Social Mapper is a Open Source Intelligence Tool that uses facial recognition to correlate social media profiles across different sites on a large scale. It takes an automated approach to searching popular social media sites for targets names and pictures to accurately detect and group a person’s presence, outputting the results into report that a human operator can quickly review.
Social Mapper has a variety of uses in the security industry, for example the automated gathering of large amounts of social media profiles for use on targeted phishing campaigns. Facial recognition aids this process by removing false positives in the search results, so that reviewing this data is quicker for a human operator.
Human-generated knowledge bases like Wikipedia have a recall problem. First, there are the articles that should be there but are entirely missing. The unknown unknowns.
Consider Joelle Pineau, the Canadian roboticist bringing scientific rigor to artificial intelligence and who directs Facebook’s new AI Research lab in Montreal. Or Miriam Adelson, an actively publishing addiction treatment researcher who happens to be a billionaire by marriage and a major funder of her own field. Or Evelyn Wang, the new head of MIT’s revered MechE department whose accomplishments include a device that generates drinkable water from sunlight and desert air. When I wrote this a few days ago, none of them had articles on English Wikipedia, though they should by any measure of notability.
(Pineau is up now thanks to my friend and fellow science crusader Jess Wade who created an article just hours after I told her about Pineau’s absence. And if the internet is in a good mood, someone will create articles for the other two soon after this post goes live.)
But I didn’t discover those people on my own. I used a machine learning system we’re building at Primer. It discovered and described them for me. It does this much as a human would, if a human could read 500 million news articles, 39 million scientific papers, all of Wikipedia, and then write 70,000 biographical summaries of scientists.
[…]
We are publicly releasing free-licensed data about scientists that we’ve been generating along the way, starting with 30,000 computer scientists. Only 15% of them are known to Wikipedia. The data set includes 1 million news sentences that quote or describe the scientists, metadata for the source articles, a mapping to their published work in the Semantic Scholar Open Research Corpus, and mappings to their Wikipedia and Wikidata entries. We will revise and add to that data as we go. (Many thanks to Oren Etzioni and AI2 for data and feedback.) Our aim is to help the open data research community build better tools for maintaining Wikipedia and Wikidata, starting with scientific content.
Fluid Knowledge
We trained Quicksilver’s models on 30,000 English Wikipedia articles about scientists, their Wikidata entries, and over 3 million sentences from news documents describing them and their work. Then we fed in the names and affiliations of 200,000 authors of scientific papers.
In the morning we found 40,000 people missing from Wikipedia who have a similar distribution of news coverage as those who do have articles. Quicksilver doubled the number of scientists potentially eligible for a Wikipedia article overnight.
It also revealed the second flavor of the recall problem that plagues human-generated knowledge bases: information decay. For most of those 30,000 scientists who are on English Wikipedia, Quicksilver identified relevant information that was missing from their articles.
A recent scientific survey off the coast of Sulawesi Island in Indonesia suggests that some shallow water corals may be less vulnerable to global warming than previously thought.
Between 2014 and 2017, the world’s reefs endured the worst coral bleaching event in history, as the cyclical El Niño climate event combined with anthropogenic warming to cause unprecedented increases in water temperature.
But the June survey, funded by Microsoft co-founder Paul Allen’s family foundation, found the Sulawesi reefs were surprisingly healthy.
In fact the reefs hadn’t appeared to decline significantly in condition than when they were originally surveyed in 2014 – a surprise for British scientist Dr Emma Kennedy, who led the research team.
A combination of 360-degree imaging tech and Artificial Intelligence (AI) allowed scientists to gather and analyse more than 56,000 images of shallow water reefs. Over the course of a six-week voyage, the team deployed underwater scooters fitted with 360 degree cameras that allowed them to photograph up to 1.5 miles of reef per dive, covering a total of 1487 square miles in total.
Researchers at the University of Queensland in Australia then used cutting edge AI software to handle the normally laborious process of identifying and cataloguing the reef imagery. Using the latest Deep Learning tech, they ‘taught’ the AI how to detect patterns in the complex contours and textures of the reef imagery and thus recognise different types of coral and other reef invertebrates.
Once the AI had shown between 400 and 600 images, it was able to process images autonomously. Says Dr Kennedy, “the use of AI to rapidly analyse photographs of coral has vastly improved the efficiency of what we do — what would take a coral reef scientist 10 to 15 minutes now takes the machine a few seconds.”
Sketch2Code is a solution that uses AI to transform a handwritten user interface design from a picture to a valid HTML markup code.
Process flow
The process of transformation of a handwritten image to HTML this solution implements is detailed as follows:
The user uploads an image through the website.
A custom vision model predicts what HTML elements are present in the image and their location.
A handwritten text recognition service reads the text inside the predicted elements.
A layout algorithm uses the spatial information from all the bounding boxes of the predicted elements to generate a grid structure that accommodates all.
