Tag Archives: humans

#434616 What Games Are Humans Still Better at ...

Artificial intelligence (AI) systems’ rapid advances are continually crossing rows off the list of things humans do better than our computer compatriots.

AI has bested us at board games like chess and Go, and set astronomically high scores in classic computer games like Ms. Pacman. More complex games form part of AI’s next frontier.

While a team of AI bots developed by OpenAI, known as the OpenAI Five, ultimately lost to a team of professional players last year, they have since been running rampant against human opponents in Dota 2. Not to be outdone, Google’s DeepMind AI recently took on—and beat—several professional players at StarCraft II.

These victories beg the questions: what games are humans still better at than AI? And for how long?

The Making Of AlphaStar
DeepMind’s results provide a good starting point in a search for answers. The version of its AI for StarCraft II, dubbed AlphaStar, learned to play the games through supervised learning and reinforcement learning.

First, AI agents were trained by analyzing and copying human players, learning basic strategies. The initial agents then played each other in a sort of virtual death match where the strongest agents stayed on. New iterations of the agents were developed and entered the competition. Over time, the agents became better and better at the game, learning new strategies and tactics along the way.

One of the advantages of AI is that it can go through this kind of process at superspeed and quickly develop better agents. DeepMind researchers estimate that the AlphaStar agents went through the equivalent of roughly 200 years of game time in about 14 days.

Cheating or One Hand Behind the Back?
The AlphaStar AI agents faced off against human professional players in a series of games streamed on YouTube and Twitch. The AIs trounced their human opponents, winning ten games on the trot, before pro player Grzegorz “MaNa” Komincz managed to salvage some pride for humanity by winning the final game. Experts commenting on AlphaStar’s performance used words like “phenomenal” and “superhuman”—which was, to a degree, where things got a bit problematic.

AlphaStar proved particularly skilled at controlling and directing units in battle, known as micromanagement. One reason was that it viewed the whole game map at once—something a human player is not able to do—which made it seemingly able to control units in different areas at the same time. DeepMind researchers said the AIs only focused on a single part of the map at any given time, but interestingly, AlphaStar’s AI agent was limited to a more restricted camera view during the match “MaNA” won.

Potentially offsetting some of this advantage was the fact that AlphaStar was also restricted in certain ways. For example, it was prevented from performing more clicks per minute than a human player would be able to.

Where AIs Struggle
Games like StarCraft II and Dota 2 throw a lot of challenges at AIs. Complex game theory/ strategies, operating with imperfect/incomplete information, undertaking multi-variable and long-term planning, real-time decision-making, navigating a large action space, and making a multitude of possible decisions at every point in time are just the tip of the iceberg. The AIs’ performance in both games was impressive, but also highlighted some of the areas where they could be said to struggle.

In Dota 2 and StarCraft II, AI bots have seemed more vulnerable in longer games, or when confronted with surprising, unfamiliar strategies. They seem to struggle with complexity over time and improvisation/adapting to quick changes. This could be tied to how AIs learn. Even within the first few hours of performing a task, humans tend to gain a sense of familiarity and skill that takes an AI much longer. We are also better at transferring skill from one area to another. In other words, experience playing Dota 2 can help us become good at StarCraft II relatively quickly. This is not the case for AI—yet.

Dwindling Superiority
While the battle between AI and humans for absolute superiority is still on in Dota 2 and StarCraft II, it looks likely that AI will soon reign supreme. Similar things are happening to other types of games.

In 2017, a team from Carnegie Mellon University pitted its Libratus AI against four professionals. After 20 days of No Limit Texas Hold’em, Libratus was up by $1.7 million. Another likely candidate is the destroyer of family harmony at Christmas: Monopoly.

Poker involves bluffing, while Monopoly involves negotiation—skills you might not think AI would be particularly suited to handle. However, an AI experiment at Facebook showed that AI bots are more than capable of undertaking such tasks. The bots proved skilled negotiators, and developed negotiating strategies like pretending interest in one object while they were interested in another altogether—bluffing.

So, what games are we still better at than AI? There is no precise answer, but the list is getting shorter at a rapid pace.

