Green neural net training, deepfake geography, and more!

Last Week in AI #114

May 3, 2021

DigitalGlobe high-resolution imagery of Hackley Park in Muskegon, Michigan. — Image Credits: DigitalGlobe via Getty Images

Mini Briefs

Google and UC Berkeley Propose Green Strategies for Large Neural Network Training

It has long been known that popular deep learning models pose a significant environmental cost, and there have been many calls for these costs to be made more transparent and mitigated. Following the line of research investigating this issue, a research team from Google and UC Berkeley examined the energy use and carbon footprint of popular large-scale models including T5 and GPT-3. In their recent paper, they introduce reduction strategies and endorse previous appeals for publication norms intended to make energy use and CO2 emissions of computationally intensive ML models more transparent. Their investigation of emissions discovers that with prudent processor, hardware, and datacenter choices, the carbon footprint of deep neural networks can be reduced by up to ~100-1000 times.

Deepfake satellite imagery poses a not-so-distant threat, warn geographers

We've heard of deepfaked ads and videos, but what about deepfake geography? Geographers are worried about the spread of AI-generated satellite imagery. Such fakery could spread hoaxes about disasters, discredit stories based on real satellite imagery, disrupt military exercises, and become a national security issue. Bo Zhao, assistant professor of geography at the University of Washington, published a paper with colleagues on "deep fake geography" which included experiments generating and detecting this imagery. Zhao aims to raise awareness of the potential influence of deepfaked geography, given that faked geographical imagery is nothing new and such AI-generated imagery is convincing enough to fool the untrained eye.

Podcast

Check out our weekly podcast covering these stories!
Website | RSS | iTunes | Spotify | YouTube

News

Advances & Business

Farming Robot Kills 100,000 Weeds per Hour With Lasers - "Carbon Robotics has unveiled the third-generation of its Autonomous Weeder, a smart farming robot that identifies weeds and then destroys them with high-power lasers."
Lyft sells self-driving unit to Toyota’s Woven Planet for $550M - "Ride-hailing company Lyft has sold off its autonomous vehicle unit to Toyota’s Woven Planet Holdings subsidiary for $550 million, the latest in a string of acquisitions spurred by the cost and lengthy timelines to commercialize the technology that have occurred within the autonomous vehicle industry."
Harvard, UCLA Researchers Use AI to Analyze Passion in Public Speaking - "Sesh is announcing a new collaborative project with the Harvard Business School and the UCLA Anderson School of Management that is using the firm’s advanced AI video analysis tools to measure how the passion of public speakers affects perception, attitudes and behaviors."
Roam Robotics introduces a smart knee brace - "There are lot of companies out there making robotic exoskeletons. In fact, it’s one of the more active categories in the space — and for good reason. These sorts of technologies have the ability to profoundly impact the future of how people work, move and rehabilitate."
AI Startup Founded and Funded by "Dream Teams" has Recent Raise leading to $5B Plus Valuation in Three plus Years - "Not often does the term 'Dream Team' refer to something outside of sports, but if one were to think of it in terms of a tech startup, SambaNova System’s founders and funders fit the bill."
Huawei trained the Chinese-language equivalent of GPT-3 - "For the better part of a year, OpenAI’s GPT-3 has remained among the largest AI language models ever created, if not the largest of its kind."
Facebook details self-supervised AI that can segment images and videos - "Facebook today announced that it developed an algorithm in collaboration with Inria called DINO that enables the training of transformers, a type of machine learning model, without labeled training data."
Vectra AI Raises $130 Million at $1.2 Billion Valuation - "The investment is coming from the growth unit at Blackstone Group (ticker: BX) growth unit, which will have a minority stake in Vectra AI. The company didn’t disclose other investors of the F round, saying only that existing investors participated in the round."
Google, Tel Aviv University launch joint program, ‘AI for Social Good’ - "Google and Tel Aviv University recently launched a program for promoting artificial intelligence-related multidisciplinary research for the benefit of society."
Transformers in Vision - "What have Vision Transformers been up to? Initially introduced with the now-famous Attention is all you need, the Transformer has dominated the field of Natural Language Processing (NLP) for years."

Concerns & Hype

We were promised Strong AI, but instead we got metadata analysis - "The late nineties dream of search engines was that they would use grand-scale Artificial Intelligence to find everything, understand most of it and help us retrieve the best of it. Not much of that has really come true. Google has always performed a wide crawl of the entire web."
Robots are animals, not humans - "In the early 2000s, a Russian man named Boris Zhurid struck a deal to sell the Iranians a large collection of weaponry. He chartered a transport aircraft to make the delivery from Sevastopol, the largest city on the Crimean Peninsula, in the Black Sea, to the Persian Gulf."
The EU’s New Proposed Rules on A.I. Are Missing Something - "Thus far, most attempts at making policy for artificial intelligence have fallen into one of two camps: either outright bans on certain applications of machine learning—for instance, the facial recognitions bans passed in a few cities in the United States—or very broad, high-level principles that offer no concrete guidance or specific rules."

That's all for this week! If you are not subscribed and liked this, feel free to subscribe below!

Last Week in AI