Tag Archives: Handle
#434786 AI Performed Like a Human on a Gestalt ...
Dr. Been Kim wants to rip open the black box of deep learning.
A senior researcher at Google Brain, Kim specializes in a sort of AI psychology. Like cognitive psychologists before her, she develops various ways to probe the alien minds of artificial neural networks (ANNs), digging into their gory details to better understand the models and their responses to inputs.
The more interpretable ANNs are, the reasoning goes, the easier it is to reveal potential flaws in their reasoning. And if we understand when or why our systems choke, we’ll know when not to use them—a foundation for building responsible AI.
There are already several ways to tap into ANN reasoning, but Kim’s inspiration for unraveling the AI black box came from an entirely different field: cognitive psychology. The field aims to discover fundamental rules of how the human mind—essentially also a tantalizing black box—operates, Kim wrote with her colleagues.
In a new paper uploaded to the pre-publication server arXiv, the team described a way to essentially perform a human cognitive test on ANNs. The test probes how we automatically complete gaps in what we see, so that they form entire objects—for example, perceiving a circle from a bunch of loose dots arranged along a clock face. Psychologist dub this the “law of completion,” a highly influential idea that led to explanations of how our minds generalize data into concepts.
Because deep neural networks in machine vision loosely mimic the structure and connections of the visual cortex, the authors naturally asked: do ANNs also exhibit the law of completion? And what does that tell us about how an AI thinks?
Enter the Germans
The law of completion is part of a series of ideas from Gestalt psychology. Back in the 1920s, long before the advent of modern neuroscience, a group of German experimental psychologists asked: in this chaotic, flashy, unpredictable world, how do we piece together input in a way that leads to meaningful perceptions?
The result is a group of principles known together as the Gestalt effect: that the mind self-organizes to form a global whole. In the more famous words of Gestalt psychologist Kurt Koffka, our perception forms a whole that’s “something else than the sum of its parts.” Not greater than; just different.
Although the theory has its critics, subsequent studies in humans and animals suggest that the law of completion happens on both the cognitive and neuroanatomical level.
Take a look at the drawing below. You immediately “see” a shape that’s actually the negative: a triangle or a square (A and B). Or you further perceive a 3D ball (C), or a snake-like squiggle (D). Your mind fills in blank spots, so that the final perception is more than just the black shapes you’re explicitly given.
Image Credit: Wikimedia Commons contributors, the free media repository.
Neuroscientists now think that the effect comes from how our visual system processes information. Arranged in multiple layers and columns, lower-level neurons—those first to wrangle the data—tend to extract simpler features such as lines or angles. In Gestalt speak, they “see” the parts.
Then, layer by layer, perception becomes more abstract, until higher levels of the visual system directly interpret faces or objects—or things that don’t really exist. That is, the “whole” emerges.
The Experiment Setup
Inspired by these classical experiments, Kim and team developed a protocol to test the Gestalt effect on feed-forward ANNs: one simple, the other, dubbed the “Inception V3,” far more complex and widely used in the machine vision community.
The main idea is similar to the triangle drawings above. First, the team generated three datasets: one set shows complete, ordinary triangles. The second—the “Illusory” set, shows triangles with the edges removed but the corners intact. Thanks to the Gestalt effect, to us humans these generally still look like triangles. The third set also only shows incomplete triangle corners. But here, the corners are randomly rotated so that we can no longer imagine a line connecting them—hence, no more triangle.
To generate a dataset large enough to tease out small effects, the authors changed the background color, image rotation, and other aspects of the dataset. In all, they produced nearly 1,000 images to test their ANNs on.
“At a high level, we compare an ANN’s activation similarities between the three sets of stimuli,” the authors explained. The process is two steps: first, train the AI on complete triangles. Second, test them on the datasets. If the response is more similar between the illusory set and the complete triangle—rather than the randomly rotated set—it should suggest a sort of Gestalt closure effect in the network.
