Making A.I. Systems that See the World as Humans Do
A Northwestern University team developed a new computational model that performs at human levels on a standard intelligence test. This work is an important step toward making artificial intelligence systems that see and understand the world as humans do.
“The model performs in the 75th percentile for American adults, making it better than average,” said Northwestern Engineering’s Ken Forbus. “The problems that are hard for people are also hard for the model, providing additional evidence that its operation is capturing some important properties of human cognition.”
The new computational model is built on CogSketch, an artificial intelligence platform previously developed in Forbus’ laboratory. The platform has the ability to solve visual problems and understand sketches in order to give immediate, interactive feedback. CogSketch also incorporates a computational model of analogy, based on Northwestern psychology professor Dedre Gentner’s structure-mapping theory. (Gentner received the 2016 David E. Rumelhart Prize for her work on this theory.)
Forbus, Walter P. Murphy Professor of Electrical Engineering and Computer Science at Northwestern’s McCormick School of Engineering, developed the model with Andrew Lovett, a former Northwestern postdoctoral researcher in psychology. Their research was published online this month in the journal Psychological Review.
The ability to solve complex visual problems is one of the hallmarks of human intelligence. Developing artificial intelligence systems that have this ability not only provides new evidence for the importance of symbolic representations and analogy in visual reasoning, but it could potentially shrink the gap between computer and human cognition.
While Forbus and Lovett’s system can be used to model general visual problem-solving phenomena, they specifically tested it on Raven’s Progressive Matrices, a nonverbal standardized test that measures abstract reasoning. All of the test’s problems consist of a matrix with one image missing. The test taker is given six to eight choices with which to best complete the matrix. Forbus and Lovett’s computational model performed better than the average American.
“The Raven’s test is the best existing predictor of what psychologists call ‘fluid intelligence, or the general ability to think abstractly, reason, identify patterns, solve problems, and discern relationships,’” said Lovett, now a researcher at the US Naval Research Laboratory. “Our results suggest that the ability to flexibly use relational representations, comparing and reinterpreting them, is important for fluid intelligence.”
The ability to use and understand sophisticated relational representations is a key to higher-order cognition. Relational representations connect entities and ideas such as “the clock is above the door” or “pressure differences cause water to flow.” These types of comparisons are crucial for making and understanding analogies, which humans use to solve problems, weigh moral dilemmas, and describe the world around them.
“Most artificial intelligence research today concerning vision focuses on recognition, or labeling what is in a scene rather than reasoning about it,” Forbus said. “But recognition is only useful if it supports subsequent reasoning. Our research provides an important step toward understanding visual reasoning more broadly.”
The neural structure we use to store and process information in verbal working memory is more complex than previously understood, finds a new study by NYU researchers–a discovery that has implications for the creation of artificial intelligence systems.
The neural structure we use to store and process information in verbal working memory is more complex than previously understood, finds a new study by researchers at New York University. It shows that processing information in working memory involves two different networks in the brain rather than one—a discovery that has implications for the creation of artificial intelligence (AI) systems, such as speech translation tools.
“Our results show there are at least two brain networks that are active when we are manipulating speech and language information in our minds,” explains Bijan Pesaran, an associate professor at New York University’s Center for Neural Science and the senior author of the research.
The work appears in the journal Nature Neuroscience.
Past studies had emphasized how a single “Central Executive” oversaw manipulations of information stored in working memory. The distinction is an important one, Pesaran observes, because current AI systems that replicate human speech typically assume computations involved in verbal working memory are performed by a single neural network.
“Artificial intelligence is gradually becoming more human like,” says Pesaran. “By better understanding intelligence in the human brain, we can suggest ways to improve AI systems. Our work indicates that AI systems with multiple working memory networks are needed.”
The paper’s first author was Greg Cogan, an NYU postdoctoral fellow at the time of the study and now a postdoctoral fellow at Duke University; other co-authors were Professor Orrin Devinsky, director of the Comprehensive Epilepsy Center at NYU Langone Medical Center, Werner Doyle, an associate professor at NYU Langone’s Department of Neurosurgery, Dan Friedman, an associate professor at NYU Langone’s Department of Neurology, and Lucia Melloni, an assistant professor at NYU Langone’s Department of Neurology.
The study focused on a form of working memory critical for thinking, planning, and creative reasoning and involves holding in mind and transforming the information necessary for speech and language.
