A team of computer scientists have built the smallest completely autonomous nano-drone that can control itself without the need for a human guidance.

Although computer vision has improved rapidly thanks to machine learning and AI, it remains difficult to deploy algorithms on devices like drones due to memory, bandwidth and power constraints.

But researchers from ETH Zurich, Switzerland and the University of Bologna, Italy have managed to build a hand-sized drone that can fly autonomously and consumes only about 94 milliWatts (0.094 W) of energy. Their efforts were published in a paper on arXiv earlier this month.

At the heart of it all is DroNet, a convolutional neural network that processes incoming images from a camera at 20 frames per second. It works out the steering angle, so that it can control the direction of the drone, and the probability of a collision, so that it know whether to keep going or stop. Training was conducted using thousands of images taken from bicycles and cars driving along different roads and streets.

[…]

But it suffers from some of the same setbacks as the older model. Since it was trained with images from a single plane, the drone can only move horizontally and cannot fly up or down.

Autonomous drones are desirable because if we’re going to use drones to do things like deliver packages, it would be grand if they could avoid obstacles instead of flying on known-safe routes. Autonomy will also help drones to monitor environments, spy on people and develop swarm intelligence for military use.

Source: Boffins build smallest drone to fly itself with AI • The Register

Now that DeepMind has solved Go, the company is applying DeepMind to navigation. Navigation relies on knowing where you are in space relative to your surroundings and continually updating that knowledge as you move. DeepMind scientists trained neural networks to navigate like this in a square arena, mimicking the paths that foraging rats took as they explored the space. The networks got information about the rat’s speed, head direction, distance from the walls, and other details. To researchers’ surprise, the networks that learned to successfully navigate this space had developed a layer akin to grid cells. This was surprising because it is the exact same system that mammalian brains use to navigate.

A few different cell populations in our brains help us make our way through space. Place cells are so named because they fire when we pass through a particular place in our environment relative to familiar external objects. They are located in the hippocampus—a brain region responsible for memory formation and storage—and are thus thought to provide a cellular place for our memories. Grid cells got their name because they superimpose a hypothetical hexagonal grid upon our surroundings, as if the whole world were overlaid with vintage tiles from the floor of a New York City bathroom. They fire whenever we pass through a node on that grid.

More DeepMind experiments showed that only the neural networks that developed layers that “resembled grid cells, exhibiting significant hexagonal periodicity (gridness),” could navigate more complicated environments than the initial square arena, like setups with multiple rooms. And only these networks could adjust their routes based on changes in the environment, recognizing and using shortcuts to get to preassigned goals after previously closed doors were opened to them.

Implications

These results have a couple of interesting ramifications. One is the suggestion that grid cells are the optimal way to navigate. They didn’t have to emerge here—there was nothing dictating their formation—yet this computer system hit upon them as the best solution, just like our biological system did. Since the evolution of any system, cell type, or protein can proceed along multiple parallel paths, it is very much not a given that the system we end up with is in any way inevitable or optimized. This report seems to imply that, with grid cells, that might actually be the case.

Another implication is the support for the idea that grid cells function to impose a Euclidian framework upon our surroundings, allowing us to find and follow the most direct route to a (remembered) destination. This function had been posited since the discovery of grid cells in 2005, but it had not yet been proven empirically. DeepMind’s findings provide a biological bolster for the idea floated by Kant in the 18th century that our perception of place is an innate ability, independent of experience.

Source: AI trained to navigate develops brain-like location tracking | Ars Technica

Ultimately, AI systems are only useful and safe as long as the goals they’ve learned actually mesh with what humans want them to do, and it can often be hard to know if they’ve subtly learned to solve the wrong problems or make bad decisions in certain conditions.

To make AI easier for humans to understand and trust, researchers at the nonprofit research organization OpenAI have proposed training algorithms to not only classify data or make decisions, but to justify their decisions in debates with other AI programs in front of a human or AI judge.

“Given a question or proposed action, two agents take turns making short statements up to a limit, then a human judges which of the agents gave the most true, useful information,” write OpenAI researchers Geoffrey Irving, Paul Christiano and Dario Amodei in a new research paper. The San Francisco-based AI lab is funded by Silicon Valley luminaries including Y Combinator President Sam Altman and Tesla CEO Elon Musk, with a goal of building safe, useful AI to benefit humanity.

Since human time is valuable and usually limited, the researchers say the AI systems can effectively train themselves in part by debating in front of an AI judge designed to mimic human decision making, similar to how software that plays games like Go or chess often trains in part by playing against itself.

In an experiment described in their paper, the researchers set up a debate where two software agents work with a standard set of handwritten numerals, attempting to convince an automated judge that a particular image is one digit rather than another digit, by taking turns revealing one pixel of the digit at a time. One bot is programmed to tell the truth, while another is programmed to lie about what number is in the image, and they reveal pixels to support their contentions that the digit is, say, a five rather than a six.