The Aim Of the Game
While AI’s mastery of games might at first glance seem an odd area to focus research on, the belief is that the way AI learn to master a game is transferrable to other areas.

For example, the Libratus poker-playing AI employed strategies that could work in financial trading or political negotiations. The same applies to AlphaStar. As Oriol Vinyals, co-leader of the AlphaStar project, told The Verge:

“First and foremost, the mission at DeepMind is to build an artificial general intelligence. […] To do so, it’s important to benchmark how our agents perform on a wide variety of tasks.”

A 2017 survey of more than 350 AI researchers predicts AI could be a better driver than humans within ten years. By the middle of the century, AI will be able to write a best-selling novel, and a few years later, it will be better than humans at surgery. By the year 2060, AI may do everything better than us.

Whether you think this is a good or a bad thing, it’s worth noting that AI has an often overlooked ability to help us see things differently. When DeepMind’s AlphaGo beat human Go champion Lee Sedol, the Go community learned from it, too. Lee himself went on a win streak after the match with AlphaGo. The same is now happening within the Dota 2 and StarCraft II communities that are studying the human vs. AI games intensely.

More than anything, AI’s recent gaming triumphs illustrate how quickly artificial intelligence is developing. In 1997, Dr. Piet Hut, an astrophysicist at the Institute for Advanced Study at Princeton and a GO enthusiast, told the New York Times that:

”It may be a hundred years before a computer beats humans at Go—maybe even longer.”

Image Credit: Roman Kosolapov / Shutterstock.com Continue reading

Posted in Human Robots

#434559 Can AI Tell the Difference Between a ...

Scarcely a day goes by without another headline about neural networks: some new task that deep learning algorithms can excel at, approaching or even surpassing human competence. As the application of this approach to computer vision has continued to improve, with algorithms capable of specialized recognition tasks like those found in medicine, the software is getting closer to widespread commercial use—for example, in self-driving cars. Our ability to recognize patterns is a huge part of human intelligence: if this can be done faster by machines, the consequences will be profound.

Yet, as ever with algorithms, there are deep concerns about their reliability, especially when we don’t know precisely how they work. State-of-the-art neural networks will confidently—and incorrectly—classify images that look like television static or abstract art as real-world objects like school-buses or armadillos. Specific algorithms could be targeted by “adversarial examples,” where adding an imperceptible amount of noise to an image can cause an algorithm to completely mistake one object for another. Machine learning experts enjoy constructing these images to trick advanced software, but if a self-driving car could be fooled by a few stickers, it might not be so fun for the passengers.

These difficulties are hard to smooth out in large part because we don’t have a great intuition for how these neural networks “see” and “recognize” objects. The main insight analyzing a trained network itself can give us is a series of statistical weights, associating certain groups of points with certain objects: this can be very difficult to interpret.

Now, new research from UCLA, published in the journal PLOS Computational Biology, is testing neural networks to understand the limits of their vision and the differences between computer vision and human vision. Nicholas Baker, Hongjing Lu, and Philip J. Kellman of UCLA, alongside Gennady Erlikhman of the University of Nevada, tested a deep convolutional neural network called VGG-19. This is state-of-the-art technology that is already outperforming humans on standardized tests like the ImageNet Large Scale Visual Recognition Challenge.

They found that, while humans tend to classify objects based on their overall (global) shape, deep neural networks are far more sensitive to the textures of objects, including local color gradients and the distribution of points on the object. This result helps explain why neural networks in image recognition make mistakes that no human ever would—and could allow for better designs in the future.

In the first experiment, a neural network was trained to sort images into 1 of 1,000 different categories. It was then presented with silhouettes of these images: all of the local information was lost, while only the outline of the object remained. Ordinarily, the trained neural net was capable of recognizing these objects, assigning more than 90% probability to the correct classification. Studying silhouettes, this dropped to 10%. While human observers could nearly always produce correct shape labels, the neural networks appeared almost insensitive to the overall shape of the images. On average, the correct object was ranked as the 209th most likely solution by the neural network, even though the overall shapes were an exact match.

