DataDownload: How the WSJ Is Using Deep Learning to Inform Content Strategy, Robots solving Rubik’s Cubes & more
Well — that was quite a week in tech-land. I have to be honest, I’ve never been good at the Rubik’s Cube. So watching a robot nail it is, well, humbling to say the least. Speaking of humbling, Ben Silbermann had a crazy idea for a curation site back when he was looking for angel investors. One of my good friends sent him to me. He invested, I passed. Ugh. Now Pinterest is doing great AI work — and Ben continues to innovate.
Today’s newsletter is chock full of amazing breakthroughs, and emerging trends. AMC — the theatre chain — is into streaming. Hmmm…. but if you only watch one video this week, don’t miss “Unsung” presenting at the NYC Media Lab Summit. Glenn Cantave and Idris Brewster are amazing and inspirational.
So — enjoy, and as always send along thoughts, feedback, questions, or critiques. DataDownload is written each week for you!
Steven Rosenbaum
Managing Director
The NYC Media Lab
Must-Read
If a Robotic Hand Solves a Rubik’s Cube, Does It Prove Something?
“In order [for OpenAI] to keep their operation going, this is what they have to do. It is their life blood.” — Zachary Lipton , professor in the machine learning group at Carnegie Mellon
OpenAI spent several months training a robotic hand (which looks like a modified version of last year’s impressive hand ) to solve a Rubik’s Cube. Besides +1 dexterity for robotkind (when it doesn’t drop the cube 8/10 times), what are the cascading benefits for the field, besides making for great headlines?
Well, the headlines are part of it — New York Times notes that the project was a way for OpenAI to promote itself “as it seeks to attract money and talent.” And robots have been able to solve the Rubik’s Cubes in under a second . What’s special about this hand is that it learned to solve the cube on its own, even dealing with randomness, like having two of its fingers tied.
This is the kind of dexterity and “intelligence” that is required to sort through a bin of items in a warehouse — something robots haven’t been able to do reliably yet, and something way more profitable than a Rubik’s Cube machine.
5 min read
How Pinterest Built One of Silicon Valley’s Most Successful Algorithms
Pinterest, which attracts 50M new users per year, has one of the least-scrutinized AI algorithms in Silicon Valley, and is the only tech unicorn who’s stock has traded consistently above its IPO price. One main reason Pinterest is able to maintain its reputation is its self-honesty — it understands the inherent bias built into its system, admits it, and is actively trying to mitigate it. The social network’s mantra comes down to three statements that place it starkly against other social giants: “embrace bias, limit virality, and become something of an anti-social network.”
Medium publication OneZero explores how Pinterest is achieving a delicate balance of automated feed curation, inclusivity, and customizability, with detailed descriptions of the algorithm’s features. One great example is how the system weighs “saves” vs. “pins”: “People don’t really save an inflammatory article about the president, [instead they click it], but they do save an outfit they want to buy in the future. So we’re biasing toward those types of interactions.”
12 min read
For the Media
How the Wall Street Journal Is Using Deep Learning to Inform Content Strategy
By using established ML methods like Doc2Vec , WSJ was able to gain a deeper level of insight into their content. The paper clustered over 20k of their articles into 361 “hyper-granular clusters,” located those with high conversion rates (topics that led readers to subscribe), and combined this data with attributes like story length and scroll depth to formulate content strategies.
To get these tools in the hands of editors, WSJ’s lead technologist John West , data scientist Mark Secada , and ML scientist Eric Bolton built an explorer tool to examine subsets of content data. Since the ML tools used to build the explorer are open source, WSJ notes that any newsroom (with a data scientist) can do it.
4 min read
Fighting fake news is excruciatingly hard. To demonstrate, MIT CSAIL researchers released two papers on how fake news detectors could be duped, building on research from last year that tried to differentiate accurate from politically prejudiced sources.