Machine Gestalt
Right off the bat, the team got their answer: yes, ANNs do seem to exhibit the law of closure.
When trained on natural images, the networks better classified the illusory set as triangles than those with randomized connection weights or networks trained on white noise.
When the team dug into the “why,” things got more interesting. The ability to complete an image correlated with the network’s ability to generalize.
Humans subconsciously do this constantly: anything with a handle made out of ceramic, regardless of shape, could easily be a mug. ANNs still struggle to grasp common features—clues that immediately tells us “hey, that’s a mug!” But when they do, it sometimes allows the networks to better generalize.
“What we observe here is that a network that is able to generalize exhibits…more of the closure effect [emphasis theirs], hinting that the closure effect reflects something beyond simply learning features,” the team wrote.
What’s more, remarkably similar to the visual cortex, “higher” levels of the ANNs showed more of the closure effect than lower layers, and—perhaps unsurprisingly—the more layers a network had, the more it exhibited the closure effect.
As the networks learned, their ability to map out objects from fragments also improved. When the team messed around with the brightness and contrast of the images, the AI still learned to see the forest from the trees.
“Our findings suggest that neural networks trained with natural images do exhibit closure,” the team concluded.
AI Psychology
That’s not to say that ANNs recapitulate the human brain. As Google’s Deep Dream, an effort to coax AIs into spilling what they’re perceiving, clearly demonstrates, machine vision sees some truly weird stuff.
In contrast, because they’re modeled after the human visual cortex, perhaps it’s not all that surprising that these networks also exhibit higher-level properties inherent to how we process information.
But to Kim and her colleagues, that’s exactly the point.
“The field of psychology has developed useful tools and insights to study human brains– tools that we may be able to borrow to analyze artificial neural networks,” they wrote.
By tweaking these tools to better analyze machine minds, the authors were able to gain insight on how similarly or differently they see the world from us. And that’s the crux: the point isn’t to say that ANNs perceive the world sort of, kind of, maybe similar to humans. It’s to tap into a wealth of cognitive psychology tools, established over decades using human minds, to probe that of ANNs.
“The work here is just one step along a much longer path,” the authors conclude.
“Understanding where humans and neural networks differ will be helpful for research on interpretability by enlightening the fundamental differences between the two interesting species.”
Image Credit: Popova Alena / Shutterstock.com Continue reading
Posted in Human Robots
#434303 Making Superhumans Through Radical ...
Imagine trying to read War and Peace one letter at a time. The thought alone feels excruciating. But in many ways, this painful idea holds parallels to how human-machine interfaces (HMI) force us to interact with and process data today.
Designed back in the 1970s at Xerox PARC and later refined during the 1980s by Apple, today’s HMI was originally conceived during fundamentally different times, and specifically, before people and machines were generating so much data. Fast forward to 2019, when humans are estimated to produce 44 zettabytes of data—equal to two stacks of books from here to Pluto—and we are still using the same HMI from the 1970s.
These dated interfaces are not equipped to handle today’s exponential rise in data, which has been ushered in by the rapid dematerialization of many physical products into computers and software.
Breakthroughs in perceptual and cognitive computing, especially machine learning algorithms, are enabling technology to process vast volumes of data, and in doing so, they are dramatically amplifying our brain’s abilities. Yet even with these powerful technologies that at times make us feel superhuman, the interfaces are still crippled with poor ergonomics.
Many interfaces are still designed around the concept that human interaction with technology is secondary, not instantaneous. This means that any time someone uses technology, they are inevitably multitasking, because they must simultaneously perform a task and operate the technology.
If our aim, however, is to create technology that truly extends and amplifies our mental abilities so that we can offload important tasks, the technology that helps us must not also overwhelm us in the process. We must reimagine interfaces to work in coherence with how our minds function in the world so that our brains and these tools can work together seamlessly.
Embodied Cognition
Most technology is designed to serve either the mind or the body. It is a problematic divide, because our brains use our entire body to process the world around us. Said differently, our minds and bodies do not operate distinctly. Our minds are embodied.