The truth-telling bots tend to reveal pixels from distinctive parts of the digit, like the horizontal line at the top of the numeral “5,” while the lying bots, in an attempt to deceive the judge, point out what amount to the most ambiguous areas, like the curve at the bottom of both a “5” and a “6.” The judge ultimately “guesses” which bot is telling the truth based on the pixels that have been revealed.The image classification task, where most of the image is invisible to the judge, is a sort of stand-in for complex problems where it wouldn’t be possible for a human judge to analyze the entire dataset to judge bot performance. The judge would have to rely on the facets of the data highlighted by debating robots, the researchers say.

“The goal here is to model situations where we have something that’s beyond human scale,” says Irving, a member of the AI safety team at OpenAI. “The best we can do there is replace something a human couldn’t possibly do with something a human can’t do because they’re not seeing an image.”

[…]

To test their hypothesis—that two debaters can lead to honest behavior even if the debaters know much more than the judge—the researchers have also devised an interactive demonstration of their approach, played entirely by humans and now available online. In the game, two human players are shown an image of either a dog or a cat and argue before a judge as to which species is represented. The contestants are allowed to highlight rectangular sections of the image to make their arguments—pointing out, for instance, a dog’s ears or cat’s paws—but the judge can “see” only the shapes and positions of the rectangles, not the actual image. While the honest player is required to tell the truth about what animal is shown, he or she is allowed to tell other lies in the course of the debate. “It is an interesting question whether lies by the honest player are useful,” the researchers write.

[…]

The researchers emphasize that it’s still early days, and the debate-based method still requires plenty of testing before AI developers will know exactly when it’s an effective strategy or how best to implement it. For instance, they may find that it may be better to use single judges or a panel of voting judges, or that some people are better equipped to judge certain debates.

It also remains to be seen whether humans will be accurate judges of sophisticated robots working on more sophisticated problems. People might be biased to rule in a certain way based on their own beliefs, and there could be problems that are hard to reduce enough to have a simple debate about, like the soundness of a mathematical proof, the researchers write.

Other less subtle errors may be easier to spot, like the sheep that Shane noticed had been erroneously labeled by Microsoft’s algorithms. “The agent would claim there’s sheep and point to the nonexistent sheep, and the human would say no,” Irving writes in an email to Fast Company.

But deceitful bots might also learn to appeal to human judges in sophisticated ways that don’t involve offering rigorous arguments, Shane suggested. “I wonder if we’d get kind of demagogue algorithms that would learn to exploit human emotions to argue their point,” she says.

Source: Why Scientists Think AI Systems Should Debate Each Other

Researchers at Endgame, a cyber-security biz based in Virginia, have published what they believe is the first large open-source dataset for machine learning malware detection known as EMBER.

EMBER contains metadata describing 1.1 million Windows portable executable files: 900,000 training samples evenly split into malicious, benign, and unlabeled categories and 200,000 files of test samples labelled as malicious and benign.

“We’re trying to push the dark arts of infosec research into an open light. EMBER will make AI research more transparent and reproducible,” Hyrum Anderson, co-author of the study to be presented at the RSA conference this week in San Francisco, told The Register.

Progress in AI is driven by data. Researchers compete with one another by building models and training them on benchmark datasets to reach ever increasing accuracies.

Computer vision is flooded with numerous datasets containing millions of annotated pictures for image recognition tasks, and natural language processing has various text-based datasets to test machine reading and comprehension skills. this has helped a lot in building AI image processing.

Although there is a strong interest in using AI for information security – look at DARPA’s Cyber Grand Challenge where academics developed software capable of hunting for security bugs autonomously – it’s an area that doesn’t really have any public datasets.

Source: Infosec brainiacs release public dataset to classify new malware using AI • The Register

researchers at software infrastructure firm Pivotal have taught AI to locate this accidentally public sensitive information in a surprising way: By looking at the code as if it were a picture. Since modern artificial intelligence is arguably better than humans at identifying minute differences in images, telling the difference between a password and normal code for a computer is just like recognizing a dog from a cat.

The best way to check whether private passwords or sensitive information has been left public today is to use hand-coded rules called “regular expressions.” These rules tell a computer to find any string of characters that meets specific criteria, like length and included characters. But passwords are all different, and this method means that the security engineer has to anticipate every kind of private data they want to guard against.

To automate the process, the Pivotal team first turned the text of passwords and code into matrixes, or lists of numbers describing each string of characters. This is the same process used when AI interprets images—similar to how the images reflected into our eyes are turned into electrical signals for the brain, images and text need to be in a simpler form for computers to process.

When the team visualized the matrices, private data looked different from the standard code. Since passwords or keys are often randomized strings of numbers, letters, and symbols—called “high entropy”—they stand out against non-random strings of letters.

Below you can see a GIF of the matrix with 100 characters of simulated secret information.

And then here’s another with 100 normal, non-secret code:

The two patterns are completely different, with patches of higher-entropy appearing lighter in the top example of “secret” data.

Pivotal then trained a deep learning algorithm typically used for images on the matrixes, and, according to Pivotal chief security officer Justin Smith, the end result performed better than the regular expressions the firm typically uses.

