Dr. T's AI brief
Daniel Tauritz
Australian Cyber Engineers Use IBM Watson To Detect Insider Threats Across Platforms
Which-50 
(10/14) reports Australian IBM cybersecurity engineers have “developed an artificial intelligence (AI) system to analyse network connections and employee communications at an enterprise scale.” The model “detects changes in users’ behaviour and can automatically triggers investigations even if the changes occur across multiple platforms” and “addresses one of cybersecurity’s biggest challenges, the insider threat.” Developed at IBM’s Gold Coast cybersecurity lab, the solution “uses AI to monitor changes in employee behaviour and flags indicators of compromise.”
Finance Sector Increasingly Turning To AI For New Skills
CNBC (9/25, Liu) reports the finance sector has been increasingly “looking for more candidates who specialize in artificial intelligence, machine learning and data science. According to reporting by Bloomberg reporting and data from LinkedIn, job listings requiring these skills in the financial industry increased nearly 60% in the past year.” According to Glassdoor, “some of the most common job openings in AI and finance are for machine learning engineers and data engineers, among other highly specialized software engineering roles.” Glassdoor Senior Economist Daniel Zhao said, “We’re also seeing job openings for workers who can help navigate the AI landscape, including consultants and researchers. As companies establish the foundations for their AI functions, we’re seeing employers hire more senior candidates to lead these new teams.”
New Set of Images That Fool AI Could Help Make It More Hacker-Proof
Technology Review
Karen Hao
June 21, 2019
The University of California, Berkeley's Dan Hendrycks has compiled an image dataset of "natural adversarial examples," capable of deceiving artificial intelligence (AI) systems into making erroneous decisions without special doctoring. Examples include a squirrel that systems commonly misidentify as a sea lion, and a dragonfly mislabeled as a manhole cover. Synthetic adversarial examples must be fully aware of an AI system's defenses to work, but Hendrycks said natural examples remain relatively effective, even when defenses shift. Hendrycks released about 6,000 such images for use as a benchmark to test image recognition systems. Said Hendrycks, "If people were to just train on this dataset, that's just memorizing these examples. That would be solving the dataset but not the task of being robust to new examples."
Researchers Develop 'Vaccine' Against Attacks on Machine Learning
CSIRO (Australia)
Chris Chelvan
June 21, 2019
Researchers at the Commonwealth Scientific and Industrial Research Organization's (CSIRO) Data61 group in Australia have developed techniques to "vaccinate" algorithms against adversarial attacks. Cyberattackers often try to fool machine learning models by adding a layer of noise to an image, in an attempt to deceive the models into misclassifying the image. The CSIRO researchers implemented a weak version of such an adversary—like small modifications or distortions of a collection of images—to create a more "difficult" training dataset so the resulting model more easily withstands adversarial attacks. Said Data61's Adrian Turner, "The new techniques … will spark a new line of machine learning research and ensure the positive use of transformative [artificial intelligence] technologies."
Florida's Latest Oddity: Semi Trucks With Nobody Inside
The Washington Post
Peter Holley
June 26, 2019
Startup Starsky Robotics is testing unmanned semi trucks on public roads in Florida. The trucks are equipped with a hybrid driving system partly governed by a remote human operator. Starsky founder Stefan Seltz-Axmacher said, "When it comes to driving a truck, a decent person paired with a decent artificial intelligence [AI] is better than the best person or the best AI." The hybrid system leaves certain decisions, like navigating off-ramps and lane changes, up to humans, while computers are better at sustaining focus during long, uninterrupted stretches of driving. Seltz-Axmacher sees this setup as benefiting remote truck operators, who otherwise would spend long stretches on the road.
*May Require Paid Registration
A Machine May Not Take Your Job, but Could Become Your Boss
The New York Times
Kevin Roose
June 23, 2019
Call centers and other workplaces are starting to use artificial intelligence (AI) programs to make workers more effective, by giving them real-time feedback. In modern workplaces, AI programs often see human workers themselves as requiring optimization. For example, Amazon uses algorithms to track worker productivity at its fulfillment centers, and automatically generate paperwork to fire workers who do not meet their targets; meanwhile, IBM has used its Watson AI platform during employee reviews to predict future performance, claiming a 96% accuracy rate. However, critics have accused companies of using algorithms for managerial tasks, arguing automated systems can dehumanize and unfairly punish employees. Workplace AI supporters insist these systems are not meant to be overbearing, but rather to make workers better, by reminding them to thank customers, empathize with frustrated callers, or avoid idling on the job.
