Discover more from Last Week in AI
Skynet This Week #9: Recycled DeepFakes, running robots, invisible elephants, and more!
Skynet This Week #9
By Viraat Aryabumi and Arnav Arora
Our bi-weekly quick take on a bunch of the most important recent media stories about AI for the period 10th September 2018 - 24th September 2018
Advances & Business
Byron Spice, CMU AI
Deep Fakes have certainly been causing a lot of concern recently, but they do have limitations - thus far it has been tricky to use the existing algorithms to generate high quality outputs with a lot of data. As summarized in this article, CMU’s new Recycle-GAN algorithm reduces some of those limitations, and once again increases the need for algorithms to be developed to detect algorithmically modified media.
Cade Metz, The New York Times
“Boston Dynamics’s robot only looks like it thinks for itself.”
A good summary of the state of the company Boston Dynamics - already famous for its impressive walking and trotting robots, but not yet a successful business. With its Spot Mini going on sale next year, it will be interesting to see if the technology is mature enough to have useful applications for generating revenue.
James Vincent, The Verge
Another day, another novel application of recent AI algorithms - this time, for controlling drones to make use of air currents similarly to birds. The methods are not ready for deployment yet, but mark yet another area in which AI could prove useful.
Natasha Lomas, Techcrunch
IBM has launched a new open source toolkit to detect and mitigate bias in datasets. The tool is said to increase interpretability of the results produced by AI algorithms as well as, in some cases, compliance with policies such as GDPR. How much the toolkit is welcomed by the AI research community remains to be seen.
Concerns & Hype
Kevin Hartnett, Quanta Magazine
Researchers from Toronto have developed a new adversarial attack for object detection systems - literally introducing an elephant into a picture of a room. The inability of current computer vision systems to go back and recheck images is a shortcoming that needs to be addressed to build robust systems. Effectively, computer vision systems will have to figure out how to do a double take.
Eliza Strickland, IEEE Spectrum
“the big AI question isn’t whether China or the United States will dominate. Instead it’s how we’ll deal with the “real AI crisis” of job losses, wealth inequality, and people’s sense of self-worth.”
Kai-Fu Lee recently wrote a new book titled AI Superpowers: China, Silicon Valley, and the New World Order, and discussed the ideas presented in it with IEEE Spectrum in this interview.
”Previous technology shifts have not had as negative effects on employment as was first feared.”
A new report from University of California, Berkeley in collaboration with Tata Communications suggests that with the rise in AI, the job satisfaction of ordinary employees will be higher, contrary to the long held belief that AI will render these people without jobs.
“Job satisfaction will be enhanced by the elimination of mundane tasks, giving people time to be more creative.”
Derek Mead, Motherboard
A journalist describes his experiments with a tool that claims to analyze a person’s personality and employability through their twitter feed. He dissects the output of the tool for his feed and embellishes by providing psychological insights, possible shortcomings, and biases in using a tool such as this.
Analysis & Policy
Will Knight, MIT Technology Review
“China might be at loggerheads with the United States over trade, but it is calling for a friendlier approach to the development of artificial intelligence.”
After an initially aggressive stance towards competing with the US for domination in the space of AI, China is now communicating a more pro-collaboration stance.
Kalev Leetaru, Forbes
“Deep learning systems learn from the training data available to them”
To move towards systems that are more robust to bias in image datasets, Google launched the Inclusive Images Challenge which promotes development of algorithms that are less prone to be biased even when skewed data is used. Such a challenge helps generalisation for underrepresented geographies and cultures and is a step towards solving an important problem in AI today.
Rebecca Hill, The Register
“We need to be very careful that if these new technologies are put into day-to-day practices, they don’t create new gaming and target cultures,”
The Royal United Services Institute (RUSI) a defense and security think tank published a report on the use of machine learning in police decision making. The report says that it is hard to predict the impact of ML-driven tools and algorithmic bias. It seems, however, that police in the UK continue to use these tools.
Expert Opinions & Discussion within the field
Kai-Fu Lee, The Wall Street Journal
This essay, an adapted excerpt from Kai-Fu Lee’s upcoming book, talks about the short-term impact that AI will have in terms of job losses. Issues relating to feelings of obsolescence, the need for a basic-income like fund and jobs that will likely not be replaced by AI are also discussed.
Gillian Hadfield, Techcrunch
Successful and safe AI that achieves our goals within the limits of socially accepted norms requires an understanding of not only how our physical systems behave, but also how human normative systems behave.
This essay argues that to ensure that AI systems are aligned with human goals they need to understand cultural norms.
Tambet Matiisen, Computational Neuroscience Lab, University of Tartu
This blog post explains the use of embeddings in OpenAI Five’s network architecture.
Yoel Zeldes, Another Datum
Variational Autoencoders (VAE) are widely used to generate new examples similar to the dataset they are trained on. This post goes into the theory behind the inner workings of a VAE.