The first paper demonstrates how fake news detectors can’t tell fake from real text if both are machine generated (in this case, by OpenAI’s GPT-2). The frustrations don’t end there.
In the second paper , the team employed the Fact Extraction and VERification ( FEVER ) fact-checking dataset, which they discovered had its own set of biases and other issues. When the researchers “de-biased” FEVER, their fact-checking model performed poorly and they had to create a new algorithm entirely. Steady, but slow progress.
4 min read
Events & Announcements
Event: The Work Awesome Conference
Date: October 24th
Location: Grant Thornton, NYC
We’d like to invite you to Work Awesome, a boutique all-day conference about the future of work on Thursday, October 24th in Midtown Manhattan. Speakers include the global head of people & leadership at Siemens, Pooja Anand, the executive transforming Danone into the world’s largest B-corp, Lorna Davis, and the co-creator of Woodstock, Michael Lang! Also, you’ll witness the launch of The Workies, a brand new awards program for people leaders.
NYC Media Lab has you covered with three complimentary tickets! Just click here and see if you got lucky. Otherwise, you can use the code MediaLabAwesome at checkout and get $100 off the ticket price. Buy Tickets Here
Internship: Cyber NYC Inventors to Founders’ Student Venture Associate
Cyber NYC Inventors to Founders, a $100M public-private partnership, is excited to launch our Student Venture Associate program. This is a paid internship program for NYC university students (undergrad and graduate) to work on our Cyber NYC Inventors to Founders team during the fall and spring semesters. Register Here
Event: All Tech Is Human NYC
Date: November 9, 9AM
Location: Thoughtworks, Madison Ave
An all-day ethical tech summit with 200 technologists, academics, advocates, students, org leaders, artists, designers, policymakers, and YOU. Join for an impactful mix of lightning talks, topical panels, strategy sessions, and tech/humanity art performance. Register Here
Event: Natural Language, Dialog and Speech (NDS) Symposium
Date: November 22, 9AM-6PM
Location: The New York Academy of Sciences
NDS2019 will convene leading researchers from academia and industry to discuss cutting-edge methodologies and computational approaches to applied and theoretical problems in dialog systems, spoken and natural language understanding, natural language generation, and speech synthesis. Register Here
What We’re Watching
Unsung, an Augmented Reality Storybox
UNSUNG is an interactive, multiplayer AR learning experience. Students can read through passages about female icons of color and answer multiple choice questions within a mobile app. Correct answers will unlock different rooms related to the subject’s lives that students can explore.
8 min watch
A Deeper Look
The State of Machine Learning Frameworks in 2019
In 2019, the war of ML frameworks has boiled down to a head-to-head battle between the industry’s two biggest players — PyTorch and TensorFlow. Cornell University student Horace He believes Facebook’s PyTorch will soon gain an edge over Google’s open source ML framework. In less than two years, PyTorch use has gone from underdog to the choice framework in conference papers (see graph below). He chronicles how PyTorch caught up to TensorFlow’s rich feature set, but in the end, the lesson is more expansive:
“It is perhaps under appreciated how much machine learning frameworks shape ML research. They don’t just enable machine learning research, they enable and restrict the ideas that researchers are able to easily explore. How many nascent ideas are crushed simply because there is no easy way to express them in a framework?”
15 min read
At some point, most of our faces became part of the public domain when they got captured by a government camera. That data could be used by AI systems to scope out your location in the future. With facial recognition systems already being deployed in US airports and against Hong Kong protestors , designers, academics, and activists are finding novel ways to protect our privacy — including stylish, CV-thwarting fashion accessories.
Business Insider’s visual essay collates a number of slightly bizarre, macabre-looking accessories that confuse facial recognition systems, including the below-pictured lens-shaped mask, an inverted projector mask, goggles fitted with LEDs (by far the most normal looking of the bunch), avant-garde makeup, headscarfs printed with images of faces, body-worn graphic prints… and this .
5 min read
Originally published at https://mailchi.mp.
