Tag Archives: Performance
#436176 We’re Making Progress in Explainable ...
Machine learning algorithms are starting to exceed human performance in many narrow and specific domains, such as image recognition and certain types of medical diagnoses. They’re also rapidly improving in more complex domains such as generating eerily human-like text. We increasingly rely on machine learning algorithms to make decisions on a wide range of topics, from what we collectively spend billions of hours watching to who gets the job.
But machine learning algorithms cannot explain the decisions they make.
How can we justify putting these systems in charge of decisions that affect people’s lives if we don’t understand how they’re arriving at those decisions?
This desire to get more than raw numbers from machine learning algorithms has led to a renewed focus on explainable AI: algorithms that can make a decision or take an action, and tell you the reasons behind it.
What Makes You Say That?
In some circumstances, you can see a road to explainable AI already. Take OpenAI’s GTP-2 model, or IBM’s Project Debater. Both of these generate text based on a large corpus of training data, and try to make it as relevant as possible to the prompt that’s given. If these models were also able to provide a quick run-down of the top few sources in that corpus of training data they were drawing information from, it may be easier to understand where the “argument” (or poetic essay about unicorns) was coming from.
This is similar to the approach Google is now looking at for its image classifiers. Many algorithms are more sensitive to textures and the relationship between adjacent pixels in an image, rather than recognizing objects by their outlines as humans do. This leads to strange results: some algorithms can happily identify a totally scrambled image of a polar bear, but not a polar bear silhouette.
Previous attempts to make image classifiers explainable relied on significance mapping. In this method, the algorithm would highlight the areas of the image that contributed the most statistical weight to making the decision. This is usually determined by changing groups of pixels in the image and seeing which contribute to the biggest change in the algorithm’s impression of what the image is. For example, if the algorithm is trying to recognize a stop sign, changing the background is unlikely to be as important as changing the sign.
Google’s new approach changes the way that its algorithm recognizes objects, by examining them at several different resolutions and searching for matches to different “sub-objects” within the main object. You or I might recognize an ambulance from its flashing lights, its tires, and its logo; we might zoom in on the basketball held by an NBA player to deduce their occupation, and so on. By linking the overall categorization of an image to these “concepts,” the algorithm can explain its decision: I categorized this as a cat because of its tail and whiskers.
Even in this experiment, though, the “psychology” of the algorithm in decision-making is counter-intuitive. For example, in the basketball case, the most important factor in making the decision was actually the player’s jerseys rather than the basketball.
Can You Explain What You Don’t Understand?
While it may seem trivial, the conflict here is a fundamental one in approaches to artificial intelligence. Namely, how far can you get with mere statistical associations between huge sets of data, and how much do you need to introduce abstract concepts for real intelligence to arise?
At one end of the spectrum, Good Old-Fashioned AI or GOFAI dreamed up machines that would be entirely based on symbolic logic. The machine would be hard-coded with the concept of a dog, a flower, cars, and so forth, alongside all of the symbolic “rules” which we internalize, allowing us to distinguish between dogs, flowers, and cars. (You can imagine a similar approach to a conversational AI would teach it words and strict grammatical structures from the top down, rather than “learning” languages from statistical associations between letters and words in training data, as GPT-2 broadly does.)
Such a system would be able to explain itself, because it would deal in high-level, human-understandable concepts. The equation is closer to: “ball” + “stitches” + “white” = “baseball”, rather than a set of millions of numbers linking various pathways together. There are elements of GOFAI in Google’s new approach to explaining its image recognition: the new algorithm can recognize objects based on the sub-objects they contain. To do this, it requires at least a rudimentary understanding of what those sub-objects look like, and the rules that link objects to sub-objects, such as “cats have whiskers.”
The issue, of course, is the—maybe impossible—labor-intensive task of defining all these symbolic concepts and every conceivable rule that could possibly link them together by hand. The difficulty of creating systems like this, which could handle the “combinatorial explosion” present in reality, helped to lead to the first AI winter.
Meanwhile, neural networks rely on training themselves on vast sets of data. Without the “labeling” of supervised learning, this process might bear no relation to any concepts a human could understand (and therefore be utterly inexplicable).
Somewhere between these two, hope explainable AI enthusiasts, is a happy medium that can crunch colossal amounts of data, giving us all of the benefits that recent, neural-network AI has bestowed, while showing its working in terms that humans can understand.