Studies using MRI scans have shown that when a person feels an emotion in their gut, blood actually moves to that area of the body. The body and the mind are linked in this way, sharing information back and forth continuously.
Current technology presents data to the brain differently from how the brain processes data. Our brains, for example, use sensory data to continually encode and decipher patterns within the neocortex. Our brains do not create a linguistic label for each item, which is how the majority of machine learning systems operate, nor do our brains have an image associated with each of these labels.
Our bodies move information through us instantaneously, in a sense “computing” at the speed of thought. What if our technology could do the same?
Using Cognitive Ergonomics to Design Better Interfaces
Well-designed physical tools, as philosopher Martin Heidegger once meditated on while using the metaphor of a hammer, seem to disappear into the “hand.” They are designed to amplify a human ability and not get in the way during the process.
The aim of physical ergonomics is to understand the mechanical movement of the human body and then adapt a physical system to amplify the human output in accordance. By understanding the movement of the body, physical ergonomics enables ergonomically sound physical affordances—or conditions—so that the mechanical movement of the body and the mechanical movement of the machine can work together harmoniously.
Cognitive ergonomics applied to HMI design uses this same idea of amplifying output, but rather than focusing on physical output, the focus is on mental output. By understanding the raw materials the brain uses to comprehend information and form an output, cognitive ergonomics allows technologists and designers to create technological affordances so that the brain can work seamlessly with interfaces and remove the interruption costs of our current devices. In doing so, the technology itself “disappears,” and a person’s interaction with technology becomes fluid and primary.
By leveraging cognitive ergonomics in HMI design, we can create a generation of interfaces that can process and present data the same way humans process real-world information, meaning through fully-sensory interfaces.
Several brain-machine interfaces are already on the path to achieving this. AlterEgo, a wearable device developed by MIT researchers, uses electrodes to detect and understand nonverbal prompts, which enables the device to read the user’s mind and act as an extension of the user’s cognition.
Another notable example is the BrainGate neural device, created by researchers at Stanford University. Just two months ago, a study was released showing that this brain implant system allowed paralyzed patients to navigate an Android tablet with their thoughts alone.
These are two extraordinary examples of what is possible for the future of HMI, but there is still a long way to go to bring cognitive ergonomics front and center in interface design.
Disruptive Innovation Happens When You Step Outside Your Existing Users
Most of today’s interfaces are designed by a narrow population, made up predominantly of white, non-disabled men who are prolific in the use of technology (you may recall The New York Times viral article from 2016, Artificial Intelligence’s White Guy Problem). If you ask this population if there is a problem with today’s HMIs, most will say no, and this is because the technology has been designed to serve them.
This lack of diversity means a limited perspective is being brought to interface design, which is problematic if we want HMI to evolve and work seamlessly with the brain. To use cognitive ergonomics in interface design, we must first gain a more holistic understanding of how people with different abilities understand the world and how they interact with technology.
Underserved groups, such as people with physical disabilities, operate on what Clayton Christensen coined in The Innovator’s Dilemma as the fringe segment of a market. Developing solutions that cater to fringe groups can in fact disrupt the larger market by opening a downward, much larger market.
Learning From Underserved Populations
When technology fails to serve a group of people, that group must adapt the technology to meet their needs.
The workarounds created are often ingenious, specifically because they have not been arrived at by preferences, but out of necessity that has forced disadvantaged users to approach the technology from a very different vantage point.
When a designer or technologist begins learning from this new viewpoint and understanding challenges through a different lens, they can bring new perspectives to design—perspectives that otherwise can go unseen.
Designers and technologists can also learn from people with physical disabilities who interact with the world by leveraging other senses that help them compensate for one they may lack. For example, some blind people use echolocation to detect objects in their environments.