An HTML generation engine uses all these pieces of information to generate an HTML markup code reflecting the result.
A group of researchers from Aalto University and the University of Padua found this out when they tested seven state-of-the-art models used to detect hate speech. All of them failed to recognize foul language when subtle changes were made, according to a paper [PDF] on arXiv.
Adversarial examples can be created automatically by using algorithms to misspell certain words, swap characters for numbers or add random spaces between words or attach innocuous words such as ‘love’ in sentences.
The models failed to pick up on adversarial examples and successfully evaded detection. These tricks wouldn’t fool humans, but machine learning models are easily blindsighted. They can’t readily adapt to new information beyond what’s been spoonfed to them during the training process.
“They perform well only when tested on the same type of data they were trained on. Based on these results, we argue that for successful hate speech detection, model architecture is less important than the type of data and labeling criteria. We further show that all proposed detection techniques are brittle against adversaries who can (automatically) insert typos, change word boundaries or add innocuous words to the original hate speech,” the paper’s abstract states.
Google is putting an artificial intelligence system in charge of its data center cooling after the system proved it could cut energy use.
Now Google and its AI company DeepMind are taking the project further; instead of recommendations being implemented by human staff, the AI system is directly controlling cooling in the data centers that run services including Google Search, Gmail and YouTube.
“This first-of-its-kind cloud-based control system is now safely delivering energy savings in multiple Google data centers,” Google said.
Data centers use vast amount of energy and as the demand for cloud computing rises even small tweaks to areas like cooling can produce significant time and cost savings. Google’s decision to use its own DeepMind-created system is also a good plug for its AI business.
Every five minutes, the AI pulls a snapshot of the data center cooling system from thousands of sensors. This data is fed into deep neural networks, which predict how different choices will affect future energy consumption.
The AI system then identifies tweaks that could reduce energy consumption, which are then sent back to the data center, checked by the local control system and implemented.
Google said giving the AI more responsibility came at the request of its data center operators who said that implementing the recommendations from the AI system required too much effort and supervision.
“We wanted to achieve energy savings with less operator overhead. Automating the system enabled us to implement more granular actions at greater frequency, while making fewer mistakes,” said Google data center operator Dan Fuenffinger.
AI is being deployed by those who set and mark exams to reduce fraud — which remains overall a small problem — and to create far greater efficiencies in preparation and marking, and to help improve teaching and studying. From a report, which may be paywalled: From traditional paper-based exam and textbook producers such as Pearson, to digital-native companies such as Coursera, online tools and artificial intelligence are being developed to reduce costs and enhance learning. For years, multiple-choice tests have allowed scanners to score results without human intervention. Now technology is coming directly into the exam hall. Coursera has patented a system to take images of students and verify their identity against scanned documents. There are plagiarism detectors that can scan essay answers and search the web — or the work of other students — to identify copying. Webcams can monitor exam locations to spot malpractice. Even when students are working, they provide clues that can be used to clamp down on cheats. They leave electronic “fingerprints” such as keyboard pressure, speed and even writing style. Emily Glassberg Sands, Cousera’s head of data science, says: “We can validate their keystroke signatures. It’s difficult to prepare for someone hell-bent on cheating, but we are trying every way possible.”
Artificial intelligent software has been trained to detect and flag up clickbait headlines.
And here at El Reg we say thank God Larry Wall for that. What the internet needs right now is software to highlight and expunge dodgy article titles about spacealien immigrants, faked moon landings, and the like.
Machine-learning eggheads continue to push the boundaries of natural language processing, and have crafted a model that can, supposedly, detect how clickbait-y a headline really is.
The system uses a convolutional neural network that converts the words in a submitted article title into vectors. These numbers are fed into a long-short-term memory network that spits out a score based on the headline’s clickbait strength. About eight times out of ten it agreed with humans on whether a title was clickbaity or not, we’re told.
The trouble is, what exactly is a clickbait headline? It’s a tough question. The AI’s team – from the International Institute of Information Technology in Hyderabad, the Manipal Institute of Technology, and Birla Institute of Technology, in India – decided to rely on the venerable Merriam-Webster dictionary to define clickbait.
One of the more frustrating aspects of Windows 10 is the operating system’s ability to start installing updates when you’re in the middle of using it. While Microsoft has tried to address this aggressive approach to updates with features to snooze installation, Windows 10 users continue to complain that updates reboot devices when they’re in use.
Reacting to this feedback, Microsoft says it’s aware of the issues. “We heard you, and to alleviate this pain, if you have an update pending we’ve updated our reboot logic to use a new system that is more adaptive and proactive,” explains Microsoft’s Windows Insider chief Dona Sarkar. Microsoft says it has trained a “predictive model” that will accurately predict when the best time to restart the device is thanks to machine learning. “We will not only check if you are currently using your device before we restart, but we will also try to predict if you had just left the device to grab a cup of coffee and return shortly after,” says Sarkar.