Image Credit: Image by Seanbatty from Pixabay Continue reading
Posted in Human Robots
#436126 Quantum Computing Gets a Boost From AI ...
Illustration: Greg Mably
Anyone of a certain age who has even a passing interest in computers will remember the remarkable breakthrough that IBM made in 1997 when its Deep Blue chess-playing computer defeated Garry Kasparov, then the world chess champion. Computer scientists passed another such milestone in March 2016, when DeepMind (a subsidiary of Alphabet, Google’s parent company) announced that its AlphaGo program had defeated world-champion player Lee Sedol in the game of Go, a board game that had vexed AI researchers for decades. Recently, DeepMind’s algorithms have also bested human players in the computer games StarCraft IIand Quake Arena III.
Some believe that the cognitive capacities of machines will overtake those of human beings in many spheres within a few decades. Others are more cautious and point out that our inability to understand the source of our own cognitive powers presents a daunting hurdle. How can we make thinking machines if we don’t fully understand our own thought processes?
Citizen science, which enlists masses of people to tackle research problems, holds promise here, in no small part because it can be used effectively to explore the boundary between human and artificial intelligence.
Some citizen-science projects ask the public to collect data from their surroundings (as eButterfly does for butterflies) or to monitor delicate ecosystems (as Eye on the Reef does for Australia’s Great Barrier Reef). Other projects rely on online platforms on which people help to categorize obscure phenomena in the night sky (Zooniverse) or add to the understanding of the structure of proteins (Foldit). Typically, people can contribute to such projects without any prior knowledge of the subject. Their fundamental cognitive skills, like the ability to quickly recognize patterns, are sufficient.
In order to design and develop video games that can allow citizen scientists to tackle scientific problems in a variety of fields, professor and group leader Jacob Sherson founded ScienceAtHome (SAH), at Aarhus University, in Denmark. The group began by considering topics in quantum physics, but today SAH hosts games covering other areas of physics, math, psychology, cognitive science, and behavioral economics. We at SAH search for innovative solutions to real research challenges while providing insight into how people think, both alone and when working in groups.
It is computationally intractable to completely map out a higher-dimensional landscape: It is called the curse of high dimensionality, and it plagues many optimization problems.
We believe that the design of new AI algorithms would benefit greatly from a better understanding of how people solve problems. This surmise has led us to establish the Center for Hybrid Intelligence within SAH, which tries to combine human and artificial intelligence, taking advantage of the particular strengths of each. The center’s focus is on the gamification of scientific research problems and the development of interfaces that allow people to understand and work together with AI.
Our first game, Quantum Moves, was inspired by our group’s research into quantum computers. Such computers can in principle solve certain problems that would take a classical computer billions of years. Quantum computers could challenge current cryptographic protocols, aid in the design of new materials, and give insight into natural processes that require an exact solution of the equations of quantum mechanics—something normal computers are inherently bad at doing.
One candidate system for building such a computer would capture individual atoms by “freezing” them, as it were, in the interference pattern produced when a laser beam is reflected back on itself. The captured atoms can thus be organized like eggs in a carton, forming a periodic crystal of atoms and light. Using these atoms to perform quantum calculations requires that we use tightly focused laser beams, called optical tweezers, to transport the atoms from site to site in the light crystal. This is a tricky business because individual atoms do not behave like particles; instead, they resemble a wavelike liquid governed by the laws of quantum mechanics.
In Quantum Moves, a player manipulates a touch screen or mouse to move a simulated laser tweezer and pick up a trapped atom, represented by a liquidlike substance in a bowl. Then the player must bring the atom back to the tweezer’s initial position while trying to minimize the sloshing of the liquid. Such sloshing would increase the energy of the atom and ultimately introduce errors into the operations of the quantum computer. Therefore, at the end of a move, the liquid should be at a complete standstill.
To understand how people and computers might approach such a task differently, you need to know something about how computerized optimization algorithms work. The countless ways of moving a glass of water without spilling may be regarded as constituting a “solution landscape.” One solution is represented by a single point in that landscape, and the height of that point represents the quality of the solution—how smoothly and quickly the glass of water was moved. This landscape might resemble a mountain range, where the top of each mountain represents a local optimum and where the challenge is to find the highest peak in the range—the global optimum.
