One of the most challenging NLP tasks is natural language understanding and reasoning. How can we construct algorithms that are able to achieve human level understanding of text and be able to answer general questions about it?

This is truly an open problem, and one with the bAbI dataset has been constructed to facilitate. bAbI presents a variety of different language understanding and reasoning tasks and exists as benchmark for comparing approaches.

In this episode, Kyle talks to Rasmus Berg Palm about his recent paper Recurrent Relational Networks

In the first half of this episode, Kyle speaks with Marc-Alexandre Côté and Wendy Tay about Text World.  Text World is an engine that simulates text adventure games.  Developers are encouraged to try out their reinforcement learning skills building agents that can programmatically interact with the generated text adventure games.

 

In the second half of this episode, Kyle interviews Kevin Patel about his paper Towards Lower Bounds on Number of Dimensions for Word Embeddings.  In this research, the explore an important question of how many hidden nodes to use when creating a word embedding.

Word2vec is an unsupervised machine learning model which is able to capture semantic information from the text it is trained on. The model is based on neural networks. Several large organizations like Google and Facebook have trained word embeddings (the result of word2vec) on large corpora and shared them for others to use.

The key algorithmic ideas involved in word2vec is the continuous bag of words model (CBOW). In this episode, Kyle uses excerpts from the 1983 cinematic masterpiece War Games, and challenges Linhda to guess a word Kyle leaves out of the transcript. This is similar to how word2vec is trained. It trains a neural network to predict a hidden word based on the words that appear before and after the missing location.

In a recent paper, Leveraging Discourse Information Effectively for Authorship Attribution, authors Su Wang, Elisa Ferracane, and Raymond J. Mooney describe a deep learning methodology for predict which of a collection of authors was the author of a given document.

The earliest efforts to apply machine learning to natural language tended to convert every token (every word, more or less) into a unique feature. While techniques like stemming may have cut the number of unique tokens down, researchers always had to face a problem that was highly dimensional. Naive Bayes algorithm was celebrated in NLP applications because of its ability to efficiently process highly dimensional data.

Of course, other algorithms were applied to natural language tasks as well. While different algorithms had different strengths and weaknesses to different NLP problems, an early paper titled Scaling to Very Very Large Corpora for Natural Language Disambiguation popularized one somewhat surprising idea. For many NLP tasks, simply providing a large corpus of examples not only improved accuracy, but it also showed that asymptotically, some algorithms yielded more improvement from working on very, very large corpora.

Although not explicitly in about NLP, the noteworthy paper The Unreasonable Effectiveness of Data emphasizes this point further while paying homage to the classic treatise The Unreasonable Effectiveness of Mathematics in the Natural Sciences.

In this episode, Kyle shares a few thoughts along these lines with Linh Da.

The discussion winds up with a brief introduction to Zipf's law. When applied to natural language, Zipf's law states that the frequency of any given word in a corpus (regardless of language) will be proportional to its rank in the frequency table.

Github is many things besides source control. It's a social network, even though not everyone realizes it. It's a vast repository of code. It's a ticketing and project management system. And of course, it has search as well.

In this episode, Kyle interviews Hamel Husain about his research into semantic code search.

This episode reboots our podcast with the theme of Natural Language Processing for the next few months.

We begin with introductions of Yoshi and Linh Da and then get into a broad discussion about natural language processing: what it is, what some of the classic problems are, and just a bit on approaches.

Finishing out the show is an interview with Lucy Park about her work on the KoNLPy library for Korean NLP in Python.

If you want to share your NLP project, please join our Slack channel.  We're eager to see what listeners are working on!

http://konlpy.org/en/latest/

 

 

Kyle shares a few thoughts on mistakes observed by job applicants and also shares a few procedural insights listeners at early stages in their careers might find value in.

Epicac by Kurt Vonnegut.

In today's episode, Kyle chats with Alexander Zhebrak, CTO of Insilico Medicine, Inc.

Insilico self describes as artificial intelligence for drug discovery, biomarker development, and aging research.

The conversation in this episode explores the ways in which machine learning, in particular, deep learning, is contributing to the advancement of drug discovery. This happens not just through research but also through software development. Insilico works on data pipelines and tools like MOSES, a benchmarking platform to support research on machine learning for drug discovery. The MOSES platform provides a standardized benchmarking dataset, a set of open-sourced models with unified implementation, and metrics to evaluate and assess their performance.

At the NeurIPS 2018 conference, Stradigi AI premiered a training game which helps players learn American Sign Language.

This episode brings the first of many interviews conducted at NeurIPS 2018.

In this episode, Kyle interviews Chief Data Scientist Carolina Bessega about the deep learning architecture used in this project.

The Stradigi AI team was exhibiting a project called the American Sign Language (ASL) Alphabet Game at the recent NeurIPS 2018 conference. They also published a detailed blog post about how they built the system found here.

 This week, Kyle interviews Scott Nestler on the topic of Data Ethics.

Today, no ubiquitous, formal ethical protocol exists for data science, although some have been proposed. One example is the INFORMS Ethics Guidelines.

Guidelines like this are rather informal compared to other professions, like the Hippocratic Oath. Yet not every profession requires such a formal commitment.

In this episode, Scott shares his perspective on a variety of ethical questions specific to data and analytics.

Kyle interviews Mick West, author of Escaping the Rabbit Hole: How to Debunk Conspiracy Theories Using Facts, Logic, and Respect about the nature of conspiracy theories, the people that believe them, and how to help people escape the belief in false information.

Mick is also the creator of metabunk.org.

The discussion explores conspiracies like chemtrails, 9/11 conspiracy theories, JFK assassination theories, and the flat Earth theory. We live in a complex world in which no person can have a sufficient understanding of all topics. It's only natural that some percentage of people will eventually adopt fringe beliefs. In this book, Mick provides a fantastic guide to helping individuals who have fallen into a rabbit hole of pseudo-science or fake news.

Fake news attempts to lead readers/listeners/viewers to conclusions that are not descriptions of reality.  They do this most often by presenting false premises, but sometimes by presenting flawed logic.

An argument is only sound and valid if the conclusions are drawn directly from all the state premises, and if there exists a path of logical reasoning leading from those premises to the conclusion.

While creating a theorem does feel to most mathematicians as a creative act of discovery, some theorems have been proven using nothing more than search.  All the "rules" of logic (like modus ponens) can be encoded into a computer program.  That program can start from the premises, applying various combinations of rules to inference new information, and check to see if the program has inference the desired conclusion or its negation.  This does seem like a mechanical process when painted in this light.  However, several challenges exist preventing any theorem prover from instantly solving all the open problems in mathematics.  In this episode, we discuss a bit about what those challenges are.

 

Fake news can be responded to with fact-checking. However, it's easier to create fake news than the fact check it.