Microsoft has been testing this new model internally, and says it has seen “promising results.”
AI can help neurologists automatically map the connections between different neurons in brain scans, a tedious task that can take hundreds and thousands of hours.
In a paper published in Nature Methods, AI researchers from Google collaborated with scientists from the Max Planck Institute of Neurobiology to inspect the brain of a Zebra Finch, a small Australian bird renowned for its singing.
Although the contents of their craniums are small, Zebra Finches aren’t birdbrains, their connectome* is densely packed with neurons. To study the connections, scientists study a slice of the brain using an electron microscope. It requires high resolution to make out all the different neurites, the nerve cells extending from neurons.
The neural circuits then have to be reconstructed by tracing out the cells. There are several methods that help neurologists flesh these out, but the error rates are high and it still requires human expertise to look over the maps. It’s a painstaking chore, a cubic millimetre of brain tissue can generate over 1,000 terabytes of data.
“A recent estimate put the amount of human labor needed to reconstruct a 1003-µm3 volume at more than 100,000 h, even with an optimized pipeline,” according to the paper.
Now, AI researchers have developed a new method using a recurrent convolutional neural network known as a “flood-filling network”. It’s essentially an algorithm that finds the edges of a neuron path and fleshes out the space in between to build up a map of the different connections.
Here’s a video showing how they work.
“The algorithm is seeded at a specific pixel location and then iteratively “fills” a region using a recurrent convolutional neural network that predicts which pixels are part of the same object as the seed,” said Viren Jain and Michal Januszewski, co-authors of the paper and AI researchers at Google.
The flood-filling network was trained using supervised learning on a small region of a Zebra Finch brain complete with annotations. It’s difficult to measure the accuracy of the network, and instead the researchers use a “expected run length” (ERL) metric that measures how far it can trace out a neuron before making a mistake.
Flood-filling networks have a longer ERL than other deep learning methods that have also been tested on the same dataset. The algorithms were better than humans at identifying dendritic spines, tiny threads jutting off dendrites that help transmit electrical signals to cells. But the level of recall, a property measuring the completeness of the map, was much lower than data collected by a professional neurologist.
Another significant disadvantage of this approach is the high computational cost. “For example, a single pass of the fully convolutional FFN over a full volume is an order of magnitude more computationally expensive than the more traditional 3D convolution-pooling architecture in the baseline approach we used for comparison,” the researchers said.
Lee Cronin, the researcher who organized the work, was kind enough to send along an image of the setup, which looks nothing like our typical conception of a robot (the researchers refer to it as “bespoke”). Most of its parts are dispersed through a fume hood, which ensures safe ventilation of any products that somehow escape the system. The upper right is a collection of tanks containing starting materials and pumps that send them into one of six reaction chambers, which can be operated in parallel.
Enlarge/ The robot in question. MS = Mass Spectrometer; IR = Infrared Spectrometer.
Lee Cronin
The outcomes of these reactions can then be sent on for analysis. Pumps can feed samples into an IR spectrometer, a mass spectrometer, and a compact NMR machine—the latter being the only bit of equipment that didn’t fit in the fume hood. Collectively, these can create a fingerprint of the molecules that occupy a reaction chamber. By comparing this to the fingerprint of the starting materials, it’s possible to determine whether a chemical reaction took place and infer some things about its products.
All of that is a substitute for a chemist’s hands, but it doesn’t replace the brains that evaluate potential reactions. That’s where a machine-learning algorithm comes in. The system was given a set of 72 reactions with known products and used those to generate predictions of the outcomes of further reactions. From there, it started choosing reactions at random from the remaining list of options and determining whether they, too, produced products. By the time the algorithm had sampled 10 percent of the total possible reactions, it was able to predict the outcome of untested reactions with more than 80-percent accuracy.
And, since the earlier reactions it tested were chosen at random, the system wasn’t biased by human expectations of what reactions would or wouldn’t work.
Once it had built a model, the system was set up to evaluate which of the remaining possible reactions was most likely to produce products and prioritize testing those. The system could continue on until it reached a set number of reactions, stop after a certain number of tests no longer produced products, or simply go until it tested every possible reaction.
Neural networking
Not content with this degree of success, the research team went on to add a neural network that was provided with data from the research literature on the yield of a class of reactions that links two hydrocarbon chains. After training on nearly 3,500 reactions, the system had an error of only 11 percent when predicting the yield on another 1,700 reactions from the literature.
This system was then integrated with the existing test setup and set loose on reactions that hadn’t been reported in the literature. This allowed the system to prioritize not only by whether the reaction was likely to make a product but also how much of the product would be produced by the reaction.
