Last Week in AI #156: DeepMind applies AI to nuclear fusion reactors, Texas sues Meta over facial recognition, SenseTime expands into manufacturing, and more!
How DeepMind uses Deep RL to control fusion plasma, Texas sues Meta over capturing facial identifiers without consent, China's SenseTime to use ML to detect defects in manufacturing
DeepMind released a paper last week on their research that applies Deep Reinforcement Learning (Deep RL) to control tokamak nuclear fusion reactors. For a bit of background: Deep RL trains a neural network that can make sequences of decisions in order to optimize some objective. Tokamak fusion reactors have a doughnut-shaped vacuum that contains a plasma, where the fusion reaction happens. Magnetic coils around the reactor are used to move this plasma to a desired shape, and also away from the reactor walls.
Manually writing software to sculpt the plasma is very hard due to the complexity of plasma physics, so using a learning-based approach can help. However, Deep RL algorithms typically need a lot of trial-and-error data. This is infeasible to obtain in the tokamak reactor DeepMind had access to, which can only maintain a plasma for 3 seconds before a 15 mins cool-down period.
To address this, researchers at EPFL developed a very accurate plasma simulator, and DeepMind trained its RL algorithm using this simulation data. While this is still challenging to do (the simulator is slow, and it doesn’t 100% match the real reactor), the trained RL agent did learn how to achieve a variety of plasma shapes, and it even worked on the real reactor.
Our take: This is an impressive demonstration of a real-world application of Deep RL. One interesting takeaway is that while there are some algorithmic modifications, the underlying RL algorithm is rather commonplace. What was really important was the accurate simulator and the system that allowed the RL agent to run at a speed fast enough to control the reactor. It is exciting that nuclear fusion is among the first real-world applications where RL is making an impact, and hopefully, there will be many more in the future.
The attorney general of Texas filed a lawsuit against Meta last week, with the claim that Facebook’s use of facial recognition violated that state’s privacy protections for personal biometric data. The suit claims Facebook is guilty of “tens of millions of violations”of Texas law and that Facebook may owe civil penalties in the hundreds of billions of dollars. There is precedent for such a lawsuit from the past:
Facebook previously settled another lawsuit over its facial-recognition practices for about $650 million. That class-action suit filed in 2015 was brought under Illinois’s biometric privacy law, which is similar in some respects to the Texas law. Both laws require individuals’ consent before their biometric identifiers can be captured.
Our take: It’s slightly surprising this happened only now, and not long ago. It seems that the outcome of the prior Illinois lawsuit, and Facebook’s failure to dismiss it, got the gears moving to set this in action. I am quite interested to see how this will play out.
Chinese AI giant SenseTime has made its name focusing on facial recognition systems, but its latest initiative focuses on the industry inspection market. In a statement on February 14, the company announced SenseSpring, an AI-enabled automated defect detection system. Its customer is Foton Cummings, the world's leading independent engine manufacturer. The project represents the upgrading of industrial practices with modern technology and will allow SenseTime to compete with other giants such as Baidu, which has launched its own AI-enabled product inspection services.
Our take: China has led combining AI and automated inspection technology, so SenseTime's foray into the scene is merely a new chapter in an already unfolding story. The market seems like an important one all around the world, with US-based companies also using AI-based systems for automated defect detection. Perhaps more "unsexy" uses for computer vision technology will help alleviate the questionable reputation it has developed for uses like facial recognition.
Machine Learning Becomes a Mathematical Collaborator - "Two recent collaborations between mathematicians and DeepMind demonstrate the potential of machine learning to help researchers generate new mathematical conjectures. Mathematicians often work together when they’re searching for insight into a hard problem."
Good News About the Carbon Footprint of Machine Learning Training - "We demonstrate four key practices that reduce the carbon (and energy) footprint of ML workloads by large margins, which we have employed to help keep ML under 15% of Google’s total energy use."
Latest LinkedIn AI Research on Using Bayesian Optimization for Balancing Metrics in Recommendation Systems - "The majority of large-scale recommender systems, such as newsfeed ranking, people recommendations, job recommendations, and so on, have many objectives that must be optimized simultaneously. User involvement, diversity, novelty, freshness, and justice are examples of these objectives."
Recognizing Chemical Formulas from Research Papers Using a Transformer-Based Artificial Neural Network - "In the last few years, deep learning has been playing an integral role in various scientific and technology areas. This development promotes AI-based tools that can help us with information retrieval."
A New Study from CMU and Bosch Center for AI Demonstrated a New Transformer Paradigm in Computer Vision - "After leveraging Convolutional Neural Network (CNN) for many years, since the advent of Transformers in Natural Language Processing (NLP), the computer vision community has also focused very carefully (and sometimes trying a little too hard) on how to apply Transformers in vision."
Hybrid AI: A new way to make machine minds that really think like us - "In the quest to make artificial intelligence that can reason and apply knowledge flexibly, many researchers are focused on fresh insights from neuroscience. Should they be looking to psychology too?"
China's Race Towards AI Research Dominance - "As our chart based on data provided by the OECD.AI project shows, China is well on its way to surpassing traditional artificial intelligence research powerhouses in the upcoming years."