From One Brain Scan, More Information for Medical AI
MIT News
Rob Matheson
June 19, 2019
Researchers at the Massachusetts Institute of Technology have developed a system to gather more information from images used to train machine-learning models, including those that can analyze medical scans to help diagnose and treat brain conditions. The new system uses a single labeled scan, along with unlabeled scans, to automatically synthesize a massive dataset of distinct training examples. This dataset can be used to better train machine learning models to find anatomical structures in new scans. The system uses a convolutional neural network to automatically generate data for the "image segmentation" process, which divides an image into regions of pixels that are more meaningful and easier to analyze. The network analyzes unlabeled scans from different patients and different equipment to "learn" anatomical, brightness, and contrast variations. Then, it applies a random combination of those learned variations to a single labeled scan to synthesize new scans that are realistic and accurately labeled.
This Robot Artist Just Became the First to Stage a Solo Exhibition. What Does That Say About Creativity?
Time
Suyin Haynes
June 17, 2019
Researchers at Oxford University in the U.K. have developed Ai-Da, possibly the world's first robot artist and the latest artificial intelligence (AI) innovation to blur the line between machine and artist. Ai-Da has a robotic arm system and human-like features, and is equipped with facial recognition technology powered by AI. The system can analyze an image, which feeds into an algorithm to dictate the movement of the arm, enabling the robot to produce sketches. For example, to create prism-like paintings, Ai-Da draws a picture, and the researchers plot the coordinates from the drawing onto a Cartesian plane. Then, they run the coordinates through an AI neural network, which creates the prism effect. Said Oxford University researcher Aidan Gomez, "The potential for technology to augment the human potential for creativity, to expand the achievable horizons of creative expression and to possess its own creative potential as an entity of its own is so fascinating and exciting."
Investors Urge AI Startups to Inject Early Dose of Ethics
The Wall Street Journal
Jared Council
June 16, 2019
Artificial intelligence (AI) startup investors are urging companies to improve their products from an ethical perspective, using a code of ethics to guide operations, a tool to explain how software makes decisions, and best practices that feature consistent, open communication and immediate feedback about algorithmic output. For example, Analytics Ventures' startups employ a tool called Klear to forensically analyze why AI systems arrive at decisions. Analytics Ventures' Andreas Roell said, "I see explainability as a core component of having an ethical guardrail around AI." Meanwhile, a tech accelerator run by Innovation Works unveiled a voluntary ethics component to its program for startups, in partnership with Carnegie Mellon University. The program targets issues like bias and data privacy, and asks each startup’s founders to craft an ethical values statement.
Daniel Tauritz
Research Paper Shows That AI Programs Can Be Sabotaged By Tainted Data
Wired (11/25, Knight) reported that a recent research paper based on a study conducted by Boston University assistant professor Wenchao Li, two BU students, and a researcher at SRI International “is the latest in a growing body of evidence suggesting that AI programs can be sabotaged by the data used to train them.” According to Wired, the group “tricked a popular reinforcement-learning algorithm from DeepMind, called Asynchronous Advantage Actor-Critic, or A3C.” Li’s team “performed the attack in several Atari games using an environment created for reinforcement-learning research” and found that by “modifying just a tiny amount of training data fed to a reinforcement learning algorithm can create a back door.” Wired states the “game example is trivial, but a reinforcement-learning algorithm could control an autonomous car or a smart manufacturing robot.”
Tech Companies Are Racing To Spot Deepfakes, Which Are Getting Better
Two stories cover the race by technology companies to learn to detect deepfakes ahead of the 2020 US election. Both stories say the sophistication of deepfakes is growing as fast as tech companies can detect them. The Wall Street Journal (11/22, Morris, Subscription Publication) examines how companies are training software to spot deepfakes. Companies at work on the problem include Facebook, Twitter, Google, Amazon, Microsoft, and Adobe. The latter is using an authentication approach, which differs from other companies. Adobe now has a system that can append attribution to content and plans to share its technology, including in Photoshop.