Full Fact is the UK's independent fact-checking organization. In this episode, Kyle interviews Mevan Babakar, head of automated fact-checking at Full Fact.

Our discussion talks about the process and challenges in doing fact-checking. Full Fact has been exploring ways in which machine learning can assist in automating parts of the fact-checking process. Progress in areas like this allows journalists to be more effective and rapid in responding to new information.

In mathematics, truth is universal.  In data, truth lies in the where clause of the query.

As large organizations have grown to rely on their data more significantly for decision making, a common problem is not being able to agree on what the data is.

As the volume and velocity of data grow, challenges emerge in answering questions with precision.  A simple question like "what was the revenue yesterday" could become mired in details.  Did your query account for transactions that haven't been finalized?  If I query again later, should I exclude orders that have been returned since the last query?  What time zone should I use?  The list goes on and on.

In any large enough organization, you are also likely to find multiple copies if the same data.  Independent systems might record the same information with slight variance.  Sometimes systems will import data from other systems; a process which could become out of sync for several reasons.

For any sufficiently large system, answering analytical questions with precision can become a non-trivial challenge.  The business intelligence community aspires to provide a "single source of truth" - one canonical place where data consumers can go to get precise, reliable, and trusted answers to their analytical questions.

Fast radio bursts are an astrophysical phenomenon first observed in 2007. While many observations have been made, science has yet to explain the mechanism for these events. This has led some to ask: could it be a form of extra-terrestrial communication?

Probably not. Kyle asks Gerry Zhang who works at the Berkeley SETI Research Center about this possibility and more importantly, about his applications of deep learning to detect fast radio bursts.

Radio astronomy captures observations from space which can be converted to a waterfall chart or spectrogram. These data structures can be formatted in a visual way and also make great candidates for applying deep learning to the task of detecting the fast radio bursts.

This episode explores the root concept of what it is to be Bayesian: describing knowledge of a system probabilistically, having an appropriate prior probability, know how to weigh new evidence, and following Bayes's rule to compute the revised distribution.

We present this concept in a few different contexts but primarily focus on how our bird Yoshi sends signals about her food preferences.

Like many animals, Yoshi is a complex creature whose preferences cannot easily be summarized by a straightforward utility function the way they might in a textbook reinforcement learning problem. Her preferences are sequential, conditional, and evolving. We may not always know what our bird is thinking, but we have some good indicators that give us clues.

This is our interview with Dorje Brody about his recent paper with David Meier, How to model fake news. This paper uses the tools of communication theory and a sub-topic called filtering theory to describe the mathematical basis for an information channel which can contain fake news.

 

Thanks to our sponsor Gartner.

Without getting into definitions, we have an intuitive sense of what a "community" is. The Louvain Method for Community Detection is one of the best known mathematical techniques designed to detect communities.

This method requires typical graph data in which people are nodes and edges are their connections. It's easy to imagine this data in the context of Facebook or LinkedIn but the technique applies just as well to any other dataset like cellular phone calling records or pen-pals.

The Louvain Method provides a means of measuring the strength of any proposed community based on a concept known as Modularity. Modularity is a value in the range  that measure the density of links internal to a community against links external to the community. The quite palatable assumption here is that a genuine community would have members that are strongly interconnected.

A community is not necessarily the same thing as a clique; it is not required that all community members know each other. Rather, we simply define a community as a graph structure where the nodes are more connected to each other than connected to people outside the community.

It's only natural that any person in a community has many connections to people outside that community. The more a community has internal connections over external connections, the stronger that community is considered to be. The Louvain Method elegantly captures this intuitively desirable quality.

In this episode, our guest is Dan Kahan about his research into how people consume and interpret science news.

In an era of fake news, motivated reasoning, and alternative facts, important questions need to be asked about how people understand new information.

Dan is a member of the Cultural Cognition Project at Yale University, a group of scholars interested in studying how cultural values shape public risk perceptions and related policy beliefs.

In a paper titled Cultural cognition of scientific consensus, Dan and co-authors Hank Jenkins‐Smith and Donald Braman discuss the "cultural cognition of risk" and establish experimentally that individuals tend to update their beliefs about scientific information through a context of their pre-existing cultural beliefs. In this way, topics such as climate change, nuclear power, and conceal-carry handgun permits often result in people.

The findings of this and other studies tell us that on topics such as these, even when people are given proper information about a scientific consensus, individuals still interpret those results through the lens of their pre-existing cultural beliefs.

The ‘cultural cognition of risk’ refers to the tendency of individuals to form risk perceptions that are congenial to their values. The study presents both correlational and experimental evidence confirming that cultural cognition shapes individuals’ beliefs about the existence of scientific consensus, and the process by which they form such beliefs, relating to climate change, the disposal of nuclear wastes, and the effect of permitting concealed possession of handguns. The implications of this dynamic for science communication and public policy‐making are discussed.

A false discovery rate (FDR) is a methodology that can be useful when struggling with the problem of multiple comparisons.

In any experiment, if the experimenter checks more than one dependent variable, then they are making multiple comparisons. Naturally, if you make enough comparisons, you will eventually find some correlation.

Classically, people applied the Bonferroni Correction. In essence, this procedure dictates that you should lower your p-value (raise your standard of evidence) by a specific amount depending on the number of variables you're considering. While effective, this methodology is strict about preventing false positives (type i errors). You aren't likely to find evidence for a hypothesis that is actually false using Bonferroni. However, your exuberance to avoid type i errors may have introduced some type ii errors. There could be some hypotheses that are actually true, which you did not notice.

This episode covers an alternative known as false discovery rates. The essence of this method is to make more specific adjustments to your expectation of what p-value is sufficient evidence. 

Digital videos can be described as sequences of still images and associated audio. Audio is easy to fake. What about video?

A video can easily be broken down into a sequence of still images replayed rapidly in sequence. In this context, videos are simply very high dimensional sequences of observations, ripe for input into a machine learning algorithm.

The availability of commodity hardware, clever algorithms, and well-designed software to implement those algorithms at scale make it possible to do machine learning on video, but to what end? There are many answers, one interesting approach being the technology called "DeepFakes".

The Deep of Deepfakes refers to Deep Learning, and the fake refers to the function of the software - to take a real video of a human being and digitally alter their face to match someone else's face. Here are two examples:

• Barack Obama via Jordan Peele

• The versatility of Nick Cage

This software produces curiously convincing fake videos. Yet, there's something slightly off about them. Surely machine learning can be used to determine real from fake... right? Siwei Lyu and his collaborators certainly thought so and demonstrated this idea by identifying a novel, detectable feature which was commonly missing from videos produced by the Deep Fakes software.

In this episode, we discuss this use case for deep learning, detecting fake videos, and the threat of fake videos in the future.

In this episode, Kyle reviews what we've learned so far in our series on Fake News and talks briefly about where we're going next.