The researchers examined human patients undergoing brain monitoring to treat drug-resistant epilepsy. Specifically, they decoded neural activity recorded from the surface of the brain of these patients as they were listening to speech sounds and speaking after a short delay. This method required the study’s subjects to use a rule provided by the researchers to transform speech sounds they heard into different spoken utterances—for example, the patients were told to repeat the same sound they had heard while at other times the researchers instructed the patients to listen to the sound and make a different utterance.
The researchers decoded the neural activity in each patient’s brain as the patients applied the rule to convert what they heard into what they needed to say. The results revealed that manipulating information held in working memory involved the operation of two brain networks. One network encoded the rule that the patients were using to guide the utterances they made (the rule network). Surprisingly, however, the rule network did not encode the details of how the subjects converted what they heard into what they said. The process of using the rule to transform the sounds into speech was handled by a second, transformation network. Activity in this network could be used to track how the input (what was heard) was being converted into the output (what was spoken) moment-by-moment.
Translating what you hear in one language to speak in another language involves applying a similar set of abstract rules. People with impairments of verbal working memory find it difficult to learn new languages. Modern intelligent machines also have trouble learning languages, the researchers add.
“One way we can enhance the development of more intelligent systems is with a fuller understanding of how the human brain and mind works,” notes Pesaran. “Diagnosing and treating working memory impairments in people involves psychological assessments. By analogy, machine psychology may one day be useful for diagnosing and treating impairments in the intelligence of our machines. This research examines a uniquely human form of intelligence, verbal working memory, and suggests new ways to make machines more intelligent.”
Image-processing system learns largely on its own, much like a human baby
Neuroscience and artificial intelligence experts from Rice University and Baylor College of Medicine have taken inspiration from the human brain in creating a new “deep learning” method that enables computers to learn about the visual world largely on their own, much as human babies do.
In tests, the group’s “deep rendering mixture model” largely taught itself how to distinguish handwritten digits using a standard dataset of 10,000 digits written by federal employees and high school students. In results presented this month at the Neural Information Processing Systems (NIPS) conference in Barcelona, Spain, the researchers described how they trained their algorithm by giving it just 10 correct examples of each handwritten digit between zero and nine and then presenting it with several thousand more examples that it used to further teach itself. In tests, the algorithm was more accurate at correctly distinguishing handwritten digits than almost all previous algorithms that were trained with thousands of correct examples of each digit.
“In deep-learning parlance, our system uses a method known as semisupervised learning,” said lead researcher Ankit Patel, an assistant professor with joint appointments in neuroscience at Baylor and electrical and computer engineering at Rice. “The most successful efforts in this area have used a different technique called supervised learning, where the machine is trained with thousands of examples: This is a one. This is a two.
“Humans don’t learn that way,” Patel said. “When babies learn to see during their first year, they get very little input about what things are. Parents may label a few things: ‘Bottle. Chair. Momma.’ But the baby can’t even understand spoken words at that point. It’s learning mostly unsupervised via some interaction with the world.”
Patel said he and graduate student Tan Nguyen, a co-author on the new study, set out to design a semisupervised learning system for visual data that didn’t require much “hand-holding” in the form of training examples. For instance, neural networks that use supervised learning would typically be given hundreds or even thousands of training examples of handwritten digits before they would be tested on the database of 10,000 handwritten digits in the Mixed National Institute of Standards and Technology (MNIST) database.
The semisupervised Rice-Baylor algorithm is a “convolutional neural network,” a piece of software made up of layers of artificial neurons whose design was inspired by biological neurons. These artificial neurons, or processing units, are organized in layers, and the first layer scans an image and does simple tasks like searching for edges and color changes. The second layer examines the output from the first layer and searches for more complex patterns. Mathematically, this nested method of looking for patterns within patterns within patterns is referred to as a nonlinear process.
“It’s essentially a very simple visual cortex,” Patel said of the convolutional neural net. “You give it an image, and each layer processes the image a little bit more and understands it in a deeper way, and by the last layer, you’ve got a really deep and abstract understanding of the image. Every self-driving car right now has convolutional neural nets in it because they are currently the best for vision.”
Like human brains, neural networks start out as blank slates and become fully formed as they interact with the world. For example, each processing unit in a convolutional net starts the same and becomes specialized over time as they are exposed to visual stimuli.
“Edges are very important,” Nguyen said. “Many of the lower layer neurons tend to become edge detectors. They’re looking for patterns that are both very common and very important for visual interpretation, and each one trains itself to look for a specific pattern, like a 45-degree edge or a 30-degree red-to-blue transition.
