Last Week in AI #131: new AI play, gigantic neural nets, undocumented healthcare AI tools
The new play 'AI', Cerebras' new AI training system that could train networks with >100 trillion parameters, studies show healthcare AI tools do not meet rigor needed for professional use, and more!
Last week, a new play titled AI was staged at London’s Young Vic theater. Directed by Jennifer Tang, the play had a unique aspect: instead of performing an existing script, its aim was to showcase the writing of a new play through the collaboration between a team of humans and the AI model GPT-3. The team of writers could query GPT-3 for ideas, and then honed and combined its outputs into a new 30 minute play that the play’s actors performed to close out the night. According to one review, the results were entirely unlike a traditional play, and highlighted the need for AI-human collaboration:
“Unlike a traditional play, whose scenes are weaved together in a very deliberate way, the performance at the end of AI was formed of loosely-connected vignettes created by GPT-3, which constructed new scenes without a memory of its previous inventions. …
In this performance, theatre’s least celebrated heroes took center stage. And far from simply deferring to AI, they found ways to manipulate it as a creative instrument, to generate sketches that could be chiseled into something altogether more worthy.”
At this year's IEEE Hot Chips conference, Cerebras Systems unveiled their new neural network training systems. The new system consists of four parts that together will enable linear imporvements in network training speed as more chips are added, and it can in theory train networks with as many as 120 trillion parameters. To put that in perspective, GPT-3 has only 175 billion parameters. Three of the four parts are hardware improvements that allow Cerebras to efficiently stream large amounts of network parameters and data to their large compute chips, each with 40GB of on-chip memory. The other improvement has to do with taking better advantage of sparsity in many neural networks --- avoiding doing computations whose results are easily known (like multiplying by zeros). These developments pave a way for bigger and more complex AI training systems in the future, potentially achieving "brain-scale" compute, which many put at the scale of 100 trillion parameters.
Generalization is a well-known issue in machine learning: can a model trained on a particular dataset generalize to unseen data? How will it perform in the "real world?" The answer, for hospital models, is not so good. University of Michigan researchers found that a widely used artificial intelligence model for spotting early signs of sepsis missed about two-thirds of actual cases and raised many false alarms. Delving into the broader problem, a new (and not yet peer-reviewed) study from Stanford researchers reveals that AI models in healthcare are not documented with anywhere near the rigor or accuracy deemed necessary by medical and AI professionals. 90% of the models examined, all developed by EPIC systems, did adhere to a dozen of the most common recommendations for documentation. But "the models complied with barely 40 percent of the total 220 individual recommendations across all 15 guidelines." As with models more generally, it seems that medical AI systems could do with a dose of transparency.
Microsoft AI Open-Source The Code For It’s Focal Transformer - "Vision Transformer has shown great promise at various computer vision tasks. The ability to capture short and long-range visual dependencies through the self-attention model is exciting, but it brings challenges due to quadratic computational overhead."
DeepMind Open-Sources Perceiver IO, A General-Purpose Deep Learning Model Architecture That Handles A Wide Range of Data and Tasks - "Recently, DeepMind has open-sourced Perceiver IO–a general-purpose deep learning model architecture that can handle many different types of inputs and outputs."
Challenges and Opportunities in NLP Benchmarking - "Over the last years, models in NLP have become much more powerful, driven by advances in transfer learning. A consequence of this drastic increase in performance is that existing benchmarks have been left behind."
Google Brain Uncovers Representation Structure Differences Between CNNs and Vision Transformers - "A Google Brain research team explores the internal representation structures of ViTs and CNNs on image classification tasks, providing insights on key differences between the two approaches."
IBM introduces Telum chips aimed at AI inferencing workloads like fraud detection - "Big Blue has unveiled Telum, its first chip with AI inferencing acceleration that will allow it to conduct tasks such as fraud detection while a transaction is occurring."
Synopsys CEO: AI-designed chips will generate 1,000X performance in 10 years - "Automation has been part of chip design since the 1980s. But now AI-designed chips are producing great results that could lead to 1,000 times better performance for chips in the next decade, according to Synopsys CEO Aart de Geus."
An Army of Grain-harvesting Robots Marches Across Russia - "The field of automated precision agriculture is based on one concept—autonomous driving technologies that guide vehicles through GPS navigation."
Astrobee Will Find Astronauts’ Lost Socks - "It'll be up to robots to keep space stations clean and functional while humans are away"
More than 50 robots are working at Singapore's high-tech hospital - "In Singapore's Changi General Hospital, there's a chance your surgeon won't have a heart. The cleaners might not have lungs, and the physiotherapist could be completely brainless."