All this, on its own, is pretty impressive. As the authors put it, “by realizing only 10 percent of the total number of reactions, we can predict the outcomes of the remaining 90 percent without needing to carry out the experiments.” But the system also helped them identify a few surprises—cases where the fingerprint of the reaction mix suggested that the product was something more than a simple combination of starting materials. These reactions were explored further by actual human chemists, who identified both ring-breaking and ring-forming reactions this way.
That last aspect really goes a long way toward explaining how this sort of capability will fit into future chemistry labs. People tend to think of robots as replacing humans. But in this context, the robots are simply taking some of the drudgery away from humans. No sane human would ever consider trying every possible combination of reactants to see what they’d do, and humans couldn’t perform the testing 24 hours a day without dangerous levels of caffeine anyway. The robots will also be good at identifying the rare cases where highly trained intuitions turn out to lead us astray about the utility of trying some reactions.
Denmark-based brewing giant Carlsberg has reported good progress in its attempts to turn Microsoft’s Azure AI into a robot beer sniffer.
The project, which kicked off earlier this year, was aimed at cutting the time a beer spends in research and development by one-third, thus getting fresh brews into the hands of drinkers faster … and their beer tokens into the pockets of Carlsberg.
The director and professor of yeast and fermentation for Carlsberg, Joch Förster, has been tasked with the seemingly enviable job of tasting a lot of beer as the brewer tries out new flavours. In reality, however, ploughing through hundreds of samples isn’t really practical. Hence Förster and his team have turned to sensors and AI to predict what a beer will taste like.
Photographers already face an uphill battle in trying to preventing people from using their digital photos without permission. But Nvidia could make protecting photos online much harder with a new advancement in artificial intelligence that can automatically remove artifacts from a photograph, including text and watermarks, no matter how obtrusive they may be.In previous advancements in automated image editing and manipulation, an AI powered by a deep learning neural network is trained on thousands of before and after example photos so that it knows what the desired output should look like. But this time, researchers at Nvidia, MIT, and Aalto University in Finland, managed to train an AI to remove noise, grain, and other visual artifacts by studying two different versions of a photo that both feature the visual defects. Fifty-thousand samples later, the AI can clean up photos better than a professional photo restorer.Practical applications for the AI include cleaning up long exposure photos of the night sky taken by telescopes, as cameras used for astrophotography often generate noise that can be mistaken for stars. The AI can also be beneficial for medical applications like magnetic resonance imaging that requires considerable post-processing to remove noise from images that are generated, so that doctors have a clear image of what’s going in someone’s body. Nvidia’s AI can cut that processing time down drastically, which in turn reduces the time needed for a diagnosis of a serious condition.
Amateur and professional musicians alike may spend hours pouring over YouTube clips to figure out exactly how to play certain parts of their favorite songs. But what if there were a way to play a video and isolate the only instrument you wanted to hear?
That’s the outcome of a new AI project out of MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL): a deep-learning system that can look at a video of a musical performance, and isolate the sounds of specific instruments and make them louder or softer.
The system, which is “self-supervised,” doesn’t require any human annotations on what the instruments are or what they sound like.
Trained on over 60 hours of videos, the “PixelPlayer” system can view a never-before-seen musical performance, identify specific instruments at pixel level, and extract the sounds that are associated with those instruments.
For example, it can take a video of a tuba and a trumpet playing the “Super Mario Brothers” theme song, and separate out the soundwaves associated with each instrument.
The researchers say that the ability to change the volume of individual instruments means that in the future, systems like this could potentially help engineers improve the audio quality of old concert footage. You could even imagine producers taking specific instrument parts and previewing what they would sound like with other instruments (i.e. an electric guitar swapped in for an acoustic one).
AI agents continue to rack up wins in the video game world. Last week, OpenAI’s bots were playing Dota 2; this week, it’s Quake III, with a team of researchers from Google’s DeepMind subsidiary successfully training agents that can beat humans at a game of capture the flag.
As we’ve seen with previous examples of AI playing video games, the challenge here is training an agent that can navigate a complex 3D environment with imperfect information. DeepMind’s researchers used a method of AI training that’s also becoming standard: reinforcement learning, which is basically training by trial and error at a huge scale.
Agents are given no instructions on how to play the game, but simply compete against themselves until they work out the strategies needed to win. Usually this means one version of the AI agent playing against an identical clone. DeepMind gave extra depth to this formula by training a whole cohort of 30 agents to introduce a “diversity” of play styles. How many games does it take to train an AI this way? Nearly half a million, each lasting five minutes.
As ever, it’s impressive how such a conceptually simple technique can generate complex behavior on behalf of the bots. DeepMind’s agents not only learned the basic rules of capture the flag (grab your opponents’ flag from their base and return it to your own before they do the same to you), but strategies like guarding your own flag, camping at your opponent’s base, and following teammates around so you can gang up on the enemy.
To make the challenge harder for the agents, each game was played on a completely new, procedurally generated map. This ensured the bots weren’t learning strategies that only worked on a single map.