“When they detect their particular pattern, they become excited and pass that on to the next layer up, which looks for patterns in their patterns, and so on,” he said. “The number of times you do a nonlinear transformation is essentially the depth of the network, and depth governs power. The deeper a network is, the more stuff it’s able to disentangle. At the deeper layers, units are looking for very abstract things like eyeballs or vertical grating patterns or a school bus.”
Nguyen began working with Patel in January as the latter began his tenure-track academic career at Rice and Baylor. Patel had already spent more than a decade studying and applying machine learning in jobs ranging from high-volume commodities training to strategic missile defense, and he’d just wrapped up a four-year postdoctoral stint in the lab of Rice’s Richard Baraniuk, another co-author on the new study. In late 2015, Baraniuk, Patel and Nguyen published the first theoretical framework that could both derive the exact structure of convolutional neural networks and provide principled solutions to alleviate some of their limitations.
Baraniuk said a solid theoretical understanding is vital for designing convolutional nets that go beyond today’s state-of-the-art.
“Understanding video images is a great example,” Baraniuk said. “If I am looking at a video, frame by frame by frame, and I want to understand all the objects and how they’re moving and so on, that is a huge challenge. Imagine how long it would take to label every object in every frame of a video. No one has time for that. And in order for a machine to understand what it’s seeing in a video, it has to understand what objects are, the concept of three-dimensional space and a whole bunch of other really complicated stuff. We humans learn those things on our own and take them for granted, but they are totally missing in today’s artificial neural networks.”
Patel said the theory of artificial neural networks, which was refined in the NIPS paper, could ultimately help neuroscientists better understand the workings of the human brain.
“There seem to be some similarities about how the visual cortex represents the world and how convolutional nets represent the world, but they also differ greatly,” Patel said. “What the brain is doing may be related, but it’s still very different. And the key thing we know about the brain is that it mostly learns unsupervised.
“What I and my neuroscientist colleagues are trying to figure out is, What is the semisupervised learning algorithm that’s being implemented by the neural circuits in the visual cortex? and How is that related to our theory of deep learning?” he said. “Can we use our theory to help elucidate what the brain is doing? Because the way the brain is doing it is far superior to any neural network that we’ve designed.”
Machine-learning system doesn’t require costly hand-annotated data.
In recent years, computers have gotten remarkably good at recognizing speech and images: Think of the dictation software on most cellphones, or the algorithms that automatically identify people in photos posted to Facebook.
But recognition of natural sounds — such as crowds cheering or waves crashing — has lagged behind. That’s because most automated recognition systems, whether they process audio or visual information, are the result of machine learning, in which computers search for patterns in huge compendia of training data. Usually, the training data has to be first annotated by hand, which is prohibitively expensive for all but the highest-demand applications.
Sound recognition may be catching up, however, thanks to researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL). At the Neural Information Processing Systems conference next week, they will present a sound-recognition system that outperforms its predecessors but didn’t require hand-annotated data during training.
Instead, the researchers trained the system on video. First, existing computer vision systems that recognize scenes and objects categorized the images in the video. The new system then found correlations between those visual categories and natural sounds.
“Computer vision has gotten so good that we can transfer it to other domains,” says Carl Vondrick, an MIT graduate student in electrical engineering and computer science and one of the paper’s two first authors. “We’re capitalizing on the natural synchronization between vision and sound. We scale up with tons of unlabeled video to learn to understand sound.”
The researchers tested their system on two standard databases of annotated sound recordings, and it was between 13 and 15 percent more accurate than the best-performing previous system. On a data set with 10 different sound categories, it could categorize sounds with 92 percent accuracy, and on a data set with 50 categories it performed with 74 percent accuracy. On those same data sets, humans are 96 percent and 81 percent accurate, respectively.
“Even humans are ambiguous,” says Yusuf Aytar, the paper’s other first author and a postdoc in the lab of MIT professor of electrical engineering and computer science Antonio Torralba. Torralba is the final co-author on the paper.
“We did an experiment with Carl,” Aytar says. “Carl was looking at the computer monitor, and I couldn’t see it. He would play a recording and I would try to guess what it was. It turns out this is really, really hard. I could tell indoor from outdoor, basic guesses, but when it comes to the details — ‘Is it a restaurant?’ — those details are missing. Even for annotation purposes, the task is really hard.”
Because it takes far less power to collect and process audio data than it does to collect and process visual data, the researchers envision that a sound-recognition system could be used to improve the context sensitivity of mobile devices.