Source: Artificial intelligence can scour code to find accidentally public passwords — Quartz

In the city of Xichang, located in the southwestern Sichuan province, there is a massive, artificial intelligence-powered roach breeding farm that is producing more than six billion cockroaches per year.

The facility, which is described by the South China Morning Post as a multi-story building about the size of two sports fields, is being operated by Chengdu-based medicine maker Gooddoctor Pharmaceutical Group. Its existence raises a number of questions like, “Oh god, why?” and “Who asked for this monstrosity?”

Inside the breeding site, the environment is described as “warm, humid, and dark” all-year round. The layout is wide open, allowing the roaches to roam around freely, find food and water, and reproduce whenever and wherever the right mood strikes.

The insect sex pit is managed by what the South China Morning Post describes as a “smart manufacturing system” that is controlled primarily by algorithms. The system is in charge of analyzing more than 80 categories of data collected from throughout the facility. Everything from the temperature to the level of food consumption is monitored by AI, which is programmed to learn from historical data to determine the best conditions for peak roach fornication.

The billions of roaches that pass through the facility each year never get to see the light of day. From their birth inside the building until their death months or years later, they are locked within the walls of the moist coitus cabin.

Each and every one of the insects is eventually fed into machines and crushed up to be used in a “healing potion” manufactured by the pharmaceutical company responsible for the facility.

The potion—which is described as having a tea-like color, a slightly sweet taste, and a fishy smell—sells for about $8 for two 100ml bottles. While it is used primarily as a fix for stomach issues, the medicine can be prescribed by doctors for just about anything.

Source: This AI-Controlled Roach Breeding Site Is a Nightmare Factory

While autonomous robots with humanlike, all-encompassing capabilities are still decades away, European lawmakers, legal experts and manufacturers are already locked in a high-stakes debate about their legal status: whether it’s these machines or human beings who should bear ultimate responsibility for their actions.

The battle goes back to a paragraph of text, buried deep in a European Parliament report from early 2017, which suggests that self-learning robots could be granted “electronic personalities.” Such a status could allow robots to be insured individually and be held liable for damages if they go rogue and start hurting people or damaging property.

Those pushing for such a legal change, including some manufacturers and their affiliates, say the proposal is common sense. Legal personhood would not make robots virtual people who can get married and benefit from human rights, they say; it would merely put them on par with corporations, which already have status as “legal persons,” and are treated as such by courts around the world.

Source: Europe divided over robot ‘personhood’ – POLITICO

People are remarkably good at focusing their attention on a particular person in a noisy environment, mentally “muting” all other voices and sounds. Known as the cocktail party effect, this capability comes natural to us humans. However, automatic speech separation — separating an audio signal into its individual speech sources — while a well-studied problem, remains a significant challenge for computers. In “Looking to Listen at the Cocktail Party”, we present a deep learning audio-visual model for isolating a single speech signal from a mixture of sounds such as other voices and background noise. In this work, we are able to computationally produce videos in which speech of specific people is enhanced while all other sounds are suppressed. Our method works on ordinary videos with a single audio track, and all that is required from the user is to select the face of the person in the video they want to hear, or to have such a person be selected algorithmically based on context. We believe this capability can have a wide range of applications, from speech enhancement and recognition in videos, through video conferencing, to improved hearing aids, especially in situations where there are multiple people speaking.

A unique aspect of our technique is in combining both the auditory and visual signals of an input video to separate the speech. Intuitively, movements of a person’s mouth, for example, should correlate with the sounds produced as that person is speaking, which in turn can help identify which parts of the audio correspond to that person. The visual signal not only improves the speech separation quality significantly in cases of mixed speech (compared to speech separation using audio alone, as we demonstrate in our paper), but, importantly, it also associates the separated, clean speech tracks with the visible speakers in the video.

The input to our method is a video with one or more people speaking, where the speech of interest is interfered by other speakers and/or background noise. The output is a decomposition of the input audio track into clean speech tracks, one for each person detected in the video.

An Audio-Visual Speech Separation Model To generate training examples, we started by gathering a large collection of 100,000 high-quality videos of lectures and talks from YouTube. From these videos, we extracted segments with a clean speech (e.g. no mixed music, audience sounds or other speakers) and with a single speaker visible in the video frames. This resulted in roughly 2000 hours of video clips, each of a single person visible to the camera and talking with no background interference. We then used this clean data to generate “synthetic cocktail parties” — mixtures of face videos and their corresponding speech from separate video sources, along with non-speech background noise we obtained from AudioSet. Using this data, we were able to train a multi-stream convolutional neural network-based model to split the synthetic cocktail mixture into separate audio streams for each speaker in the video. The input to the network are visual features extracted from the face thumbnails of detected speakers in each frame, and a spectrogram representation of the video’s soundtrack. During training, the network learns (separate) encodings for the visual and auditory signals, then it fuses them together to form a joint audio-visual representation. With that joint representation, the network learns to output a time-frequency mask for each speaker. The output masks are multiplied by the noisy input spectrogram and converted back to a time-domain waveform to obtain an isolated, clean speech signal for each speaker. For full details, see our paper.