Two weeks ago we discussed click through rates or CTRs and their usefulness and limits as a metric. Today, we discuss a related metric known as quality score.

While that phrase has probably been used to mean dozens of different things in different contexts, our discussion focuses around the idea of quality score encountered in Search Engine Marketing (SEM). SEM is the practice of purchasing keyword targeted ads shown to customers using a search engine.

Most SEM is managed via an auction mechanism - the advertiser states the price they are willing to pay, and in real time, the search engine will serve users advertisements and charge the advertiser.

But how to search engines decide who to show and what price to charge? This is a complicated question requiring a multi-part answer to address completely. In this episode, we focus on one part of that equation, which is the quality score the search engine assigns to the ad in context. This quality score is calculated via several factors including crawling the destination page (also called the landing page) and predicting how applicable the content found there is to the ad itself.

Kyle interviews Steven Sloman, Professor in the school of Cognitive, Linguistic, and Psychological Sciences at Brown University. Steven is co-author of The Knowledge Illusion: Why We Never Think Alone and Causal Models: How People Think about the World and Its Alternatives. Steven shares his perspective and research into how people process information and what this teaches us about the existence of and belief in fake news.

A Click Through Rate (CTR) is the proportion of clicks to impressions of some item of content shared online. This terminology is most commonly used in digital advertising but applies just as well to content websites might choose to feature on their homepage or in search results.

A CTR is intuitively appealing as a metric for optimization. After all, if users are disinterested in some content, under normal circumstances, it's reasonable to assume they would ignore the content, rather than clicking on it. On the other hand, the best content is likely to elicit a high CTR as users signal their interest by following the hyperlink.

In the advertising world, a website could charge per impression, per click, or per action. Both impression and action based pricing have asymmetrical results for the publisher and advertiser. However, paying per click (CPC based advertising) seems to strike a nice balance. For this and other numeric reasons, many digital advertising mechanisms (such as Google Adwords) use CPC as the payment mechanism.

When charging per click, an advertising platform will value a high CTR when selecting which ad to show. As we learned in our episode on Goodhart's Law, once a measure is turned into a target, it ceases to be a good measure. While CTR alone does not entirely drive most online advertising algorithms, it does play an important role. Thus, advertisers are incentivized to adopt strategies that maximize CTR.

On the surface, this sounds like a great idea: provide internet users what they are looking for, and be awarded with their attention and lower advertising costs. However, one possible unintended consequence of this type of optimization is the creation of ads which are designed solely to generate clicks, regardless of if the users are happy with the page they visit after clicking a link.

So, at least in part, websites that optimize for higher CTRs are going to favor content that does a good job getting viewers to click it. Getting a user to view a page is not totally synonymous with getting a user to appreciate the content of a page. The gap between the algorithmic goal and the user experience could be one of the factors that has promoted the creation of fake news.

The scale and frequency with which information can be distributed on social media makes the problem of fake news a rapidly metastasizing issue. To do any content filtering or labeling demands an algorithmic solution.

In today's episode, Kyle interviews Kai Shu and Mike Tamir about their independent work exploring the use of machine learning to detect fake news.

Kai Shu and his co-authors published Fake News Detection on Social Media: A Data Mining Perspective, a research paper which both surveys the existing literature and organizes the structure of the problem in a robust way.

Mike Tamir led the development of fakerfact.org, a website and Chrome/Firefox plugin which leverages machine learning to try and predict the category of a previously unseen web page, with categories like opinion, wiki, and fake news.

If you prepared a list of creatures regarded as highly intelligent, it's unlikely ants would make the cut. This is expected, as on an individual level, ants do not generally display behavior that most humans would regard as intelligence. In fact, it might even be true that most species of ants are unable to learn. Despite this, ant colonies have evolved excellent survival mechanisms through the careful orchestration of ants.

With publications such as "Prior exposure increases perceived accuracy of fake news", "Lazy, not biased: Susceptibility to partisan fake news is better explained by lack of reasoning than by motivated reasoning", and "The science of fake news", Gordon Pennycook is asking and answering analytical questions about the nature of human intuition and fake news.

Gordon appeared on Data Skeptic in 2016 to discuss people's ability to recognize pseudo-profound bullshit.  This episode explores his work in fake news.

Today's spam filters are advanced data driven tools. They rely on a variety of techniques to effectively and often seamlessly filter out junk email from good email.

Whitelists, blacklists, traffic analysis, network analysis, and a variety of other tools are probably employed by most major players in this area. Naturally content analysis can be an especially powerful tool for detecting spam.

Given the binary nature of the problem ( or ) its clear that this is a great problem to use machine learning to solve. In order to apply machine learning, you first need a labelled training set. Thankfully, many standard corpora of labelled spam data are readily available. Further, if you're working for a company with a spam filtering problem, often asking users to self-moderate or flag things as spam can be an effective way to generate a large amount of labels for "free".

With a labeled dataset in hand, a data scientist working on spam filtering must next do feature engineering. This should be done with consideration of the algorithm that will be used. The Naive Bayesian Classifer has been a popular choice for detecting spam because it tends to perform pretty well on high dimensional data, unlike a lot of other ML algorithms. It also is very efficient to compute, making it possible to train a per-user Classifier if one wished to. While we might do some basic NLP tricks, for the most part, we can turn each word in a document (or perhaps each bigram or n-gram in a document) into a feature.

The Naive part of the Naive Bayesian Classifier stems from the naive assumption that all features in one's analysis are considered to be independent. If  and  are known to be independent, then . In other words, you just multiply the probabilities together. Shh, don't tell anyone, but this assumption is actually wrong! Certainly, if a document contains the word algorithm, it's more likely to contain the word probability than some randomly selected document. Thus, , violating the assumption. Despite this "flaw", the Naive Bayesian Classifier works remarkably will on many problems. If one employs the common approach of converting a document into bigrams (pairs of words instead of single words), then you can capture a good deal of this correlation indirectly.

In the final leg of the discussion, we explore the question of whether or not a Naive Bayesian Classifier would be a good choice for detecting fake news.

How does fake news get spread online? Its not just a matter of manipulating search algorithms. The social platforms for sharing play a major role in the distribution of fake news. But how significant of an impact can there be? How significantly can bots influence the spread of fake news?

In this episode, Kyle interviews Filippo Menczer, Professor of Computer Science and Informatics.

Fil is part of the Observatory on Social Media ([OSoMe][https://osome.iuni.iu.edu/tools/]). OSoMe are the creators of Hoaxy, Botometer, Fakey, and other tools for studying the spread of information on social media.

The interview explores these tools and the contributions Bots make to the spread of fake news.

This episode kicks off our new theme of "Fake News" with guests Robert Sheaffer and Brad Schwartz.