When coupled with GPS data, for instance, a sound-recognition system could determine that a cellphone user is in a movie theater and that the movie has started, and the phone could automatically route calls to a prerecorded outgoing message. Similarly, sound recognition could improve the situational awareness of autonomous robots.
“For instance, think of a self-driving car,” Aytar says. “There’s an ambulance coming, and the car doesn’t see it. If it hears it, it can make future predictions for the ambulance — which path it’s going to take — just purely based on sound.”
The researchers’ machine-learning system is a neural network, so called because its architecture loosely resembles that of the human brain. A neural net consists of processing nodes that, like individual neurons, can perform only rudimentary computations but are densely interconnected. Information — say, the pixel values of a digital image — is fed to the bottom layer of nodes, which processes it and feeds it to the next layer, which processes it and feeds it to the next layer, and so on. The training process continually modifies the settings of the individual nodes, until the output of the final layer reliably performs some classification of the data — say, identifying the objects in the image.
Vondrick, Aytar, and Torralba first trained a neural net on two large, annotated sets of images: one, the ImageNet data set, contains labeled examples of images of 1,000 different objects; the other, the Places data set created by Torralba’s group, contains labeled images of 401 different scene types, such as a playground, bedroom, or conference room.
Once the network was trained, the researchers fed it the video from 26 terabytes of video data downloaded from the photo-sharing site Flickr. “It’s about 2 million unique videos,” Vondrick says. “If you were to watch all of them back to back, it would take you about two years.” Then they trained a second neural network on the audio from the same videos. The second network’s goal was to correctly predict the object and scene tags produced by the first network.
The result was a network that could interpret natural sounds in terms of image categories. For instance, it might determine that the sound of birdsong tends to be associated with forest scenes and pictures of trees, birds, birdhouses, and bird feeders.
To compare the sound-recognition network’s performance to that of its predecessors, however, the researchers needed a way to translate its language of images into the familiar language of sound names. So they trained a simple machine-learning system to associate the outputs of the sound-recognition network with a set of standard sound labels.
For that, the researchers did use a database of annotated audio — one with 50 categories of sound and about 2,000 examples. Those annotations had been supplied by humans. But it’s much easier to label 2,000 examples than to label 2 million. And the MIT researchers’ network, trained first on unlabeled video, significantly outperformed all previous networks trained solely on the 2,000 labeled examples.
“With the modern machine-learning approaches, like deep learning, you have many, many trainable parameters in many layers in your neural-network system,” says Mark Plumbley, a professor of signal processing at the University of Surrey. “That normally means that you have to have many, many examples to train that on. And we have seen that sometimes there’s not enough data to be able to use a deep-learning system without some other help. Here the advantage is that they are using large amounts of other video information to train the network and then doing an additional step where they specialize the network for this particular task. That approach is very promising because it leverages this existing information from another field.”
Plumbley says that both he and colleagues at other institutions have been involved in efforts to commercialize sound recognition software for applications such as home security, where it might, for instance, respond to the sound of breaking glass. Other uses might include eldercare, to identify potentially alarming deviations from ordinary sound patterns, or to control sound pollution in urban areas. “I really think that there’s a lot of potential in the sound-recognition area,” he says.
Researchers have discovered a way to remove specific fears from the brain, using a combination of artificial intelligence and brain scanning technology.
Their technique, published in the inaugural edition of Nature Human Behaviour, could lead to a new way of treating patients with conditions such as post-traumatic stress disorder (PTSD) and phobias.
The challenge then was to find a way to reduce or remove the fear memory, without ever consciously evoking it
Fear related disorders affect around one in 14 people and place considerable pressure on mental health services. Currently, a common approach is for patients to undergo some form of aversion therapy, in which they confront their fear by being exposed to it in the hope they will learn that the thing they fear isn’t harmful after all. However, this therapy is inherently unpleasant, and many choose not to pursue it. Now a team of neuroscientists from the University of Cambridge, Japan and the USA, has found a way of unconsciously removing a fear memory from the brain.
The team developed a method to read and identify a fear memory using a new technique called ‘Decoded Neurofeedback’. The technique used brain scanning to monitor activity in the brain, and identify complex patterns of activity that resembled a specific fear memory. In the experiment, a fear memory was created in 17 healthy volunteers by administering a brief electric shock when they saw a certain computer image. When the pattern was detected, the researchers over-wrote the fear memory by giving their experimental subjects a reward.