Our multi-stream, neural network-based model architecture.

Here are some more speech separation and enhancement results by our method, playing first the input video with mixed or noisy speech, then our results. Sound by others than the selected speakers can be entirely suppressed or suppressed to the desired level.

Application to Speech Recognition Our method can also potentially be used as a pre-process for speech recognition and automatic video captioning. Handling overlapping speakers is a known challenge for automatic captioning systems, and separating the audio to the different sources could help in presenting more accurate and easy-to-read captions.

You can similarly see and compare the captions before and after speech separation in all the other videos in this post and on our website, by turning on closed captions in the YouTube player when playing the videos (“cc” button at the lower right corner of the player). On our project web page you can find more results, as well as comparisons with state-of-the-art audio-only speech separation and with other recent audio-visual speech separation work. Indeed, with recent advances in deep learning, there is a clear growing interest in the academic community in audio-visual analysis. For example, independently and concurrently to our work, this work from UC Berkeley explored a self-supervised approach for separating speech of on/off-screen speakers, and this work from MIT addressed the problem of separating the sound of multiple on-screen objects (e.g., musical instruments), while locating the image regions from which the sound originates. We envision a wide range of applications for this technology. We are currently exploring opportunities for incorporating it into various Google products. Stay tuned!

Source: Research Blog: Looking to Listen: Audio-Visual Speech Separation

Transporting yourself into a video game, body and all, just got easier. Artificial intelligence has been used to create 3D models of people’s bodies for virtual reality avatars, surveillance, visualizing fashion, or movies. But it typically requires special camera equipment to detect depth or to view someone from multiple angles. A new algorithm creates 3D models using standard video footage from one angle.

The system has three stages. First, it analyzes a video a few seconds long of someone moving—preferably turning 360° to show all sides—and for each frame creates a silhouette separating the person from the background. Based on machine learning techniques—in which computers learn a task from many examples—it roughly estimates the 3D body shape and location of joints. In the second stage, it “unposes” the virtual human created from each frame, making them all stand with arms out in a T shape, and combines information about the T-posed people into one, more accurate model. Finally, in the third stage, it applies color and texture to the model based on recorded hair, clothing, and skin.

The researchers tested the method with a variety of body shapes, clothing, and backgrounds and found that it had an average accuracy within 5 millimeters, they will report in June at the Computer Vision and Pattern Recognition conference in Salt Lake City. The system can also reproduce the folding and wrinkles of fabric, but it struggles with skirts and long hair. With a model of you, the researchers can change your weight, clothing, and pose—and even make you perform a perfect pirouette. No practice necessary.

Source: Watch artificial intelligence create a 3D model of a person—from just a few seconds of video | Science | AAAS

Researchers have successfully trained artificial intelligence to generate new clips of the prehistoric animated series based on nothing but random text descriptions of what’s happening in a scene.

A team of researchers from the Allen Institute for Artificial Intelligence, and the University of Illinois Urbana-Champaign, trained an AI by feeding it over 25,000 three-second clips of the cartoon, which hasn’t seen any new episodes in over 50 years. Most AI experiments as of late have involved generating freaky images based on what was learned, but this time the researchers included detailed descriptions and annotations of what appeared, and what was happening, in every clip the AI ingested.

As a result, the new Flintstones animations generated by the Allen Institute’s AI aren’t just random collages of chopped up cartoons. Instead, the researchers are able to feed the AI a very specific description of a scene, and it outputs a short clip featuring the characters, props, and locations specified—most of the time.

The quality of the animations that are generated is awful at best; no one’s going to be fooled into thinking these are the Hanna-Barbera originals. But seeing an AI generate a cartoon, featuring iconic characters, all by itself, is a fascinating sneak peek at how some films and TV shows might be made one day.

Source: This AI Can Automatically Animate New Flintstones Cartoons

30.3 million Americans have diabetes according to a 2015 CDC study. An additional 84.1 million have prediabetes, which often leads to the full disease within five years. It’s important to detect diabetes early to avoid health complications like heart disease, stroke, amputation of extremities and vision loss. Technology increasingly plays an important role in early detection, too. In that vein, the US Food and Drug Administration (FDA) has just approved an AI-powered device that can be used by non-specialists to detect diabetic retinopathy in adults with diabetes.

Diabetic retinopathy occurs when the high levels of blood sugar in the bloodstream cause damage to your retina’s blood vessels. It’s the most common cause of vision loss, according to the FDA. The approval comes for a device called IDx-DR, a software program that uses an AI algorithm to analyze images of the eye that can be taken in a regular doctor’s office with a special camera, the Topcon NW400.

The photos are then uploaded to a server that runs IDx-DR, which can then tell the doctor if there is a more than mild level of diabetic retinopathy present. If not, it will advise a re-screen in 12 months. The device and software can be used by health care providers who don’t normally provide eye care services. The FDA warns that you shouldn’t be screened with the device if you have had laser treatment, eye surgery or injections, as well as those with other conditions, like persistent vision loss, blurred vision, floaters, previously diagnosed macular edema and more.