Fake news is a new label for an old idea. For our purposes, we will define fake news information created to deliberately mislead while masquerading as a legitimate, journalistic source of truth. It's become a modern topic of discussion as our cultures evolve to the fledgling mechanisms of communication introduced by online platforms.

What was the earliest incident of fake news? That's a question for which we may never find a satisfying answer. While not the earliest, we present a dramatization of an early example of fake news, which leads us into a discussion with UFO Skeptic Robert Sheaffer. Following that we get into our main interview with Brad Schwartz, author of Broadcast Hysteria: Orson Welles's War of the Worlds and the Art of Fake News.

We revisit the 2018 Microsoft Build in this episode, focusing on the latest ideas in DevOps. Kyle interviews Cloud Developer Advocates Damien Brady, Paige Bailey, and Donovan Brown to talk about DevOps and data science and databases.

For a data scientist, what does it even mean to “build”? Packaging and deployment are things that a data scientist doesn't normally have to consider in their day-to-day work. The process of making an AI app is usually divided into two streams of work: data scientists building machine learning models and app developers building the application for end users to consume.

DevOps includes all the parties involved in getting the application deployed and maintained and thinking about all the phases that follow and precede their part of the end solution. So what does DevOps mean for data science? Why should you adopt DevOps best practices?

In the first half, Paige and Damian share their views on what DevOps for data science would look like and how it can be introduced to provide continuous integration, delivery, and deployment of data science models. In the second half, Donovan and Damian talk about the DevOps life cycle of putting a database under version control and carrying out deployments through a release pipeline.

Logic is a fundamental of mathematical systems. It's roots are the values true and false and it's power is in what it's rules allow you to prove. Prepositional logic provides it's user variables. This episode gets into First Order Logic, an extension to prepositional logic.

In this week’s episode, our host Kyle interviews Gokula Krishnan from ETH Zurich, about his recent contributions to defenses against adversarial attacks. The discussion centers around his latest paper, titled “Defending Against Adversarial Attacks by Leveraging an Entire GAN,” and his proposed algorithm, aptly named ‘Cowboy.’

On a long car ride, Linhda and Kyle record a short episode. This discussion is about transfer learning, a technique using in machine learning to leverage training from one domain to have a head start learning in another domain.

Transfer learning has some obvious appealing features. Take the example of an image recognition problem. There are now many widely available models that do general image recognition. Detecting that an image contains a "sofa" is an impressive feat. However, for a furniture company interested in more specific details, this classifier is absurdly general. Should the furniture company build a massive corpus of tagged photos, effectively starting from scratch? Or is there a way they can transfer the learnings from the general task to the specific one.

A general definition of transfer learning in machine learning is the use of taking some or all aspects of a pre-trained model as the basis to begin training a new model which a specific and potentially limited dataset.

Thanks to our sponsor Galvanize

A Kalman Filter is a technique for taking a sequence of observations about an object or variable and determining the most likely current state of that object. In this episode, we discuss it in the context of tracking our lilac crowned amazon parrot Yoshi.

Kalman filters have many applications but the one of particular interest under our current theme of artificial intelligence is to efficiently update one's beliefs in light of new information.

The Kalman filter is based upon the Gaussian distribution. This distribution is described by two parameters:  (the mean) and standard deviation. The procedure for updating these values in light of new information has a closed form. This means that it can be described with straightforward formulae and computed very efficiently.

You may gain a greater appreciation for Kalman filters by considering what would happen if you could not rely on the Gaussian distribution to describe your posterior beliefs. If determining the probability distribution over the variables describing some object cannot be efficiently computed, then by definition, maintaining the most up to date posterior beliefs can be a significant challenge.

Kyle will be giving a talk at Skeptical 2018 in Berkeley, CA on June 10.

Today's interview is with the authors of the textbook Artificial Intelligence and Games.

Thanks to our sponsor The Great Courses.

This week's episode is a short primer on game theory.

For tickets to the free Data Skeptic meetup in Chicago on Tuesday, May 15 at the Mendoza College of Business (224 South Michigan Avenue, Suite 350), click here,

This week on Data Skeptic, we begin with a skit to introduce the topic of this show: The Imitation Game. We open with a scene in the distant future. The year is 2027, and a company called Shamony is announcing their new product, Ada, the most advanced artificial intelligence agent. To prove its superiority, the lead scientist announces that it will use the Turing Test that Alan Turing proposed in 1950. During this we introduce Turing’s “objections” outlined in his famous paper, “Computing Machinery and Intelligence.”

Following that, we talk with improv coach Holly Laurent on the art of improvisation and Peter Clark from the Allen Institute for Artificial Intelligence about question and answering algorithms.

In this episode, Kyle shares his perspective on the chatbot Eugene Goostman which (some claim) "passed" the Turing Test. As a second topic Kyle also does an intro of the Winograd Schema Challenge.

In this episode, Kyle and Linhda discuss the theory of formal languages. Any language can (theoretically) be a formal language. The requirement is that the language can be rigorously described as a set of strings which are considered part of the language. Those strings are any combination of alphabet characters in the given language.

Read more

 

The Loebner Prize is a competition in the spirit of the Turing Test.  Participants are welcome to submit conversational agent software to be judged by a panel of humans.  This episode includes interviews with Charlie Maloney, a judge in the Loebner Prize, and Bruce Wilcox, a winner of the Loebner Prize.

In this episode, Kyle chats with Vince from iv.ai and Heather Shapiro who works on the Microsoft Bot Framework. We solicit their advice on building a good chatbot both creatively and technically.

Our sponsor today is Warby Parker.

In this week’s episode, Kyle Polich interviews Pedro Domingos about his book, The Master Algorithm: How the quest for the ultimate learning machine will remake our world. In the book, Domingos describes what machine learning is doing for humanity, how it works and what it could do in the future. He also hints at the possibility of an ultimate learning algorithm, in which the machine uses it will be able to derive all knowledge — past, present, and future.

What's the best machine learning algorithm to use? I hear that XGBoost wins most of the Kaggle competitions that aren't won with deep learning. Should I just use XGBoost all the time? That might work out most of the time in practice, but a proof exists which tells us that there cannot be one true algorithm to rule them.

For a long time, physicians have recognized that the tools they have aren't powerful enough to treat complex diseases, like cancer. In addition to data science and models, clinicians also needed actual products — tools that physicians and researchers can draw upon to answer questions they regularly confront, such as “what clinical trials are available for this patient that I'm seeing right now?” In this episode, our host Kyle interviews guests Alex Grigorenko and Iker Huerga from Memorial Sloan Kettering Cancer Center to talk about how data and technology can be used to prevent, control and ultimately cure cancer.

In a previous episode, we discussed Markov Decision Processes or MDPs, a framework for decision making and planning. This episode explores the generalization Partially Observable MDPs (POMDPs) which are an incredibly general framework that describes most every agent based system.