Unlike OpenAI’s Dota 2 bots, DeepMind’s agents also didn’t have access to raw numerical data about the game — feeds of numbers that represents information like the distance between opponents and health bars. Instead, they learned to play just by looking at the visual input from the screen, the same as a human. However, this does not necessarily mean that DeepMind’s bots faced a greater challenge; Dota 2 is overall a much more complex game than the stripped-down version of Quake III that was used in this research.
To test the AI agents’ abilities, DeepMind held a tournament, with two-player teams of only bots, only humans, and a mixture of bots and humans squaring off against one another. The bot-only teams were most successful, with a 74 percent win probability. This compared to 43 precent probability for average human players, and 52 percent probability for strong human players. So: clearly the AI agents are the better players.
A graph showing the Elo (skill) rating of various players. The “FTW” agents are DeepMind’s, which played against themselves in a team of 30.Credit: DeepMind
However, it’s worth noting that the greater the number of DeepMind bots on a team, the worse they did. A team of four DeepMind bots had a win probability of 65 percent, suggesting that while the researchers’ AI agents did learn some elements of cooperative play, these don’t necessarily scale up to more complex team dynamics.
As ever with research like this, the aim is not to actually beat humans at video games, but to find new ways of teaching agents to navigate complex environments while pursuing a shared goal. In other words, it’s about teaching collective intelligence — something that has (despite abundant evidence to the contrary) been integral to humanity’s success as a species. Capture the flag is just a proxy for bigger games to come.
Bedrijven worden emotioneler: gebruikersinterfaces, chatbots en andere componenten zijn steeds beter in staat om de emotionele staat van gebruikers in te schatten en emotie te simuleren als ze terug praten. Volgens een Gartner-rapport eerder dit jaar weten apparaten over vier jaar “meer over je emotionele staat dan je eigen familie”.
Herkennen van emotie
Deep learning kan geavanceerd emoties herkennen zoals geluk, verrassing, woede, verdriet, angst en afschuw – tot meer dan twintig subtielere emoties zoals bewondering, blije verrassing en haat. (Psychologen beweren dat mensen 27 verschillende emoties hebben.)
De Universiteit van Ohio ontwikkelde een programma dat 21 emoties herkent op basis van gezichtsuitdrukkingen op foto’s. Het schokkende: De onderzoekers beweren dat hun systeem deze emoties beter detecteert dan mensen. Er is een goede reden en een geweldige reden voor emotionele interfaces in de organisatie.
Kwaliteitsinteracties
Ten eerste de goede reden. De “empathie economie” is de monetaire of zakelijke waarde die door AI wordt gecreëerd en die menselijke emoties detecteert en simuleert, een vermogen dat klantenservice, virtuele assistenten, robotica, fabrieksveiligheid, gezondheidszorg en transport zal transformeren.
Uit een Cogito-onderzoek van Frost & Sullivan gaf 93% van de ondervraagden aan dat interacties met de klantenservice van invloed zijn op hun perceptie van een bedrijf. En empathie is één van de belangrijkste factoren in kwaliteitsinteracties, volgens het bedrijf. Cogito’s AI-software, die uitgebreid is gebaseerd op gedragswetenschappelijk onderzoek van MIT’s Human Dynamics Lab, analyseert de emotionele toestand van klanten en geeft directe feedback aan menselijke call center agents, waardoor ze gemakkelijker meevoelen met klanten.
Zorg en andere toepassingen
Dit soort technologie geeft callcentermedewerkers empathische vermogens, die de publieke perceptie van een bedrijf sterk kunnen verbeteren. Bedrijven als Affectiva en Realeyes bieden cloud-gebaseerde oplossingen die webcams gebruiken om gezichtsuitdrukkingen en hartslag te volgen (door de polsslag in de huid van het gezicht te detecteren). Een van de toepassingen is marktonderzoek: consumenten kijken naar advertenties, en de technologie detecteert hoe ze denken over de beelden of woorden in de advertenties.
De ondernemingen zijn op zoek naar andere gebieden, zoals de gezondheidszorg, waar geautomatiseerde call centers depressie of pijn in de stem van de beller zou kunnen detecteren, zelfs als de beller niet in staat is deze emoties uit te drukken.
Stemming detecteren
Een robot met de naam Forpheus, gemaakt door Omron Automation in Japan en gedemonstreerd tijdens CES in januari, speelt pingpong. Een deel van haar arsenaal van tafeltennisvaardigheden is haar vermogen om lichaamstaal te lezen om zowel de stemming en vaardigheid niveau van de menselijke tegenstander te achterhalen.
Het gaat natuurlijk niet om pingpong, maar het doel is industriële machines die “in harmonie” met de mens werken, wat zowel de productiviteit als de veiligheid verhoogt. Door bijvoorbeeld de lichaamstaal van fabrieksarbeiders te lezen, konden industriële robots anticiperen op hoe en waar mensen zich zouden kunnen bewegen.
