on it is unlocking new ways of thinking about evolution, genomic diseases like cancer, and what really makes each of us different for everyone else. This episode touches on some of the things that make machine learning on genomics data so challenging, and the algorithms designed to do it anyway.

Hot enough for you? Climate models suggest that it's only going to get warmer in the coming years. This episode unpacks those models, so you understand how they work. A lot of the episodes we do are about fun studies we hear about, like "if you're interested, this is kinda cool"--this episode is much more important than that. Understanding these models, and taking action on them where appropriate, will have huge implications in the years to come. Relevant links: https://climatesight.org/

Last week’s episode on artificial intelligence gets a huge payoff this week—we’ll explore a wonderful couple of papers about all the ways that artificial intelligence can go wrong. Malevolent actors? You bet. Collateral damage? Of course. Reward hacking? Naturally! It’s fun to think about, and the discussion starting now will have reverberations for decades to come. https://www.technologyreview.com/s/601519/how-to-create-a-malevolent-artificial-intelligence/ http://arxiv.org/abs/1605.02817 https://arxiv.org/abs/1606.06565

There’s a ton of excitement about reinforcement learning, a form of semi-supervised machine learning that underpins a lot of today’s cutting-edge artificial intelligence algorithms. Here’s a crash course in the algorithmic machinery behind AlphaGo, and self-driving cars, and major logistical optimization projects—and the robots that, tomorrow, will clean our houses and (hopefully) not take over the world…

Apple wants to study iPhone users' activities and use it to improve performance. Google collects data on what people are doing online to try to improve their Chrome browser. Do you like the idea of this data being collected? Maybe not, if it's being collected on you--but you probably also realize that there is some benefit to be had from the improved iPhones and web browsers. Differential privacy is a set of policies that walks the line between individual privacy and better data, including even some old-school tricks that scientists use to get people to answer embarrassing questions honestly. Relevant links: http://static.googleusercontent.com/media/research.google.com/en//pubs/archive/42852.pdf

Something a little different in this episode--we'll be talking about the technical plumbing that gets our podcast from our brains to your ears. As it turns out, it's a multi-step bucket brigade process of RSS feeds, links to downloads, and lots of hand-waving when it comes to trying to figure out how many of you (listeners) are out there.

Machine learning on imbalanced classes: surprisingly tricky. Many (most?) algorithms tend to just assign the majority class label to all the data and call it a day. SMOTE is an algorithm for manufacturing new minority class examples for yourself, to help your algorithm better identify them in the wild. Relevant links: https://www.jair.org/media/953/live-953-2037-jair.pdf

Conjoint analysis is like AB tester, but more bigger more better: instead of testing one or two things, you can test potentially dozens of options. Where might you use something like this? Well, if you wanted to design an entire hotel chain completely from scratch, and to do it in a data-driven way. You'll never look at Courtyard by Marriott the same way again. Relevant link: https://marketing.wharton.upenn.edu/files/?whdmsaction=public:main.file&fileID=466