In many real world situations, a person/agent doesn't necessarily know their own objectives or the mechanics of the world they're interacting with. However, if the agent receives rewards which are correlated with the both their actions and the state of the world, then reinforcement learning can be used to discover behaviors that maximize the reward earned.

Formally, an MDP is defined as the tuple containing states, actions, the transition function, and the reward function. This podcast examines each of these and presents them in the context of simple examples.  Despite MDPs suffering from the curse of dimensionality, they're a useful formalism and a basic concept we will expand on in future episodes.

Last week on Data Skeptic, we visited the Laboratory of Neuroimaging, or LONI, at USC and learned about their data-driven platform that enables scientists from all over the world to share, transform, store, manage and analyze their data to understand neurological diseases better. We talked about how neuroscientists measure the brain using data from MRI scans, and how that data is processed and analyzed to understand the brain. This week, we'll continue the second half of our two-part episode on LONI.

In artificial intelligence, the term 'agent' is used to mean an autonomous, thinking agent with the ability to interact with their environment. An agent could be a person or a piece of software. In either case, we can describe aspects of the agent in a standard framework.

This episode kicks off the next theme on Data Skeptic: artificial intelligence.  Kyle discusses what's to come for the show in 2018, why this topic is relevant, and how we intend to cover it.

We break format from our regular programming today and bring you an excerpt from Max Tegmark's book "Life 3.0".  The first chapter is a short story titled "The Tale of the Omega Team".  Audio excerpted courtesy of Penguin Random House Audio from LIFE 3.0 by Max Tegmark, narrated by Rob Shapiro.  You can find "Life 3.0" at your favorite bookstore and the audio edition via penguinrandomhouseaudio.com.

Kyle will be giving a talk at the Monterey County SkeptiCamp 2018.

This episode features an interview with Rigel Smiroldo recorded at NIPS 2017 in Long Beach California.  We discuss data privacy, machine learning use cases, model deployment, and end-to-end machine learning.

When computers became commodity hardware and storage became incredibly cheap, we entered the era of so-call "big" data. Most definitions of big data will include something about not being able to process all the data on a single machine. Distributed computing is required for such large datasets.

Getting an algorithm to run on data spread out over a variety of different machines introduced new challenges for designing large-scale systems. First, there are concerns about the best strategy for spreading that data over many machines in an orderly fashion. Resolving ambiguity or disagreements across sources is sometimes required.

This episode discusses how such algorithms related to the complexity class NC.

In this week's episode, Scott Aaronson, a professor at the University of Texas at Austin, explains what a quantum computer is, various possible applications, the types of problems they are good at solving and much more. Kyle and Scott have a lively discussion about the capabilities and limits of quantum computers and computational complexity.

I sat down with Ali Ghodsi, CEO and found of Databricks, and John Chirapurath, GM for Data Platform Marketing at Microsoft related to the recent announcement of Azure Databricks.

When I heard about the announcement, my first thoughts were two-fold.  First, the possibility of optimized integrations with existing Azure services.  This would be a big benefit to heavy Azure users who also want to use Spark.  Second, the benefits of active directory to control Databricks access for large enterprise.

Hear Ali and JG's thoughts and comments on what makes Azure Databricks a novel offering.

 

In this episode we discuss the complexity class of EXP-Time which contains algorithms which require $O(2^{p(n)})$ time to run.  In other words, the worst case runtime is exponential in some polynomial of the input size.  Problems in this class are even more difficult than problems in NP since you can't even verify a solution in polynomial time.

We mostly discuss Generalized Chess as an intuitive example of a problem in EXP-Time.  Another well-known problem is determining if a given algorithm will halt in k steps.  That extra condition of restricting it to k steps makes this problem distinct from Turing's original definition of the halting problem which is known to be intractable.

Algorithms with similar runtimes are said to be in the same complexity class. That runtime is measured in the how many steps an algorithm takes relative to the input size.

The class P contains all algorithms which run in polynomial time (basically, a nested for loop iterating over the input).  NP are algorithms which seem to require brute force.  Brute force search cannot be done in polynomial time, so it seems that problems in NP are more difficult than problems in P.  I say it "seems" this way because, while most people believe it to be true, it has not been proven.  This is the famous P vs. NP conjecture.  It will be discussed in more detail in a future episode.

Given a solution to a particular problem, if it can be verified/checked in polynomial time, that problem might be in NP.  If someone hands you a completed Sudoku puzzle, it's not difficult to see if they made any mistakes.  The effort of developing the solution to the Sudoku game seems to be intrinsically more difficult.  In fact, as far as anyone knows, in the general case of all possible examples of the game, it seems no strategy can do better on average than just random guessing.

This notion of random guessing the solution is where the N in NP comes from: Non-deterministic.  Imagine a machine with a random input already written in its memory.  Given enough such machines, one of them will have the right answer.  If they all ran in parallel, one of them could verify it's input in polynomial time.  This guess / provided input is often called a witness string.

NP is an important concept for many reasons.  To me, the most reason to know about NP is a practical one.  Depending on your goals or the goals of your employer, there are many challenging problems you may attempt to solve.  If a problem you are trying to solve happens to be in NP, then you should consider the implications very carefully.  Perhaps you'll be lucky and discover that your particular instance of the problem is easy.  Sudoku is pretty easy if only 2 remaining squares need to be filled in.  The traveling salesman problem is easy to solve if you live in a country where all roads for a ring with exactly one road in and out.

If the problem you wish to solve is not trivial, or if you will face many instances of the problem and expect some will not be trivial, then it's unlikely you'll be able to find the exact solution.  Sure, maybe you can grab a bunch of commodity servers and try to scale the heck out of your attempt.  Depending on the problem you're solving, that might just work.  If you can out-purchase your problem in computing power, then problems in NP will surrender to you.  But if your input size ever grows, it's unlikely you'll be able to keep up.

If your problem is intractable in this way, all is not lost.  You might be able to find an approximate solution to your problem.  Good enough is better than no solution at all, right?  Most of the time, probably.  However, some tremendous work has also been done studying topics like this.  Are there problems which are not even approximable in polynomial time?  What approximation techniques work best?  Alas, those answers lie elsewhere.

This episode avoids a discussion of a few key points in order to keep the material accessible.  If you find this interesting, you should next familiarize yourself with the notions of NP-Complete, NP-Hard, and co-NP.  These are topics we won't necessarily get to in future episodes.  Michael Sipser's Introduction to the Theory of Computation is a good resource.

 

In this episode, Professor Michael Kearns from the University of Pennsylvania joins host Kyle Polich to talk about the computational complexity of machine learning, complexity in game theory, and algorithmic fairness. Michael's doctoral thesis gave an early broad overview of computational learning theory, in which he emphasizes the mathematical study of efficient learning algorithms by machines or computational systems.