Dr. Ben Seymour, of the University of Cambridge’s Engineering Department, was one of the authors on the study. He explained the process:
“The way information is represented in the brain is very complicated, but the use of artificial intelligence (AI) image recognition methods now allow us to identify aspects of the content of that information. When we induced a mild fear memory in the brain, we were able to develop a fast and accurate method of reading it by using AI algorithms. The challenge then was to find a way to reduce or remove the fear memory, without ever consciously evoking it.
“We realised that even when the volunteers were simply resting, we could see brief moments when the pattern of fluctuating brain activity had partial features of the specific fear memory, even though the volunteers weren’t consciously aware of it. Because we could decode these brain patterns quickly, we decided to give subjects a reward – a small amount of money – every time we picked up these features of the memory.”
The team repeated the procedure over three days. Volunteers were told that the monetary reward they earned depended on their brain activity, but they didn’t know how. By continuously connecting subtle patterns of brain activity linked to the electric shock with a small reward, the scientists hoped to gradually and unconsciously override the fear memory.
Dr Ai Koizumi, of the Advanced Telecommunicatons Research Institute International, Kyoto and Centre of Information and Neural Networks, Osaka, led the research:
“In effect, the features of the memory that were previously tuned to predict the painful shock, were now being re-programmed to predict something positive instead.”
The team then tested what happened when they showed the volunteers the pictures previously associated with the shocks.
“Remarkably, we could no longer see the typical fear skin-sweating response. Nor could we identify enhanced activity in the amygdala – the brain’s fear centre,” she continued. “This meant that we’d been able to reduce the fear memory without the volunteers ever consciously experiencing the fear memory in the process.”
Although the sample size in this initial study was relatively small, the team hopes the technique can be developed into a clinical treatment for patients with PTSD or phobias.
“To apply this to patients, we need to build a library of the brain information codes for the various things that people might have a pathological fear of, say, spiders” adds Dr Seymour. “Then, in principle, patients could have regular sessions of Decoded Neurofeedback to gradually remove the fear response these memories trigger.”
Such a treatment could have major benefits over traditional drug based approaches. Patients could also avoid the stress associated with exposure therapies, and any side-effects resulting from those drugs.
Learn more: Reconditioning the brain to overcome fear
New research published by the University of Surrey in Boston College Law Review is calling for inventions by computers to be legally granted patents.
- New research published by the University of Surrey in Boston College Law Review says the law has failed to address the issue of computer inventorship
- Inventions generated by Artificial Intelligence are rising exponentially without the legal framework to manage the issue of patents, which could result in less innovation and uncertainty about invention ownership
- Expert in patent law proposes acknowledging computers as inventors in order to incentivise the development of creative computers – without which, some inventions may never be realised
- Computers could overtake humans as the primary source of new inventions in the foreseeable future
The research states that the rapid increase in computer power is posing new challenges when it comes to patenting an invention. Artificial Intelligence is playing an ever larger role in innovation – with major players such as IBM, Pfizer and Google investing heavily in creative computing – but current patent law does not recognise computers as inventors.
Without a change in the law, the findings warn that there will be less innovation, caused by uncertainty, which would prevent industry from capitalising on the huge potential of creative computers. We are also likely to see disputes over inventorship, with individuals taking credit for inventions that are not genuinely theirs.
Ryan Abbott, Professor of Law and Health Sciences at the University of Surrey’s School of Law proposes that non-humans should be allowed to be named as inventors on patents as this would incentivise the creation of intellectual property by encouraging the development of creative computers. By assigning ownership of a computer’s invention to a computer’s owner, he argues, it would be possible to reward inventive activity which happens before the invention itself.
Professor Abbott commented, “While some patent prosecutors say the ability of machines to create patentable inventions on their own is well off in the future, artificial intelligence has actually been generating inventive ideas for decades. In just one example, an Artificial Intelligence system named ‘The Creativity Machine’ invented the first cross-bristled toothbrush design.
“Soon computers will be routinely inventing, and it may only be a matter of time until computers are responsible for most innovation. To optimise innovation – and the positive impact this will have on our economies – it is critical that we extend the laws around inventorship to include computers.”
The study also examines the implications of computer inventorship for other areas of patent law – for example whether computers should replace the ‘skilled person’ conventionally used to judge a patent’s inventiveness, since a computer would have an unlimited knowledge of the particular field in question.
The elusive and complex components of creativity have been identified by computer experts at the University of Kent.
Dr Anna Jordanous, lecturer in the School of Computing, worked with language expert Dr Bill Keller (University of Sussex) on how to define the language people use when talking about creativity, known in the field as computational creativity. With that knowledge it becomes possible to make computer programs use this language too.