This AI Generates Fake New Pokémon, And It's Actually Pretty Good - "A machine-learning engineer has harnessed a powerful machine-learning technology to develop a web application that lets you create images of fake pocket monsters. The results are hilarious, and I love how simple the tool is."
Who Is Behind QAnon? Linguistic Detectives Find Fingerprints - "Using machine learning, separate teams of computer scientists identified the same two men as likely authors of messages that fueled the viral movement. "
AI generated this moody, entertaining animated music video - "Aza Raskin used OpenAI tools to render a film that looks like it was animated by hand. He says the potential of AI-enhanced creativity is mind-bending."
Paving the Way for Machine Learning in Psychiatry - "there are places where we can accelerate the use of technology, as well as machine learning, to improve diagnostic accuracy and treatment trajectory in psychiatry. "
Nuclear regulatory body wants to use AI to detect cyberattacks on power plants - "The Nuclear Regulatory Commission wants outside researchers to determine how machine learning and artificial intelligence can detect and respond to cyberattacks at nuclear power plants, according to a SAM.gov notice."
CPU-Based Deep Learning Breakthrough Could Ease Pressure on GPU Market - "Israel-based deep learning and artificial intelligence (AI) specialist Deci announced this week that it achieved "breakthrough deep learning performance" using CPUs. "
Machine Learning for Mechanical Ventilation Control - "we present exploratory research into the design of a deep learning–based algorithm to improve medical ventilator control for invasive ventilation"
AI Arrives for Serious Photo Editing, Not Just Smartphone Snapshots - "Software maker Skylum released on Thursday a new version of its Luminar Neo photo editing tool that uses artificial intelligence to handle what used to be mundane but time-consuming tasks."
Listen to an AI voice actor try and flirt with you - "The quality of AI-generated voices has improved rapidly in recent years, but there are still aspects of human speech that escape synthetic imitation."
The Deep Learning Toolbox Now Apparently Includes Shattered Glass - "Clever speckle trick optically encodes selfies for encrypted, quantum-safe facial-recognition system"
White Castle Rolling Out Fry Station Robots To 100 More Locations - "Today White Castle announced they are rolling out the Flippy 2 fry station robot to 100 new locations. The expansion follows a trial of the original Flippy in the Chicagoland White Castle and 10 additional White Castle locations in 2020, and an upgrade to the Flippy 2 last fall."
Intel’s Mobileye to launch self-driving shuttles in U.S. in 2024 - "Intel Corp’s Mobileye unit plans to build and deploy self-driving electric shuttles with its partners in the United States in 2024, in a bid to scale up its automated driving system beyond taxis and delivery vehicles, executives told Reuters."
Is the Future of Food Quality in the Hands of Machine Learning? - "Berk Birand of Fero Labs discusses how machine learning can help food companies streamline manufacturing processes and improve product quality."
Toronto AI star Cohere lands Tiger-led $125-million venture funding - "Cohere Inc., a Toronto artificial intelligence startup looking to make it easier for humans to talk to machines, has raised US$125-million from Tiger Global Management, LLC the latest in a string of large-scale investments in Canada by the prolific New York financier."
C.H. Robinson and Waymo launch strategic partnership focused on autonomous trucking - "they have teamed up on a long-term strategic partnership focused on mutually exploring the practical application of autonomous driving technology in both logistics and supply chains."
Virtual Beings Startup Soul Machines Raises $70M - "Soul Machines has been developing virtual humans since 2016 with an eye toward branded creations representing companies in the real world in what is now called the metaverse. "
Facial recognition firm Clearview AI tells investors it is seeking massive expansion beyond law enforcement - "The facial recognition company Clearview AI is telling investors it’s on track to have 100 billion facial photos in its database within a year, enough to ensure “almost everyone in the world will be identifiable,” according to a financial presentation from December obtained by The Washington Post."
Humans Find AI-Generated Faces More Trustworthy Than the Real Thing - "Viewers struggle to distinguish images of sophisticated machine-generated faces from actual humans"
Tesla’s FSD Beta - An Experiment On Public Roads - "Tesla’s FSD Beta, which stands for ‘full-self driving’ beta, can best be summarized as a host of new driver assistant features that are not yet debugged. Chief among them is “autosteer on city streets,” which enables drivers to automatically navigate around complex urban environments witho"
Face Recognition Is Out. So How Will the IRS Verify Identity? - "Fighting fraud is important. But so is respecting privacy and guarding against bias. It's a “no-win situation,” one former official says."
Cow, Bull, and the Meaning of AI Essays - "Essays can be marvelous and sublime; they can also be churned from content mills. Where does the AI-written variety fit?"
Europe's bid for AI standard faces long road, EU lawmakers say - "STRASBOURG, Feb 16 (Reuters) - Europe's effort to set a standard for artificial intelligence will likely take more than a year, with the debate focusing on whether facial recognition should be banned and who should enforce the rules, two key lawmakers said on Wednesday."
That's all for this week! If you are not subscribed and liked this, feel free to subscribe below!