The New York Times (11/24, Metz) has a similar story about detection efforts using artificial intelligence systems, which “learn on their own how to build fake images by analyzing thousands of real images. That means they can handle a portion of the workload that once fell to trained technicians. And that means people can create far more fake stuff than they used to.” That cycle will continue, so “the question is: Which side will improve more quickly?” The article quotes Arizona State University computer science professor Subbarao Kambhampati, who said, “Even with current technology, it is hard for some people to tell what is real and what is not.” Kambhampati also said, “In the short term, detection will be reasonably effective,” but “in the longer term, I think it will be impossible to distinguish between the real pictures and the fake pictures.”
AI Poker Bot Is First to Beat Professionals at Multiplayer Game
Nature
Douglas Heaven
July 11, 2019
Carnegie Mellon University (CMU) researchers developed an artificial intelligence (AI) program that beat elite professional poker players at six-player no-limit Texas hold'em poker. CMU's Noam Brown and Tuomas Sandholm created the Pluribus AI by updating an earlier program, Libratus, which only plays two-player matches. The researchers revamped Libratus' search algorithm, which searches to the end of a game before selecting an action. Adding more players negated the practicality of this approach, so Brown and Sandholm invented a technique that permitted Pluribus to make good choices after looking ahead only a few moves. Pluribus trained itself by initially playing poker randomly, and improved as it ascertained which actions won more money; after each hand, it reevaluated its moves, and checked whether it would have won more with different actions, which it will be more likely to utilize later on.
New Face of the Bank of England's £50 Note Revealed as Alan Turing
BBC News
Kevin Peachey
July 15, 2019
Alan Turing, a celebrated computer pioneer and codebreaker, will be featured on the new design of the Bank of England's £50 note, slated to enter circulation by the end of 2021. Turing is known as the father of computer science and artificial intelligence (AI), and his work helped accelerate Allied efforts to read German Naval messages encrypted with the Enigma machine during World War II. Turing's work helped cement the concept of the algorithm—a set of instructions used to perform computations. Said Bank of England Governor Mark Carney, "As the father of computer science and AI, as well as a war hero, Alan Turing's contributions were far-ranging and path-breaking. Turing is a giant on whose shoulders so many now stand."
Intel Packs 8 Million Digital Neurons into Its Brain-Like Computer |
This AI Can Solve a Rubik's Cube Super-Fast
Forbes
Jennifer Kite-Powell
July 15, 2019
University of California, Irvine (UCI) researchers programmed an artificial intelligence (AI) system to solve a Rubik's cube in one second, without any domain knowledge or in-game coaching from humans. The DeepCubeA algorithm solved 100% of all test configurations and found the shortest path to the goal—all six sides displaying a solid color—about 60% of the time. The algorithm also works on other combinatorial games. The team started with a computer simulation of the completed puzzle and then scrambled the Rubik's cube. After the code was running, DeepCubeA trained in isolation for two days, solving an increasingly difficult series of combinations, during which time it began to learn on its own. Said UCI researcher Pierre Baldi, "This work is part of a general effort to bridge machine learning AI and symbolic AI to address complex problems that humans solve through planning and reasoning."
At Zappos, Algorithms Teach Themselves
The Wall Street Journal
Jared Council
July 8, 2019
Online shoe and clothing retailer Zappos sees promise in a self-learning algorithm's ability to address the problem of its search engine producing irrelevant results. Zappos' chief data scientist Ameen Kazerouni said several years ago his team began testing a genetic algorithm, which has since become critical to boosting the search engine's relevancy. Genetic algorithms generate various solutions to a problem, using natural-selection principles like reproduction and mutation to return the optimal or "fittest" solution. The algorithms were designed to parse out the intent of a search phrase, with those that perform best on an internal "relevance test," which models how users engage with search results, having the greatest odds of having their traits inherited by the next generation. Zappos uses three genetic algorithm engines in parallel to generate better search results.
AI Can Edit Photos with Zero Experience
IEEE Spectrum
Matthew Hutson
July 10, 2019
Researchers at the Weizmann Institute of Science in Israel are using deep internal learning, in which a machine learning algorithm ascertains the internal structure of a single image from scratch, to edit photos without previous training. This achievement builds on research from a team at the Skolkovo Institute of Science and Technology in Russia, involving Deep Image Prior (DIP), a technique in which a multi-neural network is trained to replicate a specific image by looking for hierarchies of repeating features. The Weizmann researchers' Double-DIP process has two DIPs running in parallel, with each converting a random input into an image, and both images superimposed on and compared to a target image. The DIPs then independently modify their parameters so their combined image comes closer to the target. Dmitry Ulyanov of Moscow’s Skolkovo Institute of Science and Technology said he and his collaborators designed DIP to study the importance of network architecture (versus data).