A particularly striking example arose when they tried to get the neural networks to classify glass figurines of objects they could already recognize. While you or I might find it easy to identify a glass model of an otter or a polar bear, the neural network classified them as “oxygen mask” and “can opener” respectively. By presenting glass figurines, where the texture information that neural networks relied on for classifying objects is lost, the neural network was unable to recognize the objects by shape alone. The neural network was similarly hopeless at classifying objects based on drawings of their outline.

If you got one of these right, you’re better than state-of-the-art image recognition software. Image Credit: Nicholas Baker, Hongjing Lu, Gennady Erlikhman, Philip J. Kelman. “Deep convolutional networks do not classify based on global object shape.” Plos Computational Biology. 12/7/18. / CC BY 4.0
When the neural network was explicitly trained to recognize object silhouettes—given no information in the training data aside from the object outlines—the researchers found that slight distortions or “ripples” to the contour of the image were again enough to fool the AI, while humans paid them no mind.

The fact that neural networks seem to be insensitive to the overall shape of an object—relying instead on statistical similarities between local distributions of points—suggests a further experiment. What if you scrambled the images so that the overall shape was lost but local features were preserved? It turns out that the neural networks are far better and faster at recognizing scrambled versions of objects than outlines, even when humans struggle. Students could classify only 37% of the scrambled objects, while the neural network succeeded 83% of the time.

Humans vastly outperform machines at classifying object (a) as a bear, while the machine learning algorithm has few problems classifying the bear in figure (b). Image Credit: Nicholas Baker, Hongjing Lu, Gennady Erlikhman, Philip J. Kelman. “Deep convolutional networks do not classify based on global object shape.” Plos Computational Biology. 12/7/18. / CC BY 4.0
“This study shows these systems get the right answer in the images they were trained on without considering shape,” Kellman said. “For humans, overall shape is primary for object recognition, and identifying images by overall shape doesn’t seem to be in these deep learning systems at all.”

Naively, one might expect that—as the many layers of a neural network are modeled on connections between neurons in the brain and resemble the visual cortex specifically—the way computer vision operates must necessarily be similar to human vision. But this kind of research shows that, while the fundamental architecture might resemble that of the human brain, the resulting “mind” operates very differently.

Researchers can, increasingly, observe how the “neurons” in neural networks light up when exposed to stimuli and compare it to how biological systems respond to the same stimuli. Perhaps someday it might be possible to use these comparisons to understand how neural networks are “thinking” and how those responses differ from humans.

But, as yet, it takes a more experimental psychology to probe how neural networks and artificial intelligence algorithms perceive the world. The tests employed against the neural network are closer to how scientists might try to understand the senses of an animal or the developing brain of a young child rather than a piece of software.

By combining this experimental psychology with new neural network designs or error-correction techniques, it may be possible to make them even more reliable. Yet this research illustrates just how much we still don’t understand about the algorithms we’re creating and using: how they tick, how they make decisions, and how they’re different from us. As they play an ever-greater role in society, understanding the psychology of neural networks will be crucial if we want to use them wisely and effectively—and not end up missing the woods for the trees.

Image Credit: Irvan Pratama / Shutterstock.com Continue reading

Posted in Human Robots

#434544 This Week’s Awesome Stories From ...

ARTIFICIAL INTELLIGENCE
DeepMind Beats Pros at Starcraft in Another Triumph for Bots
Tom Simonite | Wired
“DeepMind’s feat is the most complex yet in a long train of contests in which computers have beaten top humans at games. Checkers fell in 1994, chess in 1997, and DeepMind’s earlier bot AlphaGo became the first to beat a champion at the board game Go in 2016. The StarCraft bot is the most powerful AI game player yet; it may also be the least unexpected.”

GENETICS
Complete Axolotl Genome Could Pave the Way Toward Human Tissue Regeneration
George Dvorsky | Gizmodo
“Now that researchers have a near-complete axolotl genome—the new assembly still requires a bit of fine-tuning (more on that in a bit)—they, along with others, can now go about the work of identifying the genes responsible for axolotl tissue regeneration.”