Once aboard, CIMON—short for Crew Interactive MObile companioN—will assist the crew with its many activities. The point of this pilot project is to see if an artificially intelligent bot can improve crew efficiency and morale during longer missions, including a possible mission to Mars. What’s more, activities and tasks performed by ISS crew members are starting to get more complicated, so an AI could help. CIMON doesn’t have any arms or legs, so it can’t assist with any physical tasks, but it features a language user interface, allowing crew members to verbally communicate with it. The bot can display repair instructions on its screen, and even search for objects in the ISS. With a reduced workload, astronauts will hopefully experience less stress and have more time to relax.
CIMON with its development team prior to launch.
Image: DLR
CIMON was built by Airbus under a contract awarded by the German Aerospace Center (DLR). It has 12 internal fans, which allows the bot to move in all directions as it floats in microgravity. CIMON can move freely, and perform rotational movements such as shaking its head back-and-forth in disapproval. CIMON’s AI language and comprehension system is derived from IBM’s Watson Technology, and it responds to commands in English. CIMON cost less than $6 million to build, and less than two years to develop.
The pilot project will be led by DLR astronaut Alexander Gerst, who arrived on the ISS about a month ago. CIMON is already familiar with Gerst’s face and voice, so the bot will work best with him, at least initially. The German astronaut will use CIMON to see if the bot will increase his efficiency and effectiveness as he works on various experiments.
Indeed, with CIMON floating nearby, the ISS astronauts could easily call upon the bot for assistance, which they can do by calling out its name. They can request that CIMON display documents and media in their field of view, or record and playback experiments with its onboard camera. In general, the bot should speed up tasks on the ISS that require hands-on work.
The round robot features no sharp edges, so it poses no threat to equipment or crew. Should it start to go squirrely and use it’s best HAL-9000 imitation to say something like, “I’m sorry, Alexander, I’m afraid I can’t do that,” the bot is equipped with a kill switch. But hopefully it won’t come to that; unlike HAL, CIMON has been programmed with an ISTJ personality, meaning “introverted, sensing, thinking, and judging.” Its developers chose a face to make it more personable and relatable, and it can even sense the tone of the crew’s conversation. CIMON smiles when the mood is upbeat, and frowns or cries when things are sad. It supposedly behaves like R2D2, and can even quote famous sci-fi movies like E.T. the Extra-Terrestrial.
The search giant said Wednesday it’s beginning public testing of the software, which debuted in May and which is designed to make calls to businesses and book appointments. Duplex instantly raised questions over the ethics and privacy implications of using an AI assistant to hold lifelike conversations for you.
Google says its plan is to start its public trial with a small group of “trusted testers” and businesses that have opted into receiving calls from Duplex. Over the “coming weeks,” the software will only call businesses to confirm business and holiday hours, such as open and close times for the Fourth of July. People will be able to start booking reservations at restaurants and hair salons starting “later this summer.”
It took nearly a century of trial and error for human scientists to organize the periodic table of elements, arguably one of the greatest scientific achievements in chemistry, into its current form.
A new artificial intelligence (AI) program developed by Stanford physicists accomplished the same feat in just a few hours.
Called Atom2Vec, the program successfully learned to distinguish between different atoms after analyzing a list of chemical compound names from an online database. The unsupervised AI then used concepts borrowed from the field of natural language processing – in particular, the idea that the properties of words can be understood by looking at other words surrounding them – to cluster the elements according to their chemical properties.
“We wanted to know whether an AI can be smart enough to discover the periodic table on its own, and our team showed that it can,” said study leader Shou-Cheng Zhang, the J. G. Jackson and C. J. Wood Professor of Physics at Stanford’s School of Humanities and Sciences.
The bots learn from self-play, meaning two bots playing each other and learning from each side’s successes and failures. By using a huge stack of 256 graphics processing units (GPUs) with 128,000 processing cores, the researchers were able to speed up the AI’s gameplay so that they learned from the equivalent of 180 years of gameplay for every day it trained. One version of the bots were trained for four weeks, meaning they played more than 5,000 years of the game.
[…]
In a match, the OpenAI team initially gives each bot a mandate to do as well as it can on its own, meaning that the bots learned to act selfishly and steal kills from each other. But by turning up a simple metric, a weighted average of the team’s success, the bots soon begin to work together and execute team attacks quicker than humanly possible. The metric was dubbed by OpenAI as “team spirit.”
“They start caring more about team fighting, and saving one another, and working together in these skirmishes in order to make larger advances towards the group goal,” says Brooke Chan, an engineer at OpenAI.
Right now, the bots are restricted to playing certain characters, can’t use certain items like wards that allow players to see more of the map or anything that grants invisibility, or summon other units to help them fight with spells. OpenAI hopes to lift those restrictions by the competition in August.