Source: FDA approves AI-powered software to detect diabetic retinopathy

Using well-established machine learning techniques, researchers from University of California, Berkeley have taught simulated humanoids to perform over 25 natural motions, from somersaults and cartwheels through to high leg kicks and breakdancing. The technique could lead to more realistic video gameplay and more agile robots.

[…]

UC Berkeley graduate student Xue Bin “Jason” Peng, along with his colleagues, have combined two techniques—motion-capture technology and deep-reinforcement computer learning—to create something completely new: a system that teaches simulated humanoids how to perform complex physical tasks in a highly realistic manner. Learning from scratch, and with limited human intervention, the digital characters learned how to kick, jump, and flip their way to success. What’s more, they even learned how to interact with objects in their environment, such as barriers placed in their way or objects hurled directly at them.

[…]

The new system, dubbed DeepMimic, works a bit differently. Instead of pushing the simulated character towards a specific end goal, such as walking, DeepMimic uses motion-capture clips to “show” the AI what the end goal is supposed to look like. In experiments, Bin’s team took motion-capture data from more than 25 different physical skills, from running and throwing to jumping and backflips, to “define the desired style and appearance” of the skill, as Peng explained at the Berkeley Artificial Intelligence Research (BAIR) blog.

Results didn’t happen overnight. The virtual characters tripped, stumbled, and fell flat on their faces repeatedly until they finally got the movements right. It took about a month of simulated “practice” for each skill to develop, as the humanoids went through literally millions of trials trying to nail the perfect backflip or flying leg kick. But with each failure came an adjustment that took it closer to the desired goal.

Using this technique, the researchers were able to produce agents who behaved in a highly realistic, natural manner. Impressively, the bots were also able to manage never-before-seen conditions, such as challenging terrain or obstacles. This was an added bonus of the reinforcement learning, and not something the researchers had to work on specifically.

“We present a conceptually simple [reinforcement learning] framework that enables simulated characters to learn highly dynamic and acrobatic skills from reference motion clips, which can be provided in the form of mocap data [i.e. motion capture] recorded from human subjects,” writes Peng. “Given a single demonstration of a skill, such as a spin-kick or a backflip, our character is able to learn a robust policy to imitate the skill in simulation. Our policies produce motions that are nearly indistinguishable from mocap,” adding that “We’re moving toward a virtual stuntman.”

Not to be outdone, the researchers used DeepMimic to create realistic movements from simulated lions, dinosaurs, and mythical beasts. They even created a virtual version of ATLAS, the humanoid robot voted most likely to destroy humanity. This platform could conceivably be used to produce more realistic computer animation, but also for virtual testing of robots.

Source: After Millions of Trials, These Simulated Humans Learned to Do Perfect Backflips and Cartwheels

Researchers can estimate your expected lifespan based on physiological traits like your genes or your circulating blood factor, but that’s not very practical on a grand scale. There may be a shortcut, however: the devices you already have on your body. Russian scientists have crafted an AI-based algorithm that uses the activity tracking from smartphones and smartwatches to estimate your lifespan with far greater precision than past models.

The team used a convolutional neural network to find the “biologically relevant” motion patterns in a large set of US health survey data and correlate that to both lifespans and overall health. It would look for not just step counts, but how often you switch between active and inactive periods — many of the other factors in your life, such as your sleeping habits and gym visits, are reflected in those switches. After that, it was just a matter of applying the understanding to a week’s worth of data from test subjects’ phones. You can even try it yourself through Gero Lifespan, an iPhone app that uses data from Apple Health, Fitbit and Rescuetime (a PC productivity measurement app) to predict your longevity.

This doesn’t provide a full picture of your health, as it doesn’t include your diet, genetics and other crucial factors. Doctors would ideally use both mobile apps and clinical analysis to give you a proper estimate, and the scientists are quick to acknowledge that what you see here isn’t completely ready for medical applications. The AI is still more effective than past approaches, though, and it could be useful for more accurate health risk models that help everything from insurance companies (which already use activity tracking as an incentive) to the development of anti-aging treatments.

Source: AI predicts your lifespan using activity tracking apps

No idea what the percentages are though

On Thursday, Emmanuel Macron, the president of France, gave a speech laying out a new national strategy for artificial intelligence in his country. The French government will spend €1.5 billion ($1.85 billion) over five years to support research in the field, encourage startups, and collect data that can be used, and shared, by engineers. The goal is to start catching up to the US and China and to make sure the smartest minds in AI—hello Yann LeCun—choose Paris over Palo Alto.Directly after his talk, he gave an exclusive and extensive interview, entirely in English, to WIRED Editor-in-Chief Nicholas Thompson about the topic and why he has come to care so passionately about it.