When we look at machine learning algorithms they are almost like meta-algorithms in some sense. For example, given a machine learning algorithm, it will look at some data and build some model, and it’s going to behave presumably very differently under different inputs. But does that mean we need new analytical tools? Or is a machine learning algorithm just the same thing as any deterministic algorithm, but just a little bit more tricky to figure out anything complexity-wise? In other words, is there some overlap between the good old-fashioned analysis of algorithms with the analysis of machine learning algorithms from a complexity viewpoint? And what is the difference between strategies for determining the complexity bounds on samples versus algorithms?

A big area of machine learning (and in the analysis of learning algorithms in general) Michael and Kyle discuss is the topic known as complexity regularization. Complexity regularization asks: How should one measure the goodness of fit and the complexity of a given model? And how should one balance those two, and how can one execute that in a scalable, efficient way algorithmically? From this, Michael and Kyle discuss the broader picture of why one should care whether a learning algorithm is efficiently learnable if it's learnable in polynomial time.

Another interesting topic of discussion is the difference between sample complexity and computational complexity. An active area of research is how one should regularize their models so that they're balancing the complexity with the goodness of fit to fit their large training sample size.

As mentioned, a good resource for getting started with correlated equilibria is: https://www.cs.cornell.edu/courses/cs684/2004sp/feb20.pdf

Thanks to our sponsors:

Mendoza College of Business - Get your Masters of Science in Business Analytics from Notre Dame.

brilliant.org - A fun, affordable, online learning tool.  Check out their Computer Science Algorithms course.

TMs are a model of computation at the heart of algorithmic analysis.  A Turing Machine has two components.  An infinitely long piece of tape (memory) with re-writable squares and a read/write head which is programmed to change it's state as it processes the input.  This exceptionally simple mechanical computer can compute anything that is intuitively computable, thus says the Church-Turing Thesis.

Attempts to make a "better" Turing Machine by adding things like additional tapes can make the programs easier to describe, but it can't make the "better" machine more capable.  It won't be able to solve any problems the basic Turing Machine can, even if it perhaps solves them faster.

An important concept we didn't get to in this episode is that of a Universal Turing Machine.  Without the prefix, a TM is a particular algorithm.  A Universal TM is a machine that takes, as input, a description of a TM and an input to that machine, and subsequently, simulates the inputted machine running on the given input.

Turing Machines are a central idea in computer science.  They are central to algorithmic analysis and the theory of computation.

How long an algorithm takes to run depends on many factors including implementation details and hardware.  However, the formal analysis of algorithms focuses on how they will perform in the worst case as the input size grows.  We refer to an algorithm's runtime as it's "O" which is a function of its input size "n".  For example, O(n) represents a linear algorithm - one that takes roughly twice as long to run if you double the input size.  In this episode, we discuss a few everyday examples of algorithmic analysis including sorting, search a shuffled deck of cards, and verifying if a grocery list was successfully completed.

Thanks to our sponsor Brilliant.org, who right now is featuring a related problem as their Brilliant Problem of the Week.

One Shot Learning is the class of machine learning procedures that focuses learning something from a small number of examples.  This is in contrast to "traditional" machine learning which typically requires a very large training set to build a reasonable model.

In this episode, Kyle presents a coded message to Linhda who is able to recognize that many of these new symbols created are likely to be the same symbol, despite having extremely few examples of each.  Why can the human brain recognize a new symbol with relative ease while most machine learning algorithms require large training data?  We discuss some of the reasons why and approaches to One Shot Learning.

Recommender systems play an important role in providing personalized content to online users. Yet, typical data mining techniques are not well suited for the unique challenges that recommender systems face. In this episode, host Kyle Polich joins Dr. Joseph Konstan from the University of Minnesota at a live recording at FARCON 2017 in Minneapolis to discuss recommender systems and how machine learning can create better user experiences. 

Thanks to our sponsor brilliant.org/dataskeptics

A Long Short Term Memory (LSTM) is a neural unit, often used in Recurrent Neural Network (RNN) which attempts to provide the network the capacity to store information for longer periods of time. An LSTM unit remembers values for either long or short time periods. The key to this ability is that it uses no activation function within its recurrent components. Thus, the stored value is not iteratively modified and the gradient does not tend to vanish when trained with backpropagation through time.

Zillow is a leading real estate information and home-related marketplace. We interviewed Andrew Martin, a data science Research Manager at Zillow, to learn more about how Zillow uses data science and big data to make real estate predictions.

Our guest Pranav Rajpurkar and his coauthored recently published Cardiologist-Level Arrhythmia Detection with Convolutional Neural Networks, a paper in which they demonstrate the use of Convolutional Neural Networks which outperform board certified cardiologists in detecting a wide range of heart arrhythmias from ECG data.

RNNs are a class of deep learning models designed to capture sequential behavior.  An RNN trains a set of weights which depend not just on new input but also on the previous state of the neural network.  This directed cycle allows the training phase to find solutions which rely on the state at a previous time, thus giving the network a form of memory.  RNNs have been used effectively in language analysis, translation, speech recognition, and many other tasks.

Thanks to our sponsor Springboard.

In this week's episode, guest Andre Natal from Mozilla joins our host, Kyle Polich, to discuss a couple exciting new developments in open source speech recognition systems, which include Project Common Voice.

In June 2017, Mozilla launched a new open source project, Common Voice, a novel complementary project to the TensorFlow-based DeepSpeech implementation. DeepSpeech is a deep learning-based voice recognition system that was designed by Baidu, which they describe in greater detail in their research paper. DeepSpeech is a speech-to-text engine, and Mozilla hopes that, in the future, they can use Common Voice data to train their DeepSpeech engine.

A Bayesian Belief Network is an acyclic directed graph composed of nodes that represent random variables and edges that imply a conditional dependence between them. It's an intuitive way of encoding your statistical knowledge about a system and is efficient to propagate belief updates throughout the network when new information is added.

In this episode, Tony Beltramelli of UIzard Technologies joins our host, Kyle Polich, to talk about the ideas behind his latest app that can transform graphic design into functioning code, as well as his previous work on spying with wearables.

In statistics, two random variables might depend on one another (for example, interest rates and new home purchases). We call this conditional dependence. An important related concept exists called conditional independence. This phrase describes situations in which two variables are independent of one another given some other variable.

For example, the probability that a vendor will pay their bill on time could depend on many factors such as the company's market cap. Thus, a statistical analysis would reveal many relationships between observable details about the company and their propensity for paying on time. However, if you know that the company has filed for bankruptcy, then we might assume their chances of paying on time have dropped to near 0, and the result is now independent of all other factors in light of this new information.

We discuss a few real world analogies to this idea in the context of some chance meetings on our recent trip to New York City.

Animals can't tell us when they're experiencing pain, so we have to rely on other cues to help treat their discomfort. But it is often difficult to tell how much an animal is suffering. The sheep, for instance, is the most inscrutable of animals. However, scientists have figured out a way to understand sheep facial expressions using artificial intelligence.

