Deep INFOMAX, Image to Image Translation, FEVER, Perception Engines, QuAC, Best 150 ML Tutorials,…

Great day and welcome to the 27th Issue of the NLP Newsletter! I am Elvis from Belize, Editor of, and a PhD researcher in AI and NLP. Here is this week’s notable NLP news: lots of QA tutorials and data releases; more on AI and health applications; unsupervised representation learning; fake news detection dataset; uncertainty in deep learning; and much more.

On People…

Several research institutions and startups, in collaboration with OpenAI, are building an AI that can learn curiosity without any guidance by only feeding it TV shows — link

Perception Engines are a series of machine vision algorithms used to create different art prints that represent abstract shapes and the machines’ understanding of the world — link

Joel Grus shares his thoughts on why he doesn’t like notebooks for programming (slides) — link

According to new report, NLP and AI are being combined to help transform healthcare from physical interaction to high-quality clinical service free from physical co-location of the doctor. Essentially, NLP can be used to mine unstructured documents and make fast and efficient suggestions around clinical decision support — link

Rwanda young technology entrepreneurs, together with the Rwanda ICT Chamber, are organizing a massive event dubbed “Building machine NLP for Kinyarwanda” to build NLP technology that can speak their language — link

On Research…

A recent paper presents a method for motion transfer using what’s called a “per-frame image-to-image translation with spatio-temporal smoothing”. The authors also created a demonstration that received huge attention on YouTube showing how professional dancing moves can be transferred to a target (amateur dancer) using pose detection — link

Deep INFOMAX is a proposed method to learn unsupervised representations with the training objective to maximize mutual information between part or all of the input and a high-level feature vector. The idea is that representations should be learned from the information content or a structural constraint rather than all the bits at once — decision that matters at the semantic level — link

Volodymyr Kuleshov presents on the topic of estimating uncertainty in deep learning methods, which is crucial in many applications such as computer vision, healthcare, and robotics (slides) — link

The Gradient releases new article where Ana Marasović discusses how different research strategies can be used to tackle several of the limitations encountered in current deep learning methods for NLP — link

A health organization leveraged weakly supervised learning approaches to label large clinical datasets; then they were able to use deep learning, computer vision techniques to perform disease detection in CT scans, 150 times faster than human radiologists — link releases new article discussing an overview of the recent trends in deep learning for natural language processing — link

Learn more about frame-semantics and semantic role labeling in this nice tutorial by Swabha Swayamdipta (presented at COLING 2018) — link

On Code and Data…

Alexa scientist releases FEVER, a public dataset containing 185,000 data instances for fact extraction and verification — link

Question Answering in Context (QuAC), is a new dataset consisting of data instances representing dialog between two crowd workers. It is useful for “modeling, understanding, and participating in information seeking dialog” — link

This article gives a light introduction on how to perform named-entity recognition (NER), considered a challenging and important NLP task, using Wikipedia — link

New dataset released by Rada Mihalcea can be used to build fake news detection systems — link

Shervine Amidi releases neat website containing several deep learning and machine learning cheat-sheets — link

Check out this nice detailed tutorial on constructing a visual question answering system from FloydHub — link

On Industry…

OpenAI bots lose to Dota 2 human professional players at The International 2018 — link

Will conversational AI become mainstream? According to a new research, these types of technologies will save businesses more than $8 billion annually by 2022 — link | link

Learn how data scientists at Networked Insights make sense of billions of daily social media posts by categorizing text information with over 30,000 classifications — link

How question answering systems are built at LinkedIn (slides presented at KDD 2018) — link

According to a new report, New York is evolving to become the capital of a booming artificial intelligence industry — link, an Indonesian AI startup, gets seed funding to build more robust NLP/NLU technology to help businesses deploy their chatbot strategy — link

Quotes of the day…

Just a great idea for an NLP project by Professor Yoav Goldberg — link

Illustrations of the day…

Learn how to classify AI technologies based on this neat AI Knowledge Map — link

Worthy Mentions…

Learn about the science behind voice interfaces — link

Fake sentence detecting based on a simple training task for encoding — link

The NLP Newsletter (Issue 26) — link

Learn how multi-lingual text processing is done in this very friendly tutorial — link

A comprehensive list of some of the best machine learning, NLP, and Python tutorials by Robbie Allen — link

Tutorial on how to deploy machine learning models with Flask, Docker, and Google Cloud (presented at PyconGhana 2018) — link

If you spot any errors or inaccuracies in this newsletter please leave a comment below. If you have any questions regarding any of the headline above, reach out to me @omarsar0.