Among our exciting announcements at //Build, one of the things I was thrilled to launch is the AI Lab – a collection of AI projects designed to help developers explore, experience, learn about and code with the latest Microsoft AI Platform technologies.
What is AI Lab?
AI Lab helps our large fast-growing community of developers get started on AI. It currently houses five projects that showcase the latest in custom vision, attnGAN (more below), Visual Studio tools for AI, Cognitive Search, machine reading comprehension and more. Each lab gives you access to the experimentation playground, source code on GitHub, a crisp developer-friendly video, and insights into the underlying business problem and solution. One of the projects we highlighted at //Build was the search and rescue challenge which gave the opportunity to developers worldwide to use AI School resources to build and deploy their first AI model for a problem involving aerial drones.
A group of scientists have built a neural network to sniff out any unusual nuclear activity. Researchers from the Pacific Northwest National Laboratory (PNNL), one of the United States Department of Energy national laboratories, decided to see if they could use deep learning to sort through the different nuclear decay events to identify any suspicious behavior.
The lab, buried beneath 81 feet of concrete, rock and earth, is blocked out from energy from cosmic rays, electronics and other sources. It means that the data collected is less noisy, making it easier to pinpoint unusual activity.
The system looks for electrons emitted and scattered from radioactive particles decaying, and monitor the abundance of argon-37, a radioactive isotope of argon-39 that is created synthetically through nuclear explosions.
Argon-37 which has a half-life of 35 days, is emitted when calcium captures excess neutrons and decays by emitting an alpha particle. Emily Mace, a scientist at PNNL, said she looks for the energy, timing, duration and other features of the decay events to see if it’s from nuclear testing.
“Some pulse shapes are difficult to interpret,” said Mace. “It can be challenging to differentiate between good and bad data.”
Deep learning makes that process easier. Computer scientists collected 32,000 pulses and annotated their properties, teaching the system to spot any odd features that might classify a signal as ‘good’ or ‘bad’.
“Signals can be well behaved or they can be poorly behaved,” said Jesse Ward. “For the network to learn about the good signals, it needs a decent amount of bad signals for comparison.” When the researchers tested their system with 50,000 pulses and asked human experts to differentiate signals, the neural network agreed with them 100 per cent of the time.
It also correctly identified 99.9 per cent of the pulses compared to 96.1 per cent from more conventional techniques.
The AI, called Project Debater, appeared on stage in a packed conference room at IBM’s San Francisco office embodied in a 6ft tall black panel with a blue, animated “mouth”. It was a looming presence alongside the human debaters Noa Ovadia and Dan Zafrir, who stood behind a podium nearby.
Although the machine stumbled at many points, the unprecedented event offered a glimpse into how computers are learning to grapple with the messy, unstructured world of human decision-making.
For each of the two short debates, participants had to prepare a four-minute opening statement, followed by a four-minute rebuttal and a two-minute summary. The opening debate topic was “we should subsidize space exploration”, followed by “we should increase the use of telemedicine”.
In both debates, the audience voted Project Debater to be worse at delivery but better in terms of the amount of information it conveyed. And in spite of several robotic slip-ups, the audience voted the AI to be more persuasive (in terms of changing the audience’s position) than its human opponent, Zafrir, in the second debate.
It’s worth noting, however, that there were many members of IBM staff in the room and they may have been rooting for their creation.
IBM hopes the research will eventually enable a more sophisticated virtual assistant that can absorb massive and diverse sets of information to help build persuasive arguments and make well-informed decisions – as opposed to merely responding to simple questions and commands.
Project Debater was a showcase of IBM’s ability to process very large data sets, including millions of news articles across dozens of subjects, and then turn snippets of arguments into full flowing prose – a challenging task for a computer.
[…]
Once an AI is capable of persuasive arguments, it can be applied as a tool to aid human decision-making.
“We believe there’s massive potential for good in artificial intelligence that can understand us humans,” said Arvind Krishna, director of IBM Research.
One example of this might be corporate boardroom decisions, where there are lots of conflicting points of view. The AI system could, without emotion, listen to the conversation, take all of the evidence and arguments into account and challenge the reasoning of humans where necessary.
“This can increase the level of evidence-based decision-making,” said Reed, adding that the same system could be used for intelligence analysis in counter-terrorism, for example identifying if a particular individual represents a threat.
In both cases, the machine wouldn’t make the decision but would contribute to the discussion and act as another voice at the table.
Essentially, Project Debater assigns a confidence score to every piece of information it understands. As in: how confident is the system that it actually understands the content of what’s being discussed? “If it’s confident that it got that point right, if it really believes it understands what that opponent was saying, it’s going to try to make a very strong argument against that point specifically,” Welser explains.
”If it’s less confident,” he says, “it’ll do it’s best to make an argument that’ll be convincing as an argument even if it doesn’t exactly answer that point. Which is exactly what a human does too, sometimes.”