FUTURE
We Analyzed 16,625 Papers to Figure Out Where AI Is Headed Next
Karen Hao | MIT Technology Review
“…though deep learning has singlehandedly thrust AI into the public eye, it represents just a small blip in the history of humanity’s quest to replicate our own intelligence. It’s been at the forefront of that effort for less than 10 years. When you zoom out on the whole history of the field, it’s easy to realize that it could soon be on its way out.”

COMPUTING
Apple’s Finger-Controller Patent Is a Glimpse at Mixed Reality’s Future
Mark Sullivan | Fast Company
“[Apple’s] engineers are now looking past the phone touchscreen toward mixed reality, where the company’s next great UX will very likely be built. A recent patent application gives some tantalizing clues as to how Apple’s people are thinking about aspects of that challenge.”

GOVERNANCE
How Do You Govern Machines That Can Learn? Policymakers Are Trying to Figure That Out
Steve Lohr | The New York Times
“Regulation is coming. That’s a good thing. Rules of competition and behavior are the foundation of healthy, growing markets. That was the consensus of the policymakers at MIT. But they also agreed that artificial intelligence raises some fresh policy challenges.”

Image Credit: Victoria Shapiro / Shutterstock.com Continue reading

Posted in Human Robots

#434492 Black Mirror’s ‘Bandersnatch’ ...

When was the last time you watched a movie where you could control the plot?

Bandersnatch is the first interactive film in the sci fi anthology series Black Mirror. Written by series creator Charlie Brooker and directed by David Slade, the film tells the story of young programmer Stefan Butler, who is adapting a fantasy choose-your-own-adventure novel called Bandersnatch into a video game. Throughout the film, viewers are given the power to influence Butler’s decisions, leading to diverging plots with different endings.

Like many Black Mirror episodes, this film is mind-bending, dark, and thought-provoking. In addition to innovating cinema as we know it, it is a fascinating rumination on free will, parallel realities, and emerging technologies.

Pick Your Own Adventure
With a non-linear script, Bandersnatch is a viewing experience like no other. Throughout the film viewers are given the option of making a decision for the protagonist. In these instances, they have 10 seconds to make a decision until a default decision is made. For example, in the early stage of the plot, Butler is given the choice of accepting or rejecting Tuckersoft’s offer to develop a video game and the viewer gets to decide what he does. The decision then shapes the plot accordingly.

The video game Butler is developing involves moving through a graphical maze of corridors while avoiding a creature called the Pax, and at times making choices through an on-screen instruction (sound familiar?). In other words, it’s a pick-your-own-adventure video game in a pick-your-own-adventure movie.

Many viewers have ended up spending hours exploring all the different branches of the narrative (though the average viewing is 90 minutes). One user on reddit has mapped out an entire flowchart, showing how all the different decisions (and pseudo-decisions) lead to various endings.

However, over time, Butler starts to question his own free will. It’s almost as if he’s beginning to realize that the audience is controlling him. In one branch of the narrative, he is confronted by this reality when the audience indicates to him that he is being controlled in a Netflix show: “I am watching you on Netflix. I make all the decisions for you”. Butler, as you can imagine, is horrified by this message.

But Butler isn’t the only one who has an illusion of choice. We, the seemingly powerful viewers, also appear to operate under the illusion of choice. Despite there being five main endings to the film, they are all more or less the same.

The Science Behind Bandersnatch
The premise of Bandersnatch isn’t based on fantasy, but hard science. Free will has always been a widely-debated issue in neuroscience, with reputable scientists and studies demonstrating that the whole concept may be an illusion.

In the 1970s, a psychologist named Benjamin Libet conducted a series of experiments that studied voluntary decision making in humans. He found that brain activity imitating an action, such as moving your wrist, preceded the conscious awareness of the action.

Psychologist Malcom Gladwell theorizes that while we like to believe we spend a lot of time thinking about our decisions, our mental processes actually work rapidly, automatically, and often subconsciously, from relatively little information. In addition to this, thinking and making decisions are usually a byproduct of several different brain systems, such as the hippocampus, amygdala, and prefrontal cortex working together. You are more conscious of some information processes in the brain than others.