Dr Jordanous and Dr Keller looked at what people say when they talk about “what is creativity” in academic discussions, from various disciplines – psychology, arts, business, and computational creativity.
In an article entitled Modelling Creativity: Identifying key components through a corpus-based approach, published by PLOS ONE, they describe a unique approach to developing a suitable model of how creative behaviour emerges that is based on the words people use to describe it. Computational creativity is a relatively new field of research into computer systems that exhibit creative behaviours.
Using language-analysis software they identified the creative words and grouped them into clusters. These are considered to be 14 components of creativity. These clusters have been used to evaluate the creativity of computational systems, and are expected to be a useful resource for other researchers in computational creativity, as well as forming a basis for the automated evaluation of creative systems.
With the help of artificial intelligence, chemists from the University of Basel in Switzerland have computed the characteristics of about two million crystals made up of four chemical elements. The researchers were able to identify 90 previously unknown thermodynamically stable crystals that can be regarded as new materials. They report on their findings in the scientific journal Physical Review Letters.
Elpasolite is a glassy, transparent, shiny and soft mineral with a cubic crystal structure. First discovered in El Paso County (Colorado, USA), it can also be found in the Rocky Mountains, Virginia and the Apennines (Italy). In experimental databases, elpasolite is one of the most frequently found quaternary crystals (crystals made up of four chemical elements). Depending on its composition, it can be a metallic conductor, a semi-conductor or an insulator, and may also emit light when exposed to radiation.
These characteristics make elpasolite an interesting candidate for use in scintillators (certain aspects of which can already be demonstrated) and other applications. Its chemical complexity means that, mathematically speaking, it is practically impossible to use quantum mechanics to predict every theoretically viable combination of the four elements in the structure of elpasolite.
Machine learning aids statistical analysis
Thanks to modern artificial intelligence, Felix Faber, a doctoral student in Prof. Anatole von Lilienfeld’s group at the University of Basel’s Department of Chemistry, has now succeeded in solving this material design problem. First, using quantum mechanics, he generated predictions for thousands of elpasolite crystals with randomly determined chemical compositions. He then used the results to train statistical machine learning models (ML models). The improved algorithmic strategy achieved a predictive accuracy equivalent to that of standard quantum mechanical approaches.
ML models have the advantage of being several orders of magnitude quicker than corresponding quantum mechanical calculations. Within a day, the ML model was able to predict the formation energy – an indicator of chemical stability – of all two million elpasolite crystals that theoretically can be obtained from the main group elements of the periodic table. In contrast, performance of the calculations by quantum mechanical means would have taken a supercomputer more than 20 million hours.
Unknown materials with interesting characteristics
An analysis of the characteristics computed by the model offers new insights into this class of materials. The researchers were able to detect basic trends in formation energy and identify 90 previously unknown crystals that should be thermodynamically stable, according to quantum mechanical predictions.
On the basis of these potential characteristics, elpasolite has been entered into the Materials Project material database, which plays a key role in the Materials Genome Initiative. The initiative was launched by the US government in 2011 with the aim of using computational support to accelerate the discovery and the experimental synthesis of interesting new materials.
Some of the newly discovered elpasolite crystals display exotic electronic characteristics and unusual compositions. “The combination of artificial intelligence, big data, quantum mechanics and supercomputing opens up promising new avenues for deepening our understanding of materials and discovering new ones that we would not consider if we relied solely on human intuition,” says study director von Lilienfeld.
This year-long exercise in scientific introspection yields a report meant to spur discussion about ‘how the fruits of an AI-dominated economy should be shared’
A panel of academic and industrial thinkers has looked ahead to 2030 to forecast how advances in artificial intelligence (AI) might affect life in a typical North American city – in areas as diverse as transportation, health care and education ¬- and to spur discussion about how to ensure the safe, fair and beneficial development of these rapidly emerging technologies.
Titled “Artificial Intelligence and Life in 2030,” this year-long investigation is the first product of the One Hundred Year Study on Artificial Intelligence (AI100), an ongoing project hosted by Stanford to inform societal deliberation and provide guidance on the ethical development of smart software, sensors and machines.
“We believe specialized AI applications will become both increasingly common and more useful by 2030, improving our economy and quality of life,” said Peter Stone, a computer scientist at the University of Texas at Austin and chair of the 17-member panel of international experts. “But this technology will also create profound challenges, affecting jobs and incomes and other issues that we should begin addressing now to ensure that the benefits of AI are broadly shared.”
