Kyle interviews Prasanth Pulavarthi about the Onyx format for deep neural networks.
Kyle and Linhda discuss some high level theory of mind and overview the concept machine learning concept of catastrophic forgetting.
Sebastian Ruder is a research scientist at DeepMind. In this episode, he joins us to discuss the state of the art in transfer learning and his contributions to it.
In 2017, Facebook published a paper called Deal or No Deal? End-to-End Learning for Negotiation Dialogues. In this research, the reinforcement learning agents developed a mechanism of communication (which could be called a language) that made them able to optimize their scores in the negotiation game. Many media sources reported this as if it were a first step towards Skynet taking over. In this episode, Kyle discusses bargaining agents and the actual results of this research.
Priyanka Biswas joins us in this episode to discuss natural language processing for languages that do not have as many resources as those that are more commonly studied such as English. Successful NLP projects benefit from the availability of like large corpora, well-annotated corpora, software libraries, and pre-trained models. For languages that researchers have not paid as much attention to, these tools are not always available.
Kyle and Linh Da discuss the class of approaches called "Named Entity Recognition" or NER. NER algorithms take any string as input and return a list of "entities" - specific facts and agents in the text along with a classification of the type (e.g. person, date, place).
Kyle and Linh Da discuss the concepts behind the neural Turing machine.
In this episode, Kyle interviews Laura Edell at MS Build 2019. The conversation covers a number of topics, notably her NCAA Final 4 prediction model.
Kyle and Linhda discuss attention and the transformer - an encoder/decoder architecture that extends the basic ideas of vector embeddings like word2vec into a more contextual use case.
When users on Twitter post with geographic tags, it creates the opportunity for a variety of interesting questions to be posed having to do with language, dialects, and location. In this episode, Kyle interviews Bruno Gonçalves about his work studying language in this way.
This is an interview with Ellen Loeshelle, Director of Product Management at Clarabridge. We primarily discuss sentiment analysis.
A gentle introduction to the very high-level idea of "attention" in machine learning, as it will play a major role in some upcoming episodes over the next few weeks.
Modern messaging technology has facilitated a trend towards highly compact, short messages send by users who can presume a great amount of context held between the communicating parties. The rules of grammar may be discarded and often visible errors are a normal part of the conversation.
>>> Good mornink
Yet such short messages are also important for businesses whose users are unlikely to read a large block of text upon completing an order. Similarly, a business might want to offer assistance and effective question and answering solutions in an automated and ideally multi-lingual way. In this episode, we discuss techniques for designing solutions like that.
ELMo (Embeddings from Language Models) introduced the idea of deep contextualized word representations. It extends previous ideas like word2vec and GloVe. The ELMo model is a neural network able to map natural language into a vector space. This vector space, out of box, proved to be incredibly useful in a wide variety of seemingly unrelated NLP tasks like sentiment analysis and name entity recognition.
Bilingual evaluation understudy (or BLEU) is a metric for evaluating the quality of machine translation using human translation as examples of acceptable quality results. This metric has become a widely used standard in the research literature. But is it the perfect measure of quality of machine translation?
While at NeurIPS 2018, Kyle chatted with Liang Huang about his work with Baidu research on simultaneous translation, which was demoed at the conference.
Machine transcription (the process of translating audio recordings of language to text) has come a long way in recent years. But how do the errors made during machine transcription compare to the errors made by a human transcriber? Find out in this episode!
A sequence to sequence (or seq2seq) model is neural architecture used for translation (and other tasks) which consists of an encoder and a decoder.
The encoder/decoder architecture has obvious promise for machine translation, and has been successfully applied this way. Encoding an input to a small number of hidden nodes which can effectively be decoded to a matching string requires machine learning to learn an efficient representation of the essence of the strings.
In addition to translation, seq2seq models have been used in a number of other NLP tasks such as summarization and image captioning.
Kyle interviews Julia Silge about her path into data science, her book Text Mining with R, and some of the ways in which she's used natural language processing in projects both personal and professional.Related Links https://stack-survey-2018.glitch.me/ https://stackoverflow.blog/2017/03/28/realistic-developer-fiction/
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 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!
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.
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.
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.
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:
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.