Google Sheets’ formula suggestions are like autofill for math - "Google has announced that Google Sheets is getting the ability to intelligently suggest formulas and functions for your spreadsheet, based on the data you're trying to analyze."
A.I. for A.I.—US Patent Regulator Uses Machine Learning to Analyze Complex A.I. Patents - "The number of A.I. patent applications received by the USPTO doubled from 2002 to 2018. Markus Winkler/UnsplashEvery year, the U.S. Patent and Trademark Office (USPTO) reviews thousands of patent applications involving artificial intelligence (A.I."
Waymo starts offering autonomous rides in San Francisco - "Waymo is going to start shuttling a wider group of passengers around in its autonomous vehicles in San Francisco, California — though they’ll have to sign nondisclosure agreements, and there still will be a human safety driver behind the wheel."
Fired From Google After Critical Work, AI Researcher Mitchell to Join Startup - "The former co-head of Google’s Ethical AI research group, Margaret Mitchell, who was fired in February after a controversy over a critical paper she co-authored, will join artificial intelligence startup Hugging Face to create tools that help companies make sure their algorithms are fair."
Musk says Tesla's self-driving software update 'not great' - "Elon Musk said on Monday said that the electric-car maker was working on improving the much-awaited update to its self-driving software "as fast as possible.""
Sweetgreen buys robotic-restaurant chain Spyce as it prepares to go public - "Restaurant unicorn Sweetgreen, valued at $1.8 billion, said Tuesday it plans to buy the robotics-focused restaurant concept Spyce. The bowl and salad chain, backed by celebrity chef Daniel Boulud, opened in Boston in 2018 with an automated kitchen that replaced cooks."
I don’t know what to do with those tossed salads and robot legs - "One of the most fascinating aspects of Boston Dynamics’ transition into a commercial organization is watching the company — and its partners — figure out real-world jobs for Spot."
The Secret Bias Hidden in Mortgage-Approval Algorithms - " In fact, high-earning Black applicants with less debt were rejected more often than high-earning White applicants who have more debt."
Toyota pauses Paralympics self-driving buses after one hits visually impaired athlete - "Toyota has apologised for the “overconfidence” of a self-driving bus after it ran over a Paralympic judoka in the athletes’ village and said it would temporarily suspend the service."
Now That Machines Can Learn, Can They Unlearn? - "Companies of all kinds use machine learning to analyze people’s desires, dislikes, or faces. Some researchers are now asking a different question: How can we make machines forget?"
Real Talk: Intersectionality and AI - "In 1989, Kimberlé Crenshaw, now a law professor at UCLA and the Columbia School of Law, first proposed the concept of intersectionality. In an article published in the University of Chicago Legal Forum, she critiqued the inability of the law to protect working Black women against discrimination."
Benefits outweigh risks for autonomous vehicles - if they are regulated - "An interdisciplinary panel of experts has assessed the risks and potential benefits associated with deploying autonomous vehicles (AVs) on U.S. roads and predicts that the benefits will substantially outweigh potential harms -- but only if the AVs are well regulated."
US government agencies plan to increase their use of facial recognition technology - "A 90-page report published Tuesday by the US Government Accountability Office (GAO) details how federal agencies currently use, and plan to expand their use of, facial recognition systems. Ten of 24 agencies surveyed plan to broaden their use of the technology by 2023."
Tax not the robots - "While the arguments in favor of a robot tax may be well-intentioned, robot taxes are a misguided idea that would have negative consequences for firms, their workers, and ultimately the economy."
Open Source Is Throwing AI Policymakers For A Loop - "Machine learning isn't just for big companies any more"
Computers suck at ‘common sense’ — AI expert explains why - "An independent news and commentary website produced by academics and journalists."
The Bouba/Kiki Effect And Sound Symbolism In CLIP - "The bouba/kiki effect is the phenomenon where humans show a preference for certain mappings between shapes and their corresponding labels/sounds. The above image of 2 theoretical objects is shown to a participant who is then asked which one is called a ‘bouba’ and which is called a ‘kiki’."
Staying visual in a CLIP world - "When CLIP based image synthesis started to emerge in January/February 2021, I was torn between curiousity and perplexity. Suddenly it was possible to synthesise images from mere text prompts."
Copyright © 2021 Skynet Today, All rights reserved.