Researchers Cast Neural Nets to Simulate Molecular Motion |
AI Invents More Effective Flu Vaccine, Adelaide Researchers Say
ABC Online (Australia)
David Sparkes; Rhett Burnie
July 2, 2019
An artificial intelligence (AI) program created a "turbocharged" flu vaccine, marking the first time a computer program has created a new drug on its own, according to researchers at Flinders University in Australia. The researchers took existing drugs that are known to work, as well as examples of drugs that do not work or have failed, and fed that information to an AI program called Sam. The program generated a suggestion of what might be an effective therapy, which the researchers tested and found that it worked. A 12-month clinical trial of the vaccine will soon be conducted in the U.S., according to Flinders researcher Nikolai Petrovsky.
Teaching AI to Create Visuals with More Common Sense
MIT News
Adam Conner-Simons
July 1, 2019
Researchers at IBM and the Massachusetts Institute of Technology (MIT) have developed a system that can automatically generate realistic photographic images and edit objects inside them. The system, called GANpaint Studio, could help computer scientists identify "fake" images, as well as helping artists and designers make quick adjustments to visuals. GANpaint Studio also could be used to improve and debug other generative adversarial networks (GANs) under development by analyzing them for "artifact" units that need to be removed. MIT Ph.D. student David Bau said the project was one of the first times computer scientists have been able to “paint with the neurons” of a neural network.
AI Job Market Cools to a Steady Boil
IEEE Spectrum
Tekla S. Perry
June 28, 2019
In its annual review of artificial intelligence (AI) job postings, job search site Indeed found the number of AI jobs listed from May 2018 to May 2019 increased by 29% over the same period a year earlier. However, that figure was significantly less than the increase over the previous year (58%). Also during that prior year, AI job postings jumped 136% over the previous year's review, a rate with declined significantly in the year through May 2019. The Indeed study did not quantify the gap between job openings and job seekers, but the company says its data suggests that gap is growing and the shortage of AI-related workers is worsening. This trend is good news for engineers with AI expertise, and bad news for companies that need to hire them.
dtau...@gmail.com
Intel Expects To Blend More AI Circuitry Into Its Products
CNET News (11/14, Shankland) reports that although computer and software makers “haven’t started clamoring for hardware that accelerates artificial intelligence tasks in a personal computer,” that will change, according to Intel, which “expects to blend more AI circuitry into its products as it catches on.” At Intel’s 2019 AI Summit this week, Intel AI GM Naveen Rao said, “You’re going to see this benefiting everybody, because the whole purpose of the computer is shifting to be an AI machine. A lot of your experiences are going to start relying on AI capabilities, even in your laptop.” The company has “already added some modest AI capability to its new Ice Lake laptop processors,” such as DLBoost, which is “designed to speed AI tasks like figuring out what’s in a photo.” Intel Silicon Engineering Group GM Jim Keller, in particular, sees “AI hardware as powering the next major boost in computing performance.” In this context, Intel expects “$3.5 billion in revenue from AI products in 2019. AI chips deliver the next wave of computing power.”
With $1 Billion from Microsoft, an AI Lab Wants to Mimic the Brain
The New York Times
Cade Metz
July 23, 2019
Microsoft has made a $1-billion investment in OpenAI, the artificial intelligence lab created by Elon Musk and Sam Altman in 2015. Musk left the nonprofit last year, and Altman has since transformed OpenAI into a for-profit company so it could more aggressively pursue financing. OpenAI researchers will use the funding to develop artificial general intelligence (AGI), which would match the capabilities of the human brain. Said Geoffrey Hinton, a Google researcher and co-recipient of the 2018 ACM A.M. Turing Award, “It's too big a problem. I'd much rather focus on something where you can figure out how you might solve it.”