So: the human says that government should have specific criteria surrounding basic human needs to justify subsidization. Project Debater responds that space is awesome and good for the economy. A human might choose that tactic as a sneaky way to avoid debating on the wrong terms. Project Debater had different motivations in its algorithms, but not that different.
The point of this experiment wasn’t to make me think that I couldn’t trust that a computer is arguing in good faith — though it very much did that. No, the point is that IBM showing off that it can train AI in new areas of research that could eventually be useful in real, practical contexts.
The first is parsing a lot of information in a decision-making context. The same technology that can read a corpus of data and come up with a bunch of pros and cons for a debate could be (and has been) used to decide whether or not a stock might be worth investing in. IBM’s system didn’t make the value judgement, but it did provide a bunch of information to the bank showing both sides of a debate about the stock.
As for the debating part, Welser says that it “helps us understand how language is used,” by teaching a system to work in a rhetorical context that’s more nuanced than the usual Hey Google give me this piece of information and turn off my lights. Perhaps it might someday help a lawyer structure their arguments, “not that Project Debater would make a very good lawyer,” he joked. Another IBM researcher suggested that this technology could help judge fake news.
How close is this to being something IBM turns into a product? “This is still a research level project,” Welser says, though “the technologies underneath it right now” are already beginning to be used in IBM projects.
The system listened to four minutes of its human opponent’s opening remarks, then parsed that data and created an argument that highlighted and attempted to debunk information shared by the opposing side. That’s incredibly impressive because it has to understand not only the words but the context of those words. Parroting back Wikipedia entries is easy, taking data and creating a narrative that’s based not only on raw data but also takes into account what it’s just heard? That’s tough.
In a world where emotion and bias colors all our decisions, Project Debater could help companies and governments see through the noise of our life experiences and produce mostly impartial conclusions. Of course, the data set it pulls from is based on what humans have written and those will have their own biases and emotion.
While the goal is an unbiased machine, during the discourse Project Debate wasn’t completely sterile. Amid its rebuttal against debater Dan Zafrir, while they argued about telemedicine expansion, the system stated that Zafrir had not told the truth during his opening statement about the increase in the use of telemedicine. In other words, it called him a liar.
When asked about the statement, Slonim said that the system has a confidence threshold during rebuttals. If it’s feeling very confident it creates a more complex statement. If it’s feeling less confident, the statement is less impressive.
An artificially intelligent system has been demonstrated generating URLs for phishing websites that appear to evade detection by security tools.
Essentially, the software can come up with URLs for webpages that masquerade as legit login pages for real websites, when in actual fact, the webpages simply collect the entered username and passwords to later hijack accounts.
Blacklists and algorithms – intelligent or otherwise – can be used to automatically identify and block links to phishing pages. Humans should be able to spot that the web links are dodgy, but not everyone is so savvy.
Using the Phishtank database, a group of computer scientists from Cyxtera Technologies, a cybersecurity biz based in Florida, USA, have built <a target=”_blank” rel=”nofollow” href=”“>DeepPhish, which is machine-learning software that, allegedly, generates phishing URLs that beat these defense mechanisms.
[…]
The team inspected more than a million URLs on Phishtank to identify three different phishing miscreants who had generated webpages to steal people’s credentials. The team fed these web addresses into AI-based phishing detection algorithm to measure how effective the URLs were at bypassing the system.
The first scumbag of the trio used 1,007 attack URLs, and only 7 were effective at avoiding setting off alarms, across 106 domains, making it successful only 0.69 per cent of the time. The second one had 102 malicious web addresses, across 19 domains. Only five of them bypassed the threat detection algorithm and it was effective 4.91 per cent of the time.
Next, they fed this information into a Long-Short Term Memory network (LSTM) to learn the general structure and extract features from the malicious URLs – for example the second threat actor commonly used “tdcanadatrustindex.html” in its address.
All the text from effective URLs were taken to create sentences and encoded into a vector and fed into the LSTM, where it is trained to predict the next character given the previous one.
Over time it learns to generate a stream of text to simulate a list of pseudo URLs that are similar to the ones used as input. When DeepPhish was trained on data from the first threat actor, it also managed to create 1,007 URLs, and 210 of them were effective at evading detection, bumping up the score from 0.69 per cent to 20.90 per cent.
When it was following the structure from the second threat actor, it also produced 102 fake URLs and 37 of them were successful, increasing the likelihood of tricking the existent defense mechanism from 4.91 per cent to 36.28 per cent.
The effectiveness rate isn’t very high as a lot of what comes out the LSTM is effective gibberish, containing strings of forbidden characters.
“It is important to automate the process of retraining the AI phishing detection system by incorporating the new synthetic URLs that each threat actor may create,” the researchers warned. ®
:format(webp)/cdn.vox-cdn.com/uploads/chorus_asset/file/11641941/CTF_Fig_Tagging_180703_r01.width_1500.png)