[…]

: AI will raise a lot of issues in ethics, in politics, it will question our democracy and our collective preferences. For instance, if you take healthcare: you can totally transform medical care making it much more predictive and personalized if you get access to a lot of data. We will open our data in France. I made this decision and announced it this afternoon. But the day you start dealing with privacy issues, the day you open this data and unveil personal information, you open a Pandora’s Box, with potential use cases that will not be increasing the common good and improving the way to treat you. In particular, it’s creating a potential for all the players to select you. This can be a very profitable business model: this data can be used to better treat people, it can be used to monitor patients, but it can also be sold to an insurer that will have intelligence on you and your medical risks, and could get a lot of money out of this information. The day we start to make such business out of this data is when a huge opportunity becomes a huge risk. It could totally dismantle our national cohesion and the way we live together. This leads me to the conclusion that this huge technological revolution is in fact a political revolution.

When you look at artificial intelligence today, the two leaders are the US and China. In the US, it is entirely driven by the private sector, large corporations, and some startups dealing with them. All the choices they will make are private choices that deal with collective values. That’s exactly the problem you have with Facebook and Cambridge Analytica or autonomous driving. On the other side, Chinese players collect a lot of data driven by a government whose principles and values are not ours. And Europe has not exactly the same collective preferences as US or China. If we want to defend our way to deal with privacy, our collective preference for individual freedom versus technological progress, integrity of human beings and human DNA, if you want to manage your own choice of society, your choice of civilization, you have to be able to be an acting part of this AI revolution . That’s the condition of having a say in designing and defining the rules of AI. That is one of the main reasons why I want to be part of this revolution and even to be one of its leaders. I want to frame the discussion at a global scale.

[…]

I want my country to be the place where this new perspective on AI is built, on the basis of interdisciplinarity: this means crossing maths, social sciences, technology, and philosophy. That’s absolutely critical. Because at one point in time, if you don’t frame these innovations from the start, a worst-case scenario will force you to deal with this debate down the line. I think privacy has been a hidden debate for a long time in the US. Now, it emerged because of the Facebook issue. Security was also a hidden debate of autonomous driving. Now, because we’ve had this issue with Uber, it rises to the surface. So if you don’t want to block innovation, it is better to frame it by design within ethical and philosophical boundaries. And I think we are very well equipped to do it, on top of developing the business in my country.

But I think as well that AI could totally jeopardize democracy. For instance, we are using artificial intelligence to organize the access to universities for our students That puts a lot of responsibility on an algorithm. A lot of people see it as a black box, they don’t understand how the student selection process happens. But the day they start to understand that this relies on an algorithm, this algorithm has a specific responsibility. If you want, precisely, to structure this debate, you have to create the conditions of fairness of the algorithm and of its full transparency. I have to be confident for my people that there is no bias, at least no unfair bias, in this algorithm. I have to be able to tell French citizens, “OK, I encouraged this innovation because it will allow you to get access to new services, it will improve your lives—that’s a good innovation to you.” I have to guarantee there is no bias in terms of gender, age, or other individual characteristics, except if this is the one I decided on behalf of them or in front of them. This is a huge issue that needs to be addressed. If you don’t deal with it from the very beginning, if you don’t consider it is as important as developing innovation, you will miss something and at a point in time, it will block everything. Because people will eventually reject this innovation.

[…]

your algorithm and be sure that this is trustworthy.” The power of consumption society is so strong that it gets people to accept to provide a lot of personal information in order to get access to services largely driven by artificial intelligence on their apps, laptops and so on. But at some point, as citizens, people will say, “I want to be sure that all of this personal data is not used against me, but used ethically, and that everything is monitored. I want to understand what is behind this algorithm that plays a role in my life.” And I’m sure that a lot of startups or labs or initiatives which will emerge in the future, will reach out to their customers and say “I allow you to better understand the algorithm we use and the bias or non-bias.” I’m quite sure that’s one of the next waves coming in AI. I think it will increase the pressure on private players. These new apps or sites will be able to tell people: “OK! You can go to this company or this app because we cross-check everything for you. It’s safe,” or on the contrary: “If you go to this website or this app or this research model, it’s not OK, I have no guarantee, I was not able to check or access the right information about the algorithm”.

Source: Emmanuel Macron Q&A: France’s President Discusses Artificial Intelligence Strategy | WIRED

A team of astroboffins have built artificial neural networks that estimate the probability of exoplanets harboring alien life.

The research was presented during a talk on Wednesday at the European Week of Astronomy and Space Science in Liverpool, United Kingdom.

The neural network works by classifying planets into five different conditions: the present-day Earth, the early Earth, Mars, Venus or Saturn’s moon Titan. All of these objects have a rocky core and an atmosphere, two requirements scientists believe are necessary for sustaining the right environments for life to blossom.

To train the system, researchers collected the spectral data that describes what chemical elements are present in a planet’s atmosphere of a planet. They then created hundreds of these “atmospheric profiles” as inputs and the neural network then gives a rough estimate of the probability that a particular planet might support life by classifying it into those five types.

If a planet is judged as Earth-like, it means it has a high probability of life. But if it’s classified as being closer to Venus, then the chances are lower.

“We’re currently interested in these artificial neural networks (ANNs) for prioritising exploration for a hypothetical, intelligent, interstellar spacecraft scanning an exoplanet system at range,” said Christopher Bishop, a PhD student at Plymouth University.

