Our bi-weekly quick take on a bunch of the most important recent media stories about AI for the period 08/13/18 - 08/27/18.
Advances & Business
Katyanna Quach, The Register
A good summary of the match between OpenAI Five, a Dota bot and a top 20 professional team that took place at The International competition this past week.
Michael Kan, PC Mag
Researchers at UC Berkeley have developed a system that is able to transfer the pose and motion from one person to another. A fun application of this is to render images of anyone performing the dance moves of a ballet dancer or Bruno Mars.
A Waymo engineer told us why a virtual-world simulation is crucial to the future of self-driving cars
Matthew DeBord, Business Insider
While testing in the real world is very important to validate Self-Driving technology, a lot more training and testing happens in a simulated world. Business Insider takes an inside look into how one engineer’s 20% project became Carcraft, Waymo’s simulated world where the company’s cars rack up 8 million miles everyday.
Concerns & Hype
Simone Stolzoff, Quartz
While there are lots of benefits of integrating AI technology in schools, these systems also raise privacy and consent concerns for students. Unclear to what extent ‘AI’ is actually involved in these systems, but certainly a sign of things to come.
Tom Simonite, Wired
An article highlighting the potential of AI systems trained to optimize a particular objective to learn ‘cheats’ – shortcuts that don’t actually accomplish the objective but do technically optimize the metric it is measured by.
“The specimens collected by Krakovna and fellow bug hunters point to a communication problem between humans and machines: Given a clear goal, an algorithm can master complex tasks, such as beating a world champion at Go. But even with logical parameters, it turns out that mathematical optimization empowers bots to develop shortcuts humans didn’t think to deem off-limits. Teach a learning algorithm to fish, and it might just drain the lake.”
Julia Gong, Skynet Today
We take a look at the hype and misconceptions surrounding Google Translate.
Analysis & Policy
Pedro Domingos, Scientific American
A well made case that we ought not worry about AI’s impact on our future so much, because “Artificial intelligence is just the ability to solve hard problems—a task that does not require free will… Like any other technology, AIs will always be extensions of us.”
James Vincent, The Verge
Tom White, a lecturer in computational design at the University of Wellington in New Zealand, uses a “Perception Engine” to make abstract shapes that computer vision systems identify as objects from the ImageNet dataset. James Vincent from The Verge lookx at the ways in which White’s algorithmic art is being perceived by researchers, artists and the general public.
Janelle Shane, aiweirdness
A recent paper developed a system that can generate images from text descriptions. While it works well when trained only on a dataset of bird images, when trained on a larger dataset with various objects the images generated are weird, to say the least.
Expert Opinions & Discussion within the field
Ana Marasović, The Gradient
A summary of the evidence that many learned NLP systems face dont really learn to understand language, and how NLP researchers suggest this may be fixed.
Joel Grus, JupyterCon 2018
Joel Grus, a researcher at the Allen Institute of AI, gave a talk at JupyterCon 2018 on why he thinks Jupyter Notebooks are not the best solution to all things data science and machine learning. Whether you agree with him or not, his slides are thought-provoking (with amazing memes) and share some good ideas about how to structure projects.
Evan Pu, Medium
A neat explainer in layman’s terms about how OpenAI Five Dota bot works along with fundamental flaws and limitations of the methods used.
Mihir Garimella, Prathik Naidu, The Gradient
A summary of various approaches to learning from 3D data, which self driving cars and robotics in general need to be able to do. Turns out, this is much harder than the much more mature problem of learning from 2D images.
Caroline Chan, Shiry Ginosar, Tinghui Zhou, Alexei A. Efros, UC Berkeley
That’s all for this digest! If you are not subscribed and liked this, feel free to subscribe below!