On this week's episode, Dr. Marwa Mahmoud from the University of Cambridge joins us to discuss her recent study, "Estimating Sheep Pain Level Using Facial Action Unit Detection." Marwa and her colleague's at Cambridge's Computer Laboratory developed an automated system using machine learning algorithms to detect and assess when a sheep is in pain. We discuss some details of her work, how she became interested in studying sheep facial expression to measure pain, and her future goals for this project.

If you're able to be in Minneapolis, MN on August 23rd or 24th, consider attending Farcon. Get your tickets today via https://farcon2017.eventbrite.com.

This episode discusses the vanishing gradient - a problem that arises when training deep neural networks in which nearly all the gradients are very close to zero by the time back-propagation has reached the first hidden layer. This makes learning virtually impossible without some clever trick or improved methodology to help earlier layers begin to learn.

hen faced with medical issues, would you want to be seen by a human or a machine? In this episode, guest Edward Choi, co-author of the study titled Doctor AI: Predicting Clinical Events via Recurrent Neural Network shares his thoughts. Edward presents his team’s efforts in developing a temporal model that can learn from human doctors based on their collective knowledge, i.e. the large amount of Electronic Health Record (EHR) data.

In a neural network, the output value of a neuron is almost always transformed in some way using a function. A trivial choice would be a linear transformation which can only scale the data. However, other transformations, like a step function allow for non-linear properties to be introduced.

Activation functions can also help to standardize your data between layers. Some functions such as the sigmoid have the effect of "focusing" the area of interest on data. Extreme values are placed close together, while values near it's point of inflection change more quickly with respect to small changes in the input. Similarly, these functions can take any real number and map all of them to a finite range such as [0, 1] which can have many advantages for downstream calculation.

In this episode, we overview the concept and discuss a few reasons why you might select one function verse another.

This episode recaps the Microsoft Build Conference.  Kyle recently attended and shares some thoughts on cloud, databases, cognitive services, and artificial intelligence.  The episode includes interviews with Rohan Kumar and David Carmona.

 

Max-pooling is a procedure in a neural network which has several benefits. It performs dimensionality reduction by taking a collection of neurons and reducing them to a single value for future layers to receive as input. It can also prevent overfitting, since it takes a large set of inputs and admits only one value, making it harder to memorize the input. In this episode, we discuss the intuitive interpretation of max-pooling and why it's more common than mean-pooling or (theoretically) quartile-pooling.

This episode is an interview with Tinghui Zhou.  In the recent paper "Unsupervised Learning of Depth and Ego-motion from Video", Tinghui and collaborators propose a deep learning architecture which is able to learn depth and pose information from unlabeled videos.  We discuss details of this project and its applications.

CNNs are characterized by their use of a group of neurons typically referred to as a filter or kernel.  In image recognition, this kernel is repeated over the entire image.  In this way, CNNs may achieve the property of translational invariance - once trained to recognize certain things, changing the position of that thing in an image should not disrupt the CNN's ability to recognize it.  In this episode, we discuss a few high-level details of this important architecture.

Despite the success of GANs in imaging, one of its major drawbacks is the problem of 'mode collapse,' where the generator learns to produce samples with extremely low variety.

To address this issue, today's guests Arnab Ghosh and Viveka Kulharia proposed two different extensions. The first involves tweaking the generator's objective function with a diversity enforcing term that would assess similarities between the different samples generated by different generators. The second comprises modifying the discriminator objective function, pushing generations corresponding to different generators towards different identifiable modes.

GANs are an unsupervised learning method involving two neural networks iteratively competing. The discriminator is a typical learning system. It attempts to develop the ability to recognize members of a certain class, such as all photos which have birds in them. The generator attempts to create false examples which the discriminator incorrectly classifies. In successive training rounds, the networks examine each and play a mini-max game of trying to harm the performance of the other.

In addition to being a useful way of training networks in the absence of a large body of labeled data, there are additional benefits. The discriminator may end up learning more about edge cases than it otherwise would be given typical examples. Also, the generator's false images can be novel and interesting on their own.

The concept was first introduced in the paper Generative Adversarial Networks.

Recently, we've seen opinion polls come under some skepticism.  But is that skepticism truly justified?  The recent Brexit referendum and US 2016 Presidential Election are examples where some claims the polls "got it wrong".  This episode explores this idea.

No reliable, complete database cataloging home sales data at a transaction level is available for the average person to access. To a data scientist interesting in studying this data, our hands are complete tied. Opportunities like testing sociological theories, exploring economic impacts, study market forces, or simply research the value of an investment when buying a home are all blocked by the lack of easy access to this dataset. OpenHouse seeks to correct that by centralizing and standardizing all publicly available home sales transactional data. In this episode, we discuss the achievements of OpenHouse to date, and what plans exist for the future.

Check out the OpenHouse gallery.

I also encourage everyone to check out the project Zareen mentioned which was her Harry Potter word2vec webapp and Joy's project doing data visualization on Jawbone data.

Guests

Thanks again to @iamzareenf, @blueplastic, and @joytafty for coming on the show. Thanks to the numerous other volunteers who have helped with the project as well!

Announcements and details

• If you're interested in getting involved in OpenHouse, check out the OpenHouse contributor's quickstart page.

• Kyle is giving a machine learning talk in Los Angeles on May 25th, 2017 at Zehr.

Sponsor

Thanks to our sponsor for this episode Periscope Data. The blog post demoing their maps option is on our blog titled Periscope Data Maps.

To start a free trial of their dashboarding too, visit http://periscopedata.com/skeptics

Kyle recently did a youtube video exploring the Data Skeptic podcast download numbers using Periscope Data. Check it out at https://youtu.be/aglpJrMp0M4.

Supplemental music is Lee Rosevere's Let's Start at the Beginning.

 

There's more than one type of computer processor. The central processing unit (CPU) is typically what one means when they say "processor". GPUs were introduced to be highly optimized for doing floating point computations in parallel. These types of operations were very useful for high end video games, but as it turns out, those same processors are extremely useful for machine learning. In this mini-episode we discuss why.

Backpropagation is a common algorithm for training a neural network.  It works by computing the gradient of each weight with respect to the overall error, and using stochastic gradient descent to iteratively fine tune the weights of the network.  In this episode, we compare this concept to finding a location on a map, marble maze games, and golf.

 

In this week's episode of Data Skeptic, host Kyle Polich talks with guest Maura Church, Patreon's data science manager. Patreon is a fast-growing crowdfunding platform that allows artists and creators of all kinds build their own subscription content service. The platform allows fans to become patrons of their favorite artists- an idea similar the Renaissance times, when musicians would rely on benefactors to become their patrons so they could make more art. At Patreon, Maura's data science team strives to provide creators with insight, information, and tools, so that creators can focus on what they do best-- making art.