This Deep Neural Network Fights Deepfakes
University of California, Riverside
Holly Ober
July 18, 2019
Researchers at the University of California, Riverside (UCR) have developed a deep neural network architecture that can identify manipulated images at the pixel level with high precision. The researchers labeled nonmanipulated images and the relevant pixels in boundary regions of manipulated images in a large dataset of photos. The team trained the network on general knowledge about the manipulated and original regions of photos. Then, they tested the neural network on a set of images it had never seen before, and found it was able to detect altered images most of the time. The researchers say their methodology could be adapted to detect deepfake videos, though there are challenges to overcome. Said UCR's Amit Roy-Chowdhury, “It's a challenging problem. This is kind of a cat and mouse game. This whole area of cybersecurity is in some ways trying to find better defense mechanisms, but then the attacker also finds better mechanisms."
Experian Tests AI Platform to Improve Identity Verification
The Wall Street Journal
Jared Council
July 19, 2019
Researchers at the Experian credit-reporting firm are testing whether artificial intelligence (AI) can better validate consumer identities. A new AI platform, which the company expects to introduce in 2020, uses machine learning algorithms to assign an identity score to a person based on hundreds of elements. Experian is currently testing the platform in its U.S. financial services and marketing data business units. The AI platform also could be used to help retailers and brands learn more about who saw their ad campaigns and made purchases. Said Experian's Eric Haller, "The amount of information used to describe identity is only going to expand, so we want to make sure that the solution we develop can handle that expansion without human intervention."
IBM Gives Cancer-Killing Drug AI Project to the Open Source Community
ZDNet
Charlie Osborne
July 22, 2019
IBM has released to the open source community three artificial intelligence (AI) projects designed to address the challenge of curing cancer. The projects, led by researchers at IBM's Computational Systems Biology Group in Switzerland, involve developing AI and machine learning approaches to help accelerate the understanding of the leading drivers and molecular mechanisms of different cancers. The first project, PaccMann, is working to develop an algorithm that can automatically analyze chemical compounds and predict which are most likely to overcome cancer strains. The second project, "Interaction Network infErence from vectoR representATions of words" (INtERAcT), aims to develop a tool that can automatically extract information from the thousands of papers published every year on cancer research. The third project, "pathway-induced multiple kernel learning," focuses on an algorithm that uses datasets describing what is currently known about molecular interactions to predict the prognosis of cancer patients.
AI Drug Hunters Could Give Big Pharma a Run for Its Money
Bloomberg
Robert Langreth
July 15, 2019
Using the latest neural-network algorithms, DeepMind, the artificial intelligence (AI) arm of Alphabet, beat seasoned biologists at 50 top labs from around the world in predicting the shapes of proteins. The company's win at the CASP13 meeting in Mexico in December has serious implications, as a tool able to accurately model protein structures could speed up the development of new drugs. Although DeepMind's simulation was unable to produce the atomic-level resolution necessary for drug discovery, its victory points to the potential for practical application of AI in one of the most expensive and failure-prone parts of the pharmaceutical business. AI could be used, for example, to scan millions of high-resolution cellular images to identify therapies researchers might otherwise have missed. In the short term, experts say AI-based simulations likely will be used to determine whether prospective drugs will be effective before proceeding to a full clinical trial.
The AI Technique that Could Imbue Machines with the Ability to Reason
Technology Review
Karen Hao
July 12, 2019
While deep learning algorithms have made great strides in giving machines perceptual abilities like vision, they have fallen short of giving them human-like reasoning skills. Yann LeCun, chief AI scientist at Facebook, a professor at New York University and 2018 ACM A.M. Turing Award co-recipient, suggested during a recent ACM webinar that the deep learning subcategory known as unsupervised learning could help the technology overcome this hurdle. LeCun thinks researchers should focus on temporal prediction by training large neural networks to predict the second half of a video when given the first. Ultimately, LeCun said, unsupervised learning will help computers develop a model of the world that can be used to predict future states of the world.
Researchers Improve AI that Can Tell from Your Voice if You’re Depressed
Folio (University of Alberta)
Katie Willis
July 11, 2019
Researchers at the University of Alberta in Canada have improved an artificial intelligence (AI) system to detect whether a person is depressed by analyzing the sound of their voice. The researchers developed a method combining several machine learning algorithms to recognize depression more accurately from acoustic cues. The tool was trained on two standard benchmark sets of audio recordings ranging from five to 50 minutes long. In addition, the team built on past studies suggesting the timbre of a person's voice contains information about his or her mood. The ultimate goal is to develop useful applications from the technology, such as helping people reflect on their moods over time, or working with mental healthcare providers.