As neuroscientist and philosopher Sam Harris points out in his book Free Will, “You did not pick your parents or the time and place of your birth. You didn’t choose your gender or most of your life experiences. You had no control whatsoever over your genome or the development of your brain. And now your brain is making choices on the basis of preferences and beliefs that have been hammered into it over a lifetime.” Like Butler, we may believe we are operating under full agency of our abilities, but we are at the mercy of many internal and external factors that influence our decisions.

Beyond free will, Bandersnatch also taps into the theory of parallel universes, a facet of the astronomical theory of the multiverse. In astrophysics, there is a theory that there are parallel universes other than our own, where all the choices you made are played out in alternate realities. For instance, if today you had the option of having cereal or eggs for breakfast, and you chose eggs, in a parallel universe, you chose cereal. Human history and our lives may have taken different paths in these parallel universes.

The Future of Cinema
In the future, the viewing experience will no longer be a passive one. Bandersnatch is just a glimpse into how technology is revolutionizing film as we know it and making it a more interactive and personalized experience. All the different scenarios and branches of the plot were scripted and filmed, but in the future, they may be adapted real-time via artificial intelligence.

Virtual reality may allow us to play an even more active role by making us participants or characters in the film. Data from your history of preferences and may be used to create a unique version of the plot that is optimized for your viewing experience.

Let’s also not underestimate the social purpose of advancing film and entertainment. Science fiction gives us the ability to create simulations of the future. Different narratives can allow us to explore how powerful technologies combined with human behavior can result in positive or negative scenarios. Perhaps in the future, science fiction will explore implications of technologies and observe human decision making in novel contexts, via AI-powered films in the virtual world.

Image Credit: andrey_l / Shutterstock.com

We are a participant in the Amazon Services LLC Associates Program, an affiliate advertising program designed to provide a means for us to earn fees by linking to Amazon.com and affiliated sites. Continue reading

Posted in Human Robots

#434336 These Smart Seafaring Robots Have a ...

Drones. Self-driving cars. Flying robo taxis. If the headlines of the last few years are to be believed, terrestrial transportation in the future will someday be filled with robotic conveyances and contraptions that will require little input from a human other than to download an app.

But what about the other 70 percent of the planet’s surface—the part that’s made up of water?

Sure, there are underwater drones that can capture 4K video for the next BBC documentary. Remotely operated vehicles (ROVs) are capable of diving down thousands of meters to investigate ocean vents or repair industrial infrastructure.

Yet most of the robots on or below the water today still lean heavily on the human element to operate. That’s not surprising given the unstructured environment of the seas and the poor communication capabilities for anything moving below the waves. Autonomous underwater vehicles (AUVs) are probably the closest thing today to smart cars in the ocean, but they generally follow pre-programmed instructions.

A new generation of seafaring robots—leveraging artificial intelligence, machine vision, and advanced sensors, among other technologies—are beginning to plunge into the ocean depths. Here are some of the latest and most exciting ones.

The Transformer of the Sea
Nic Radford, chief technology officer of Houston Mechatronics Inc. (HMI), is hesitant about throwing around the word “autonomy” when talking about his startup’s star creation, Aquanaut. He prefers the term “shared control.”

Whatever you want to call it, Aquanaut seems like something out of the script of a Transformers movie. The underwater robot begins each mission in a submarine-like shape, capable of autonomously traveling up to 200 kilometers on battery power, depending on the assignment.

When Aquanaut reaches its destination—oil and gas is the primary industry HMI hopes to disrupt to start—its four specially-designed and built linear actuators go to work. Aquanaut then unfolds into a robot with a head, upper torso, and two manipulator arms, all while maintaining proper buoyancy to get its job done.

The lightbulb moment of how to engineer this transformation from submarine to robot came one day while Aquanaut’s engineers were watching the office’s stand-up desks bob up and down. The answer to the engineering challenge of the hull suddenly seemed obvious.

“We’re just gonna build a big, gigantic, underwater stand-up desk,” Radford told Singularity Hub.

Hardware wasn’t the only problem the team, comprised of veteran NASA roboticists like Radford, had to solve. In order to ditch the expensive support vessels and large teams of humans required to operate traditional ROVs, Aquanaut would have to be able to sense its environment in great detail and relay that information back to headquarters using an underwater acoustics communications system that harkens back to the days of dial-up internet connections.