The new report traces its roots to a 2009 study that brought AI scientists together in a process of introspection that became ongoing in 2014, when Eric and Mary Horvitz created the AI100 endowment through Stanford. AI100 formed a standing committee of scientists and charged this body with commissioning periodic reports on different aspects of AI over the ensuing century.
“This process will be a marathon, not a sprint, but today we’ve made a good start,” said Russ Altman, a professor of bioengineering and the Stanford faculty director of AI100. “Stanford is excited to host this process of introspection. This work makes practical contribution to the public debate on the roles and implications of artificial intelligence.”
The AI100 standing committee first met in 2015, led by chairwoman and Harvard computer scientist Barbara Grosz. It sought to convene a panel of scientists with diverse professional and personal backgrounds and enlist their expertise to assess the technological, economic and policy implications of potential AI applications in a societally relevant setting.
“AI technologies can be reliable and broadly beneficial,” Grosz said. “Being transparent about their design and deployment challenges will build trust and avert unjustified fear and suspicion.”
The report investigates eight domains of human activity in which AI technologies are beginning to affect urban life in ways that will become increasingly pervasive and profound by 2030.
The 28,000-word report includes a glossary to help nontechnical readers understand how AI applications such as computer vision might help screen tissue samples for cancers or how natural language processing will allow computerized systems to grasp not simply the literal definitions, but the connotations and intent, behind words.
The report is broken into eight sections focusing on applications of AI. Five examine application arenas such as transportation where there is already buzz about self-driving cars. Three other sections treat technological impacts, like the section on employment and workplace trends which touches on the likelihood of rapid changes in jobs and incomes.
“It is not too soon for social debate on how the fruits of an AI-dominated economy should be shared,” the researchers write in the report, noting also the need for public discourse.
“Currently in the United States, at least sixteen separate agencies govern sectors of the economy related to AI technologies,” the researchers write, highlighting issues raised by AI applications: “Who is responsible when a self-driven car crashes or an intelligent medical device fails? How can AI applications be prevented from [being used for] racial discrimination or financial cheating?”
The eight sections discuss:
- Transportation: Autonomous cars, trucks and, possibly, aerial delivery vehicles may alter how we commute, work and shop and create new patterns of life and leisure in cities.
- Home/service robots: Like the robotic vacuum cleaners already in some homes, specialized robots will clean and provide security in live/work spaces that will be equipped with sensors and remote controls.
- Health care: Devices to monitor personal health and robot-assisted surgery are hints of things to come if AI is developed in ways that gain the trust of doctors, nurses, patients and regulators.
- Education: Interactive tutoring systems already help students learn languages, math and other skills. More is possible if technologies like natural language processing platforms develop to augment instruction by humans.
- Entertainment: The conjunction of content creation tools, social networks and AI will lead to new ways to gather, organize and deliver media in engaging, personalized and interactive ways.
- Low-resource communities: Investments in uplifting technologies like predictive models to prevent lead poisoning or improve food distributions could spread AI benefits to the underserved.
- Public safety and security: Cameras, drones and software to analyze crime patterns should use AI in ways that reduce human bias and enhance safety without loss of liberty or dignity.
- Employment and workplace: Work should start now on how to help people adapt as the economy undergoes rapid changes as many existing jobs are lost and new ones are created.
“Until now, most of what is known about AI comes from science fiction books and movies,” Stone said. “This study provides a realistic foundation to discuss how AI technologies are likely to affect society.”
Grosz said she hopes the AI 100 report “initiates a century-long conversation about ways AI-enhanced technologies might be shaped to improve life and societies.”
Researchers at Houston Methodist have developed an artificial intelligence (AI) software that reliably interprets mammograms, assisting doctors with a quick and accurate prediction of breast cancer risk. According to a new study published in Cancer (early online Aug. 29), the computer software intuitively translates patient charts into diagnostic information at 30 times human speed and with 99 percent accuracy.
“This software intelligently reviews millions of records in a short amount of time, enabling us to determine breast cancer risk more efficiently using a patient’s mammogram. This has the potential to decrease unnecessary biopsies,” says Stephen T. Wong, Ph.D., P.E., chair of the Department of Systems Medicine and Bioengineering at Houston Methodist Research Institute.
The team led by Wong and Jenny C. Chang, M.D., director of the Houston Methodist Cancer Center used the AI software to evaluate mammograms and pathology reports of 500 breast cancer patients. The software scanned patient charts, collected diagnostic features and correlated mammogram findings with breast cancer subtype. Clinicians used results, like the expression of tumor proteins, to accurately predict each patient’s probability of breast cancer diagnosis.