The BrainPort device developed by Wicab is an incredible example of technology leveraging one human sense to serve or compliment another. The BrainPort device captures environmental information with a wearable video camera and converts this data into soft electrical stimulation sequences that are sent to a device on the user’s tongue—the most sensitive touch receptor in the body. The user learns how to interpret the patterns felt on their tongue, and in doing so, become able to “see” with their tongue.
Key to the future of HMI design is learning how different user groups navigate the world through senses beyond sight. To make cognitive ergonomics work, we must understand how to leverage the senses so we’re not always solely relying on our visual or verbal interactions.
Radical Inclusion for the Future of HMI
Bringing radical inclusion into HMI design is about gaining a broader lens on technology design at large, so that technology can serve everyone better.
Interestingly, cognitive ergonomics and radical inclusion go hand in hand. We can’t design our interfaces with cognitive ergonomics without bringing radical inclusion into the picture, and we also will not arrive at radical inclusion in technology so long as cognitive ergonomics are not considered.
This new mindset is the only way to usher in an era of technology design that amplifies the collective human ability to create a more inclusive future for all.
Image Credit: jamesteohart / Shutterstock.com Continue reading
Posted in Human Robots
#433728 AI Is Kicking Space Exploration into ...
Artificial intelligence in space exploration is gathering momentum. Over the coming years, new missions look likely to be turbo-charged by AI as we voyage to comets, moons, and planets and explore the possibilities of mining asteroids.
“AI is already a game-changer that has made scientific research and exploration much more efficient. We are not just talking about a doubling but about a multiple of ten,” Leopold Summerer, Head of the Advanced Concepts and Studies Office at ESA, said in an interview with Singularity Hub.
Examples Abound
The history of AI and space exploration is older than many probably think. It has already played a significant role in research into our planet, the solar system, and the universe. As computer systems and software have developed, so have AI’s potential use cases.
The Earth Observer 1 (EO-1) satellite is a good example. Since its launch in the early 2000s, its onboard AI systems helped optimize analysis of and response to natural occurrences, like floods and volcanic eruptions. In some cases, the AI was able to tell EO-1 to start capturing images before the ground crew were even aware that the occurrence had taken place.
Other satellite and astronomy examples abound. Sky Image Cataloging and Analysis Tool (SKICAT) has assisted with the classification of objects discovered during the second Palomar Sky Survey, classifying thousands more objects caught in low resolution than a human would be able to. Similar AI systems have helped astronomers to identify 56 new possible gravitational lenses that play a crucial role in connection with research into dark matter.
AI’s ability to trawl through vast amounts of data and find correlations will become increasingly important in relation to getting the most out of the available data. ESA’s ENVISAT produces around 400 terabytes of new data every year—but will be dwarfed by the Square Kilometre Array, which will produce around the same amount of data that is currently on the internet in a day.
AI Readying For Mars
AI is also being used for trajectory and payload optimization. Both are important preliminary steps to NASA’s next rover mission to Mars, the Mars 2020 Rover, which is, slightly ironically, set to land on the red planet in early 2021.
An AI known as AEGIS is already on the red planet onboard NASA’s current rovers. The system can handle autonomous targeting of cameras and choose what to investigate. However, the next generation of AIs will be able to control vehicles, autonomously assist with study selection, and dynamically schedule and perform scientific tasks.
Throughout his career, John Leif Jørgensen from DTU Space in Denmark has designed equipment and systems that have been on board about 100 satellites—and counting. He is part of the team behind the Mars 2020 Rover’s autonomous scientific instrument PIXL, which makes extensive use of AI. Its purpose is to investigate whether there have been lifeforms like stromatolites on Mars.
“PIXL’s microscope is situated on the rover’s arm and needs to be placed 14 millimetres from what we want it to study. That happens thanks to several cameras placed on the rover. It may sound simple, but the handover process and finding out exactly where to place the arm can be likened to identifying a building from the street from a picture taken from the roof. This is something that AI is eminently suited for,” he said in an interview with Singularity Hub.