To tackle that problem of low bandwidth, HMI equipped Aquanaut with a machine vision system comprised of acoustic, optical, and laser-based sensors. All of that dense data is compressed using in-house designed technology and transmitted to a single human operator who controls Aquanaut with a few clicks of a mouse. In other words, no joystick required.

“I don’t know of anyone trying to do this level of autonomy as it relates to interacting with the environment,” Radford said.

HMI got $20 million earlier this year in Series B funding co-led by Transocean, one of the world’s largest offshore drilling contractors. That should be enough money to finish the Aquanaut prototype, which Radford said is about 99.8 percent complete. Some “high-profile” demonstrations are planned for early next year, with commercial deployments as early as 2020.

“What just gives us an incredible advantage here is that we have been born and bred on doing robotic systems for remote locations,” Radford noted. “This is my life, and I’ve bet the farm on it, and it takes this kind of fortitude and passion to see these things through, because these are not easy problems to solve.”

On Cruise Control
Meanwhile, a Boston-based startup is trying to solve the problem of making ships at sea autonomous. Sea Machines is backed by about $12.5 million in capital venture funding, with Toyota AI joining the list of investors in a $10 million Series A earlier this month.

Sea Machines is looking to the self-driving industry for inspiration, developing what it calls “vessel intelligence” systems that can be retrofitted on existing commercial vessels or installed on newly-built working ships.

For instance, the startup announced a deal earlier this year with Maersk, the world’s largest container shipping company, to deploy a system of artificial intelligence, computer vision, and LiDAR on the Danish company’s new ice-class container ship. The technology works similar to advanced driver-assistance systems found in automobiles to avoid hazards. The proof of concept will lay the foundation for a future autonomous collision avoidance system.

It’s not just startups making a splash in autonomous shipping. Radford noted that Rolls Royce—yes, that Rolls Royce—is leading the way in the development of autonomous ships. Its Intelligence Awareness system pulls in nearly every type of hyped technology on the market today: neural networks, augmented reality, virtual reality, and LiDAR.

In augmented reality mode, for example, a live feed video from the ship’s sensors can detect both static and moving objects, overlaying the scene with details about the types of vessels in the area, as well as their distance, heading, and other pertinent data.

While safety is a primary motivation for vessel automation—more than 1,100 ships have been lost over the past decade—these new technologies could make ships more efficient and less expensive to operate, according to a story in Wired about the Rolls Royce Intelligence Awareness system.

Sea Hunt Meets Science
As Singularity Hub noted in a previous article, ocean robots can also play a critical role in saving the seas from environmental threats. One poster child that has emerged—or, invaded—is the spindly lionfish.

A venomous critter endemic to the Indo-Pacific region, the lionfish is now found up and down the east coast of North America and beyond. And it is voracious, eating up to 30 times its own stomach volume and reducing juvenile reef fish populations by nearly 90 percent in as little as five weeks, according to the Ocean Support Foundation.

That has made the colorful but deadly fish Public Enemy No. 1 for many marine conservationists. Both researchers and startups are developing autonomous robots to hunt down the invasive predator.

At the Worcester Polytechnic Institute, for example, students are building a spear-carrying robot that uses machine learning and computer vision to distinguish lionfish from other aquatic species. The students trained the algorithms on thousands of different images of lionfish. The result: a lionfish-killing machine that boasts an accuracy of greater than 95 percent.

Meanwhile, a small startup called the American Marine Research Corporation out of Pensacola, Florida is applying similar technology to seek and destroy lionfish. Rather than spearfishing, the AMRC drone would stun and capture the lionfish, turning a profit by selling the creatures to local seafood restaurants.

Lionfish: It’s what’s for dinner.

Water Bots
A new wave of smart, independent robots are diving, swimming, and cruising across the ocean and its deepest depths. These autonomous systems aren’t necessarily designed to replace humans, but to venture where we can’t go or to improve safety at sea. And, perhaps, these latest innovations may inspire the robots that will someday plumb the depths of watery planets far from Earth.

Image Credit: Houston Mechatronics, Inc. Continue reading

Posted in Human Robots