“We’re also looking at the use of large area, deployable, planar Fresnel antennas to get data back to Earth from an interstellar probe at large distances. This would be needed if the technology is used in robotic spacecraft in the future.”

Experimental

At the moment, however, the ANN is more of a proof of concept. Angelo Cangelosi, professor of artificial intelligence and cognition at Plymouth University and the supervisor of the project, said initial results seem promising.

“Given the results so far, this method may prove to be extremely useful for categorizing different types of exoplanets using results from ground–based and near Earth observatories.”

There are a couple exoplanet-hunting telescopes that will use spectroscopy to analyze a planet’s chemical composition that are expected to be launched in the near future.

NASA’s Transiting Exoplanet Satellite Survey (TESS) will monitor the brightest stars in the sky to look for periodic dips in brightness when an orbiting planet crosses its path. The European Space Agency also announced Ariel, a mission that uses infrared to find exoplanets.

The Kepler Space Telescope is already looking for new candidates – although it’s set to retire soon – and is also looking for similar data. It is hoped by analyzing the spectral data for exoplanets, it could aid scientists in choosing better targets for future missions, where spacecraft can be sent to more detailed observations

Source: Is there alien life out there? Let’s turn to AI, problem solver du jour • The Register

The thing about ML models is that shit in leads to shit out. We have no data on inhabited planets apart from Earth, so it seems to me that the assumptions these guys are making aren’t worth a damn.

Analysis IBM boasts that machine learning is not just quicker on its POWER servers than on TensorFlow in the Google Cloud, it’s 46 times quicker.

Back in February Google software engineer Andreas Sterbenz wrote about using Google Cloud Machine Learning and TensorFlow on click prediction for large-scale advertising and recommendation scenarios.

He trained a model to predict display ad clicks on Criteo Labs clicks logs, which are over 1TB in size and contain feature values and click feedback from millions of display ads.

Data pre-processing (60 minutes) was followed by the actual learning, using 60 worker machines and 29 parameter machines for training. The model took 70 minutes to train, with an evaluation loss of 0.1293. We understand this is a rough indicator of result accuracy.

Sterbenz then used different modelling techniques to get better results, reducing the evaluation loss, which all took longer, eventually using a deep neural network with three epochs (a measure of the number of times all of the training vectors are used once to update the weights), which took 78 hours.

[…]

Thomas Parnell and Celestine Dünner at IBM Research in Zurich used the same source data – Criteo Terabyte Click Logs, with 4.2 billion training examples and 1 million features – and the same ML model, logistic regression, but a different ML library. It’s called Snap Machine Learning.

They ran their session using Snap ML running on four Power System AC922 servers, meaning eight POWER9 CPUs and 16 Nvidia Tesla V100 GPUs. Instead of taking 70 minutes, it completed in 91.5 seconds, 46 times faster.

They prepared a chart showing their Snap ML, the Google TensorFlow and three other results:

A 46x speed improvement over TensorFlow is not to be sneezed at. What did they attribute it to?

They say Snap ML features several hierarchical levels of parallelism to partition the workload among different nodes in a cluster, takes advantage of accelerator units, and exploits multi-core parallelism on the individual compute units

1. First, data is distributed across the individual worker nodes in the cluster
2. On a node data is split between the host CPU and the accelerating GPUs with CPUs and GPUs operating in parallel
3. Data is sent to the multiple cores in a GPU and the CPU workload is multi-threaded

Snap ML has nested hierarchical algorithmic features to take advantage of these three levels of parallelism.

Source: IBM claims its machine learning library is 46x faster than TensorFlow • The Register

. As research into AI grows ever more ambitious and complex, these robot brains will challenge the fundamental assumptions of how we humans do things. And, as ever, the only true law of robotics is that computers will always do literally, exactly what you tell them to.

A paper recently published to ArXiv highlights just a handful of incredible and slightly terrifying ways that algorithms think. These AI were designed to reflect evolution by simulating generations while other competing algorithms conquered problems posed by their human masters with strange, uncanny, and brilliant solutions.

My favorite example of when AI surprised us, the scientists wielding it. A corner case in our algo challenged AI to learn to walk without touching its feet to the ground. Many more in https://t.co/7690RlnTwI from @joelbot3000with @jb_mouret @CULLYAntoine Cully et al Nature 2015 pic.twitter.com/YtzV7zgwhb

— Jeff Clune (@jeffclune) March 13, 2018

The Surprising Creativity of Digital Evolution: A Collection of Anecdotes from the Evolutionary Computation and Artificial Life Research Communities covers some 27 anecdotes from various computer science projects and is worth a read on its own, but here are a few highlights:

• A study designed to evolve moving creatures generated ‘hackers’ that would break their simulation by clipping into the ground and using the “free energy” of the simulation’s correction to speed towards their goal.
• An AI project which pit programs against each other in games of five-in-a-row Tic-Tac-Toe on an infinitely expansive board surfaced the extremely successful method of requesting moves involving extremely long memory addresses which would crash the opponent’s computer and award a win by default.
• A program designed to simulate efficient ways of braking an aircraft as it landed on an aircraft carrier learned that by maximizing the force on landing—the opposite of its actual goal—the variable holding that value would overflow and flip to zero, creating a practically catastrophic, but technically perfect solution.
• A test that challenged a simulated robot to walk without allowing its feet to touch the ground saw the robot flip on its back and walk on its elbows (or knees?) as shown in the tweet above.
• A study to evolve a simulated creature that could jump as high as possible yielded top-heavy creatures on tiny poles that would fall over and spin in mid-air for a technically high ‘jump.’