AI also helps PIXL operate autonomously throughout the night and continuously adjust as the environment changes—the temperature changes between day and night can be more than 100 degrees Celsius, meaning that the ground beneath the rover, the cameras, the robotic arm, and the rock being studied all keep changing distance.
“AI is at the core of all of this work, and helps almost double productivity,” Jørgensen said.
First Mars, Then Moons
Mars is likely far from the final destination for AIs in space. Jupiter’s moons have long fascinated scientists. Especially Europa, which could house a subsurface ocean, buried beneath an approximately 10 km thick ice crust. It is one of the most likely candidates for finding life elsewhere in the solar system.
While that mission may be some time in the future, NASA is currently planning to launch the James Webb Space Telescope into an orbit of around 1.5 million kilometers from Earth in 2020. Part of the mission will involve AI-empowered autonomous systems overseeing the full deployment of the telescope’s 705-kilo mirror.
The distances between Earth and Europa, or Earth and the James Webb telescope, means a delay in communications. That, in turn, makes it imperative for the crafts to be able to make their own decisions. Examples from the Mars Rover project show that communication between a rover and Earth can take 20 minutes because of the vast distance. A Europa mission would see much longer communication times.
Both missions, to varying degrees, illustrate one of the most significant challenges currently facing the use of AI in space exploration. There tends to be a direct correlation between how well AI systems perform and how much data they have been fed. The more, the better, as it were. But we simply don’t have very much data to feed such a system about what it’s likely to encounter on a mission to a place like Europa.
Computing power presents a second challenge. A strenuous, time-consuming approval process and the risk of radiation mean that your computer at home would likely be more powerful than anything going into space in the near future. A 200 GHz processor, 256 megabytes of ram, and 2 gigabytes of memory sounds a lot more like a Nokia 3210 (the one you could use as an ice hockey puck without it noticing) than an iPhone X—but it’s actually the ‘brain’ that will be onboard the next rover.
Private Companies Taking Off
Private companies are helping to push those limitations. CB Insights charts 57 startups in the space-space, covering areas as diverse as natural resources, consumer tourism, R&D, satellites, spacecraft design and launch, and data analytics.
David Chew works as an engineer for the Japanese satellite company Axelspace. He explained how private companies are pushing the speed of exploration and lowering costs.
“Many private space companies are taking advantage of fall-back systems and finding ways of using parts and systems that traditional companies have thought of as non-space-grade. By implementing fall-backs, and using AI, it is possible to integrate and use parts that lower costs without adding risk of failure,” he said in an interview with Singularity Hub.
Terraforming Our Future Home
Further into the future, moonshots like terraforming Mars await. Without AI, these kinds of projects to adapt other planets to Earth-like conditions would be impossible.
Autonomous crafts are already terraforming here on Earth. BioCarbon Engineering uses drones to plant up to 100,000 trees in a single day. Drones first survey and map an area, then an algorithm decides the optimal locations for the trees before a second wave of drones carry out the actual planting.
As is often the case with exponential technologies, there is a great potential for synergies and convergence. For example with AI and robotics, or quantum computing and machine learning. Why not send an AI-driven robot to Mars and use it as a telepresence for scientists on Earth? It could be argued that we are already in the early stages of doing just that by using VR and AR systems that take data from the Mars rovers and create a virtual landscape scientists can walk around in and make decisions on what the rovers should explore next.
One of the biggest benefits of AI in space exploration may not have that much to do with its actual functions. Chew believes that within as little as ten years, we could see the first mining of asteroids in the Kuiper Belt with the help of AI.
“I think one of the things that AI does to space exploration is that it opens up a whole range of new possible industries and services that have a more immediate effect on the lives of people on Earth,” he said. “It becomes a relatable industry that has a real effect on people’s daily lives. In a way, space exploration becomes part of people’s mindset, and the border between our planet and the solar system becomes less important.”
Image Credit: Taily / Shutterstock.com Continue reading
Posted in Human Robots