In the United States, 12.1 million mammograms are performed annually, according to the Centers for Disease Control and Prevention (CDC). Fifty percent yield false positive results, according to the American Cancer Society (ACS), resulting in one in every two healthy women told they have cancer.
Currently, when mammograms fall into the suspicious category, a broad range of 3 to 95 percent cancer risk, patients are recommended for biopsies.
Over 1.6 million breast biopsies are performed annually nationwide, and about 20 percent are unnecessarily performed due to false-positive mammogram results of cancer free breasts, estimates the ACS.
The Houston Methodist team hopes this artificial intelligence software will help physicians better define the percent risk requiring a biopsy, equipping doctors with a tool to decrease unnecessary breast biopsies.
Manual review of 50 charts took two clinicians 50-70 hours. AI reviewed 500 charts in a few hours, saving over 500 physician hours.
“Accurate review of this many charts would be practically impossible without AI,” says Wong.
Software may appear to operate without bias because it strictly uses computer code to reach conclusions. That’s why many companies use algorithms to help weed out job applicants when hiring for a new position.
But a team of computer scientists from the University of Utah, University of Arizona and Haverford College in Pennsylvania have discovered a way to find out if an algorithm used for hiring decisions, loan approvals and comparably weighty tasks could be biased like a human being.
The researchers, led by Suresh Venkatasubramanian, an associate professor in the University of Utah’s School of Computing, have discovered a technique to determine if such software programs discriminate unintentionally and violate the legal standards for fair access to employment, housing and other opportunities. The team also has determined a method to fix these potentially troubled algorithms. Venkatasubramanian presented his findings Aug. 12 at the 21st Association for Computing Machinery’s SIGKDD Conference on Knowledge Discovery and Data Mining in Sydney, Australia.
“There’s a growing industry around doing résumé filtering and résumé scanning to look for job applicants, so there is definitely interest in this,” says Venkatasubramanian. “If there are structural aspects of the testing process that would discriminate against one community just because of the nature of that community, that is unfair.”
Many companies have been using algorithms in software programs to help filter out job applicants in the hiring process, typically because it can be overwhelming to sort through the applications manually if many apply for the same job. A program can do that instead by scanning résumés and searching for keywords or numbers (such as school grade point averages) and then assigning an overall score to the applicant.
These programs also can learn as they analyze more data. Known as machine-learning algorithms, they can change and adapt like humans so they can better predict outcomes. Amazon uses similar algorithms so they can learn the buying habits of customers or more accurately target ads, and Netflix uses them so they can learn the movie tastes of users when recommending new viewing choices.
But there has been a growing debate on whether machine-learning algorithms can introduce unintentional bias much like humans do.
“The irony is that the more we design artificial intelligence technology that successfully mimics humans, the more that A.I. is learning in a way that we do, with all of our biases and limitations,” Venkatasubramanian says.
Venkatasubramanian’s research determines if these software algorithms can be biased through the legal definition of disparate impact, a theory in U.S. anti-discrimination law that says a policy may be considered discriminatory if it has an adverse impact on any group based on race, religion, gender, sexual orientation or other protected status.
Venkatasubramanian’s research revealed that you can use a test to determine if the algorithm in question is possibly biased. If the test — which ironically uses another machine-learning algorithm — can accurately predict a person’s race or gender based on the data being analyzed, even though race or gender is hidden from the data, then there is a potential problem for bias based on the definition of disparate impact.
“I’m not saying it’s doing it, but I’m saying there is at least a potential for there to be a problem,” Venkatasubramanian says.
Read more: PROGRAMMING AND PREJUDICE
Artificial intelligence recently won out during simulated aerial combat against U.S. expert tacticians. Importantly, it did so using no more than the processing power available in a tiny, affordable computer (Raspberry Pi) that retails for as little as $35.
Not only was Lee not able to score a kill against ALPHA after repeated attempts, he was shot out of the air every time during protracted engagements in the simulator
Artificial intelligence (AI) developed by a University of Cincinnati doctoral graduate was recently assessed by subject-matter expert and retired United States Air Force Colonel Gene Lee — who holds extensive aerial combat experience as an instructor and Air Battle Manager with considerable fighter aircraft expertise — in a high-fidelity air combat simulator.
The artificial intelligence, dubbed ALPHA, was the victor in that simulated scenario, and according to Lee, is “the most aggressive, responsive, dynamic and credible AI I’ve seen to date.”