While the most amusing examples are clearly ones where algorithms abused bugs in their simulations (essentially glitches in the Matrix that gave them superpowers), the paper outlines some surprising solutions that could have practical benefits as well. One algorithm invented a spinning-type movement for robots which would minimize negative effect of inconsistent hardware between bots, for instance.

As the paper notes in its discussion—and you may already be thinking—these amusing stories also reflect the potential for evolutionary algorithms or neural networks to stumble upon solutions to problems that are outside-the-box in dangerous ways. They’re a funnier version of the classic AI nightmare where computers tasked with creating peace on Earth decide the most efficient solution is to exterminate the human race.

The solution, the paper suggests, is not fear but careful experimentation. As humans gain more experience in training these sorts of algorithms, and tweaking along the way, experts gain a better sense of intuition. Still, as these anecdotes prove, it’s basically impossible to avoid unexpected results. The key is to be prepared—and to not hand over the nuclear arsenal to a robot for its very first test.

Source: The Hilarious (and Terrifying?) Ways Algorithms Have Outsmarted Their Creators

A pair of computer scientists have created a neural network that can self-replicate.

“Self-replication is a key aspect of biological life that has been largely overlooked in Artificial Intelligence systems,” they argue in a paper popped onto arXiv this month.

It’s an important process in reproduction for living things, and is an important step for evolution through natural selection. Oscar Chang, first author of the paper and a PhD student at Columbia University, explained to The Register that the goal was to see if AI could be made to be continually self improving by mimicking the biological self-replication process.

“The primary motivation here is that AI agents are powered by deep learning, and a self-replication mechanism allows for Darwinian natural selection to occur, so a population of AI agents can improve themselves simply through natural selection – just like in nature – if there was a self-replication mechanism for neural networks.”

The researchers compare their work to quines, a type of computer program that learns to produces copies of its source code. In neural networks, however, instead of the source code it’s the weights – which determine the connections between the different neurons – that are being cloned.

The researchers set up a “vanilla quine” network, a feed-forward system that produces its own weights as outputs. The vanilla quine network can also be used to self-replicate its weights and solve a task. They decided to use it for image classification on the MNIST dataset, where computers have to identify the correct digit from a set of handwritten numbers from zero to nine.

[…]

The test network required 60,000 MNIST images for training, another 10,000 for testing. And after 30 runs, the quine network had an accuracy rate of 90.41 per cent. It’s not a bad start, but its performance doesn’t really compare to larger, more sophisticated image recognition models out there.

The paper states that the “self-replication occupies a significant portion of the neural network’s capacity.” In other words, the neural network cannot focus on the image recognition task if it also has to self-replicate.

“This is an interesting finding: it is more difficult for a network that has increased its specialization at a particular task to self-replicate. This suggests that the two objectives are at odds with each other,” the paper said.

Chang explained he wasn’t sure why this happened, but it’s what happens in nature too.

Source: AI software that can reproduce like a living thing? Yup, boffins have only gone and done it • The Register

Most recent advances in artificial-intelligence systems such as speech- or face-recognition programs have come courtesy of neural networks, densely interconnected meshes of simple information processors that learn to perform tasks by analyzing huge sets of training data.

But neural nets are large, and their computations are energy intensive, so they’re not very practical for handheld devices. Most smartphone apps that rely on neural nets simply upload data to internet servers, which process it and send the results back to the phone.

Now, MIT researchers have developed a special-purpose chip that increases the speed of neural-network computations by three to seven times over its predecessors, while reducing power consumption 94 to 95 percent. That could make it practical to run neural networks locally on smartphones or even to embed them in household appliances.

“The general processor model is that there is a memory in some part of the chip, and there is a processor in another part of the chip, and you move the data back and forth between them when you do these computations,” says Avishek Biswas, an MIT graduate student in electrical engineering and computer science, who led the new chip’s development.

“Since these machine-learning algorithms need so many computations, this transferring back and forth of data is the dominant portion of the energy consumption. But the computation these algorithms do can be simplified to one specific operation, called the dot product. Our approach was, can we implement this dot-product functionality inside the memory so that you don’t need to transfer this data back and forth?”

Source: Neural networks everywhere | MIT News

This page shows how they used Deep Dream, Neural Style Transfer and OpticalFlow tools to create this music video

Source: Music Video and Artificial Intelligence – Jean Pierre by Hardcore Anal Hydrogen