Inside AI - November 18th, 2019

Inside AI (Nov 18th, 2019)

Deutsche Bank using AI / "Tree Hole" bot / Nvidia CEO interview


---CAMPAIGN:PREVIEW--- ###PIXEL:IMG###
Subscribe | View in browser

1. Deutsche Bank is reportedly using machine learning tools to replace employees, many of whom will lose their jobs within the next three years, Financial News reported. The bank has cut more than 4,000 positions in the last year and plans to eliminate another 18,000 over the next three years, a spokesperson confirmed. In the meantime, the bank is turning more toward AI to eliminate manual work, boost productivity, and redistribute capacity, according to Mark Matthews, head of operations for Deutsche's corporate and investment bank. Bots have processed 5 million transactions in Deutsche's corporate bank, as well as 3.4 million checks within its investment bank, in the last several years, Matthews said. As part of an initiative to save $6.6 billion in three years, the company plans to automate parts of its back office as well. - FINANCIAL NEWS

2. An AI bot can scan posts on Weibo to locate phrases that suggest a person is suicidal. After finding a post, the "Tree Hole" bot automatically alerts a group of consultants, scholars, and volunteers, who then reach out to the user. (The bot is named after the "tree holes" on Weibo, where users post secretive messages for others to read). Tree Hole was trained on words and phrases that often indicate suicide notions, then uses semantic analysis programming to better analyze the phrasing. The program's creator, Huang Zhisheng, a senior AI researcher at Vrije University of Amsterdam, said the system has prevented more than 1,000 suicides since its launch in July 2018. - SOUTH CHINA MORNING POST

3. The next-generation breakthrough in AI will be in multimodality and multidomain learning, according to Nvidia CEO Jensen Huang. In an interview with Venture Beat's Dean Takahashi, Huang said the first breakthrough came in image recognition and "now, we have a breakthrough in natural language understanding." However, Huang was quick to note the many challenges in deep learning today, which include a "very high" barrier to adoption and a surplus of many different types of models, which complicate the training process. "It’s going to get harder and harder," he said. - VENTURE BEAT

4. Since 2012, the cloud computing power required for today’s machine-learning systems has been doubling every 3.4 months, leading one writer to ask: can we afford this remarkable technology? Machine learning tasks require enormous capacity for data storage, which in turn requires enormous amounts of electricity. To contextualize what’s at risk, writer James Naughton wrote in an essay for The Guardian that researchers at Nvidia, the company that manufactures the GPU processors currently used in most machine-learning systems, developed a "massive natural-language model that was 24 times bigger than its predecessor and yet was only 34 percent better at its learning task." According to Naughton, training the final model took 512, V100 GPUs running continuously for 9.2 days, which is three times the annual energy consumption of the average American. - THE GUARDIAN 

A version of this story first appeared in Inside Cloud.

5. A new machine learning system can predict when lightning will strike down to the nearest 10 to 30 minutes, as long as it's within a certain area, according to research published this month in the journal Climate and Atmospheric Science. A team from École Polytechnique Fédérale de Lausanne in Switzerland tapped data about past lightning strikes to develop a predictive algorithm about new lightning strikes. Essentially, they trained the algorithm on weather station data involving air pressure, air temperature, relative humidity, and wind speed, which all are good indicators for future lightning strikes. In post-training tests, the model correctly predicted new strikes 80 percent of the time when covering a radius of 18.6 miles. - POPULAR MECHANICS

6. The startup Assaia installed an AI system at London's Heathrow Airport that can warn managers about potential delays before flights. The system is made up of cameras installed around aircraft stands, which monitor employees while they're performing manual checks of planes in-between flights. AI compares the footage to a schedule, detecting delays in the checks and sending alerts to a manager. Assaia took part in the Hangar 51 accelerator program, launched by International Airlines Group, British Airway’s parent company, that helps startups develop and test their technology. - CNBC

7. Skylum has incorporated some next-gen AI tools in its latest release of photography program Luminar 4, according to VentureBeat. These include a new “AI Structure” tool, which taps machine learning models to isolate and then enhance certain areas of a photo-based on their context. The less subtle AI Sky Replacement uses machine learning to locate parts of the sky in a photo and change it in accordance with the lighting. As Venture Beat describes, "An image that starts out with blue skies might shift, naturally, to yellow or pink lighting, with accurate pixel-level shifts even when foliage and other fine details are impacted by the swap." Luminar 4 is available for Mac (10.12 or higher) and Windows (7 or higher). - VENTURE BEAT

8. Stu Bradley, VP of Fraud and Security Intelligence at SAS, says there are six basic components that you need to apply machine learning to fraud detection. These include quality data, multiplicity, integration, "white-boxing" or interpretability, ongoing monitoring, and experimentation, Bradley writes in a recent article published in MIT Technology Review. Machine learning systems, when designed optimally, are great for fighting fraud because they can "learn, adapt, and uncover emerging patterns without the over-adaptation that can result in too many false positives," according to Bradley. - MIT TECH REVIEW

9. Intel unveiled a new programming model over the weekend called oneAPI. According to the company, one of the biggest challenges of AI and machine learning development is the need to tightly couple middleware and frameworks directly to specific hardware. The oneAPI model is designed "to abstract that tight coupling away, allowing developers to focus on their actual project and re-use the same code when the underlying hardware changes," according to ARS Technica, and is available at Intel Devcloud. - ARS TECHNICA

10. TechCrunch is accepting applications for early-stage AI companies to pitch at TC Sessions: Robotics + AI. The single-day event, which will include speakers, breakout sessions, and Q&As, is scheduled for March 3 at UC Berkeley’s Zellerbach Hall. For the 2020 show, the publication added a new competition that will allow early-stage startups to pitch on the main stage. Through an online application process, TechCrunch will choose 10 startups to make their initial pitch during a private event the night before TC Sessions. A panel of VC judges will choose the top five teams to present on the main stage the following day. Teams can apply online through February 1. - TECHCRUNCH

Written and curated by Beth Duckett in Orange County. Beth is a former reporter for The Arizona Republic who has written for USA Today, Get Out magazine and other publications. Follow her tweets about breaking news and other topics in southern California here.

Editor: Kim Lyons (Pittsburgh-based journalist and managing editor at Inside).

Copyright © 2020 Inside.com, All rights reserved.

Our mailing address is:
Inside.com
767 Bryant St. #203
San Francisco, CA 94107



Did someone forward this email to you? Head over to inside.com to get your very own free subscription!

You received this email because you subscribed to Inside AI. Click here to unsubscribe from Inside AI list or manage your subscriptions.

Subscribe to Inside AI