DataBytes
By Jessi & Susan
DataBytesJan 24, 2020
#50: Extreme Classification: All You Need Is Some Hash (Functions)
In part 2 of this saga on extreme classification, we get into the weeds on how MACH is able to magically handle such massive classification problems. The title says it all -- hash functions are the magical ingredient. We provide a step-by-step view of how one might come up with the MACH algorithm from first principles.
#49: Extreme Classification: Going at MACH Speed (Part 1)
In this episode, Dr. Derek Feng drops by to chat about a recent paper on a divide-and-conquer approach (Merged-Averaged Classifiers via Hashing) to massive classification problems. In part 1 (of 2 episodes), we describe the general problem solved by and strategy taken by MACH, wherein the original large classification problem is broken down into smaller-sized classification problems. Next week in the second episode, we talk about more technical details of how the division of labor works, and why it works.
#48: Where Moneyball Meets Footy
We've long heard about the waves that statistics has made in baseball. But what about soccer? In this episode, we summarize a few applications of statistics in European football (or American soccer).
#47: Domoic Acid Testing -- A Crabshoot?
Domoic acid has plagued shellfish and other wildlife along the Pacific coastline in recent years. Testing for domoic acid concentration in crabs on a regular basis has become important for determining when crabs and their viscera can be safely consumed. Unlike many other common hypothesis tests, the setup used for domoic acid testing is based on the sample maximum rather than the sample mean. In this episode, we critique the testing methodology.
#46: Finding Your (Niche) Board Games
In this episode, we discuss how two statisticians used data from BoardGameGeek.com to put together their own board game recommendation engine, specifically designed to stay away from mainstream recommendations.
#45: Learning Publicly, with Private Data
In this episode, Dr. Derek Feng discusses the general issue of data privacy in the age of big data, including topics of differential privacy and federated learning.
#44: A Conversation with Jon Krohn
We sit down with Dr. Jon Krohn to chat about his work as a Chief Data Scientist at untapt, his newly published bestseller "Deep Learning Illustrated", and his teaching/research.
#43: To Google and Back
In this episode, Professor Albert Y. Kim of Smith College describes his post-PhD journey, which included a stint at Google Adwords before academic posts at Reed College, Middlebury College, Amherst College, and Smith College.
#42: Black in the Box
Dr. Derek Feng joins us again to discuss the two metrics by which we align all statistical/machine learning methods -- interpretability versus predictive ability. In a world where black box methods reign supreme, what does learning mean?
#41: What to do with Outliers
Guest Dylan O'Connell joins us today to talk about a recent surprising, but legitimate Democratic primary poll result done by Monmouth University. We discuss different perspectives on how to approach a data point that doesn't fit in with the others.
#40: Making a DIY ML-Controlled Cat Door
Outdoor-cat owners know all too well the unpleasantries of dealing with what the cat dragged in. A self-proclaimed machine learning novice proves that you don't need to be a pro to set up a smart cat door that prevents the cat from bringing prey into your home.
#39: Rolling in the Deep Patient
We take a deep dive into the poster child for black-box machine learning methods, namely Deep Patient: an unsupervised learning method that uses denoising auto-encoders as the means for extracting salient features in electronic health records, which in turn can then be used to predict health outcomes. We do our best to explain what on earth the previous sentence meant.
#38: The Misuse of Statistics in Court
In this episode, we talk about how a statistical concept that you would learn about in an introductory course was misused in court. The error led to dire consequences in the case of Sally Clark who was charged in the deaths of two of her children.
#37: Susan Starts a New Job
In this episode, we talk about Susan's new job as a Data Scientist! She recently transitioned from academia to industry and we discuss her experience with searching for positions, interviewing, and her first few weeks in her new role.
#36: What's New in Machine Learning Startups
In this episode, we talk about some machine learning startups to pay attention to this year.
#35: You Look How You Sound
Deep learning has been useful for lots of applications when it comes to prediction. Yet another is the use of a short sound clip of speech to predict the face of the speaker.
#34: Protecting Kids' Digital Privacy
In this episode, we talk about protecting kids' digital privacy.
#33: Statisticians Hate Post-Hoc Power
Statistics is key to demonstrating the effectiveness of new advancements in science and medicine, but when statistical significance is not achieved, is post-hoc power a valid justification?
#32: Amazon's 3D Body Scan Study
In this episode, we talk about Amazon's 3D body scan study.
#31: What Data Visualizations Do You Care About? It's Personal
In this episode, we talk about how data are personal for those in a rural Pennsylvania community.
#30: Some Like It Hot -- What Gender Reveals About Our Temperature Preferences
Word on the street is that women prefer warmer temperatures than men do. Researchers designed an experiment to investigate whether this is actually true, specifically, considering how men and women perform on various cognitive tasks under different temperature scenarios. In this episode, we dissect the study so you can judge whether you believe the results.
#29: Jeopardy! Meets Statistics
Jeopardy! is a weeknightly televised trivia game show. In recent months, one player, James Holzhauer has taken the Jeopardy! fandom by storm with his unusual style of play and his long run of big wins. In this episode, we discuss how statistics can help explain his betting tactics, and we discuss how some other Jeopardy! players have used statistics to help up their game.
#28: Facial Recognition Technology Update and Rating Trustworthiness of AI-Generated Airbnb Profiles
In this episode, we discuss a number of miscellaneous news updates regarding facial recognition technology (concerning San Francisco, Amazon, and pandas!). And then, we talk about how much we trust AI-generated profiles for Airbnb.