On the show, Maura talks about some of her projects with the data science team at Patreon. Among the several topics discussed during the episode include: optical music recognition (OMR) to translate musical scores to electronic format, network analysis to understand the connection between creators and patrons, growth forecasting and modeling in a new market, and churn modeling to determine predictors of long time support.

A more detailed explanation of Patreon's A/B testing framework can be found here

Other useful links to topics mentioned during the show:

OMR research

Patreon blog

Patreon HQ blog

Amanda Palmer

Fran Meneses

Feed Forward Neural Networks

In a feed forward neural network, neurons cannot form a cycle. In this episode, we explore how such a network would be able to represent three common logical operators: OR, AND, and XOR. The XOR operation is the interesting case.

Below are the truth tables that describe each of these functions.

AND Truth Table

Input 1	Input 2	Output
0	0	0
0	1	0
1	0	0
1	1	1

OR Truth Table

Input 1	Input 2	Output
0	0	0
0	1	1
1	0	1
1	1	1

XOR Truth Table

Input 1	Input 2	Output
0	0	0
0	1	1
1	0	1
1	1	0

The AND and OR functions should seem very intuitive. Exclusive or (XOR) if true if and only if exactly single input is 1. Could a neural network learn these mathematical functions?

Let's consider the perceptron described below. First we see the visual representation, then the Activation function , followed by the formula for calculating the output.

 

 

 

 

Can this perceptron learn the AND function?

Sure. Let and

What about OR?

Yup. Let and

An infinite number of possible solutions exist, I just picked values that hopefully seem intuitive. This is also a good example of why the bias term is important. Without it, the AND function could not be represented.

How about XOR?

No. It is not possible to represent XOR with a single layer. It requires two layers. The image below shows how it could be done with two laters.

 

 

In the above example, the weights computed for the middle hidden node capture the essence of why this works. This node activates when recieving two positive inputs, thus contributing a heavy penalty to be summed by the output node. If a single input is 1, this node will not activate.

Universal approximation theorem tells us that any continuous function can be tightly approximated using a neural network with only a single hidden layer and a finite number of neurons. With this in mind, a feed forward neural network should be adaquet for any applications. However, in practice, other network architectures and the allowance of more hidden layers are empirically motivated.

Other types neural networks have less strict structal definitions. The various ways one might relax this constraint generate other classes of neural networks that often have interesting properties. We'll get into some of these in future mini-episodes.

 

Check out our recent blog post on how we're using Periscope Data cohort charts.

Thanks to Periscope Data for sponsoring this episode. More about them at periscopedata.com/skeptics

In this Data Skeptic episode, Kyle is joined by guest Ruggiero Cavallo to discuss his latest efforts to mitigate the problems presented in this new world of online advertising. Working with his collaborators, Ruggiero reconsiders the search ad allocation and pricing problems from the ground up and redesigns a search ad selling system. He discusses a mechanism that optimizes an entire page of ads globally based on efficiency-maximizing search allocation and a novel technical approach to computing prices.

Today's episode overviews the perceptron algorithm. This rather simple approach is characterized by a few particular features. It updates its weights after seeing every example, rather than as a batch. It uses a step function as an activation function. It's only appropriate for linearly separable data, and it will converge to a solution if the data meets these criteria. Being a fairly simple algorithm, it can run very efficiently. Although we don't discuss it in this episode, multi-layer perceptron networks are what makes this technique most attractive.

DataRefuge is a public collaborative, grassroots effort around the United States in which scientists, researchers, computer scientists, librarians and other volunteers are working to download, save, and re-upload government data. The DataRefuge Project, which is led by the UPenn Program in Environmental Humanities and the Penn Libraries group at University of Pennsylvania, aims to foster resilience in an era of anthropogenic global climate change and raise awareness of how social and political events affect transparency.

 

If a CEO wants to know the state of their business, they ask their highest ranking executives. These executives, in turn, should know the state of the business through reports from their subordinates. This structure is roughly analogous to a process observed in deep learning, where each layer of the business reports up different types of observations, KPIs, and reports to be interpreted by the next layer of the business. In deep learning, this process can be thought of as automated feature engineering. DNNs built to recognize objects in images may learn structures that behave like edge detectors in the first hidden layer. Proceeding layers learn to compose more abstract features from lower level outputs. This episode explore that analogy in the context of automated feature engineering.

Linh Da and Kyle discuss a particular image in this episode. The image included below in the show notes is drawn from the work of Lee, Grosse, Ranganath, and Ng in their paper Convolutional Deep Belief Networks for Scalable Unsupervised Learning of Hierarchical Representations.

 

In this episode, I speak with Raghu Ramakrishnan, CTO for Data at Microsoft.  We discuss services, tools, and developments in the big data sphere as well as the underlying needs that drove these innovations.

In this episode, we talk about a high-level description of deep learning.  Kyle presents a simple game (pictured below), which is more of a puzzle really, to try and give  Linh Da the basic concept.

 

 

Thanks to our sponsor for this week, the Data Science Association. Please check out their upcoming Dallas conference at dallasdatascience.eventbrite.com

Versioning isn't just for source code. Being able to track changes to data is critical for answering questions about data provenance, quality, and reproducibility. Daniel Whitenack joins me this week to talk about these concepts and share his work on Pachyderm. Pachyderm is an open source containerized data lake.

During the show, Daniel mentioned the Gopher Data Science github repo as a great resource for any data scientists interested in the Go language. Although we didn't mention it, Daniel also did an interesting analysis on the 2016 world chess championship that complements our recent episode on chess well. You can find that post here

Supplemental music is Lee Rosevere's Let's Start at the Beginning.

 

Thanks to Periscope Data for sponsoring this episode. More about them at periscopedata.com/skeptics

 

 

 

Logistic Regression is a popular classification algorithm. In this episode, we discuss how it can be used to determine if an audio clip represents one of two given speakers. It assumes an output variable (isLinhda) is a linear combination of available features, which are spectral bands in the discussion on this episode.

 

Keep an eye on the dataskeptic.com blog this week as we post more details about this project.

 

Thanks to our sponsor this week, the Data Science Association.  Please check out their upcoming conference in Dallas on Saturday, February 18th, 2017 via the link below.

 

dallasdatascience.eventbrite.com

 

Prior work has shown that people's response to competition is in part predicted by their gender. Understanding why and when this occurs is important in areas such as labor market outcomes. A well structured study is challenging due to numerous confounding factors. Peter Backus and his colleagues have identified competitive chess as an ideal arena to study the topic. Find out why and what conclusions they reached.

Our discussion centers around Gender, Competition and Performance: Evidence from Real Tournaments from Backus, Cubel, Guid, Sanchez-Pages, and Mañas. A summary of their paper can also be found here.