#27: Does Uber/Lyft Help Or Hurt Traffic Congestion and Machine Learning Interpretability
In this episode, we look at a study about whether ride-sharing services contribute to increased or decreased traffic congestion in San Francisco. We then discuss some strategies to build interpretable machine learning models.
#26: Household Electronics That See and Google's Reservation AI
In this episode, we talk about a new innovation that enables household electronics to see what's around them. We then discuss Google Duplex, an AI designed to happily make reservations and appointments for you.
#25: DataFest 2019 and Measuring Migrations from Hurricane Maria
Susan recently served as a judge at a local DataFest competition (a weekend-long data competition for undergraduates). She shares her experiences and recommendations for future contestants. We then discuss how Facebook data might be helpful for counting the number of people how migrated from Puerto Rico to the mainland U.S. as a result of Hurricane Maria.
#24: Predictive Power of Early Polling and Did a TV Show Result in Higher Teenage Suicides?
In this episode, we discuss FiveThirtyEight.com's analysis of primary election polling over the past 40 years. In particular, we consider whether early polling is helpful for predicting election outcomes. And then, we talk about a study that potentially blames Netflix for a surge in teenage suicides in 2017.
#23: Offline Song Identification and Perceptions about AI
In this episode, we discuss how Google's Now Playing feature can identify songs that are playing around you, using embeddings. We then talk about a study that reports on America's perceptions about artificial intelligence -- who can we trust to develop AI responsibly?
#22: Betting on the Game of Thrones and the Misfortune of Lefthandedness
In this episode, we discuss how bookmakers price/take bets on outcomes in the Game of Thrones. We then discuss a study that claimed that lefthanded people have shorter life expectancies than righthanded people. Spoiler alert: lefthanders have nothing to worry about!
#21: Pitch Call Accuracy and Predicting the Outcome of the Champions League
Buckle up for a sports-filled episode! We discuss a study that analyzes the accuracy of umpire calls about strikes vs. balls and take a deep dive into FiveThirtyEight.com's statistical methods for predicting the winner of the Champions League.
#20: Thinking Like Computers and Text Mining the Mueller Report
In this episode, we discuss a study that recruits human researchers to try to predict how computers classify images. We then highlight a number of examples of natural language processing techniques applied to the Mueller Report.
#19: Seeing with AI and Detecting Exoplanets
In this episode, we discuss Microsoft's handy phone application for scanning and reporting on our surroundings, as a way of helping vision impaired individuals better interact with the world around them. We then talk about how AI can be useful in detecting exoplanets (or extrasolar planets).
#18: Statistical Anxiety and the Fight Against Statistical Significance
We discuss a survey designed to analyze the extent and root cause of statistical anxiety in the classroom, discussing the methods/limitations of the study. We then talk about yet another crusade against hypothesis testing, this time around the concept of "statistical significance".
#17: How Theranos Sinned Statistically
In this episode, Susan Wang is joined by guest Natalie Doss to consider the statistical sins committed by Theranos, the former blood testing unicorn. From arbitrary data manipulation to inappropriate data aggregation, we discuss what they did and why these practices were particularly bad. Then, we weigh in on how Theranos could have done worse, making it harder for the public to find out about their faulty tests.
#16: Machine-Generated Faces & Text, and Relating Health Outcomes to Skin Color
We discuss NVIDIA's AI-generated faces that look incredibly authentic, and relatedly, OpenAI's text generator that is so capable that it has to be kept under wraps. We then assess the study design of a recent research article that considered how health outcomes vary amongst African Americans of different skin tones.
#15: Deep Learning to Fold Proteins and Automated Journalism
We discuss opportunities for machines and humans in the prediction of protein structures, a necessary task in new drug discovery. Google's DeepMind has taken the prize in the recent iteration of CASP, a protein folding prediction challenge. We also discuss how AI has begun to revolutionize journalism.
#14: A Personality Test that Makes Sense and What Does Spotify Know?
538 has provided a free, online personality test that might make more sense than your typical online clickbaity quiz. We talk about why it calls itself the only personality test that isn't junk science. We then discuss the results of a recent study on Spotify data. Does it know too much about us (and you)? We'll let you know.
#13: IBM's Debate Machine and Adopting a 'Data Culture' in Companies
On February 11, IBM showcased its Project Debater in a face-off against debate champion Harish Natarajan. We talk about how this machine vs. human competition went. Then, we discuss a Harvard Business Review article citing a survey that discovered companies are not becoming data-oriented quickly enough.
#12: Super Bowl Stats, Confidence Intervals, and Data Sources
Three topics are featured in this episode: first, statistics about Super Bowl LIII, including what was in the bowls as the game happened; second, a fun activity for teaching confidence intervals; finally, we present some online sources for data.
#11: How Machines Might be Biased and the Job Market for Data Scientists
AI and ML algorithms are growing popular -- but they can actually perpetuate cognitive biases in our daily lives. We discuss the state of the problem and possible solutions. We also present a favorable job outlook for aspiring (or continuing!) data scientists.
#10: AI in Medicine and Racial Bias in College Admissions
Artificial intelligence is starting to make waves in medicine; we look at how technology might potentially change how medical testing works. We also bring in some statistical reasoning in the debate of whether or not there is racial discrimination in Harvard's college admissions process.