Illustration: Greg Mably
Researchers must compromise between searching the landscape for taller mountains (“exploration”) and climbing to the top of the nearest mountain (“exploitation”). Making such a trade-off may seem easy when exploring an actual physical landscape: Merely hike around a bit to get at least the general lay of the land before surveying in greater detail what seems to be the tallest peak. But because each possible way of changing the solution defines a new dimension, a realistic problem can have thousands of dimensions. It is computationally intractable to completely map out such a higher-dimensional landscape. We call this the curse of high dimensionality, and it plagues many optimization problems.
Although algorithms are wonderfully efficient at crawling to the top of a given mountain, finding good ways of searching through the broader landscape poses quite a challenge, one that is at the forefront of AI research into such control problems. The conventional approach is to come up with clever ways of reducing the search space, either through insights generated by researchers or with machine-learning algorithms trained on large data sets.
At SAH, we attacked certain quantum-optimization problems by turning them into a game. Our goal was not to show that people can beat computers in this arena but rather to understand the process of generating insights into such problems. We addressed two core questions: whether allowing players to explore the infinite space of possibilities will help them find good solutions and whether we can learn something by studying their behavior.
Today, more than 250,000 people have played Quantum Moves, and to our surprise, they did in fact search the space of possible moves differently from the algorithm we had put to the task. Specifically, we found that although players could not solve the optimization problem on their own, they were good at searching the broad landscape. The computer algorithms could then take those rough ideas and refine them.
Herbert A. Simon said that “solving a problem simply means representing it so as to make the solution transparent.” Apparently, that’s what our games can do with their novel user interfaces.
Perhaps even more interesting was our discovery that players had two distinct ways of solving the problem, each with a clear physical interpretation. One set of players started by placing the tweezer close to the atom while keeping a barrier between the atom trap and the tweezer. In classical physics, a barrier is an impenetrable obstacle, but because the atom liquid is a quantum-mechanical object, it can tunnel through the barrier into the tweezer, after which the player simply moved the tweezer to the target area. Another set of players moved the tweezer directly into the atom trap, picked up the atom liquid, and brought it back. We called these two strategies the “tunneling” and “shoveling” strategies, respectively.
Such clear strategies are extremely valuable because they are very difficult to obtain directly from an optimization algorithm. Involving humans in the optimization loop can thus help us gain insight into the underlying physical phenomena that are at play, knowledge that may then be transferred to other types of problems.
Quantum Moves raised several obvious issues. First, because generating an exceptional solution required further computer-based optimization, players were unable to get immediate feedback to help them improve their scores, and this often left them feeling frustrated. Second, we had tested this approach on only one scientific challenge with a clear classical analogue, that of the sloshing liquid. We wanted to know whether such gamification could be applied more generally, to a variety of scientific challenges that do not offer such immediately applicable visual analogies.
We address these two concerns in Quantum Moves 2. Here, the player first generates a number of candidate solutions by playing the original game. Then the player chooses which solutions to optimize using a built-in algorithm. As the algorithm improves a player’s solution, it modifies the solution path—the movement of the tweezer—to represent the optimized solution. Guided by this feedback, players can then improve their strategy, come up with a new solution, and iteratively feed it back into this process. This gameplay provides high-level heuristics and adds human intuition to the algorithm. The person and the machine work in tandem—a step toward true hybrid intelligence.
In parallel with the development of Quantum Moves 2, we also studied how people collaboratively solve complex problems. To that end, we opened our atomic physics laboratory to the general public—virtually. We let people from around the world dictate the experiments we would run to see if they would find ways to improve the results we were getting. What results? That’s a little tricky to explain, so we need to pause for a moment and provide a little background on the relevant physics.
One of the essential steps in building the quantum computer along the lines described above is to create the coldest state of matter in the universe, known as a Bose-Einstein condensate. Here millions of atoms oscillate in synchrony to form a wavelike substance, one of the largest purely quantum phenomena known. To create this ultracool state of matter, researchers typically use a combination of laser light and magnetic fields. There is no familiar physical analogy between such a strange state of matter and the phenomena of everyday life.
The result we were seeking in our lab was to create as much of this enigmatic substance as was possible given the equipment available. The sequence of steps to accomplish that was unknown. We hoped that gamification could help to solve this problem, even though it had no classical analogy to present to game players.
Images: ScienceAtHome
Fun and Games: The
Quantum Moves game evolved over time, from a relatively crude early version [top] to its current form [second from top] and then a major revision,
Quantum Moves 2 [third from top].
Skill Lab: Science Detective games [bottom] test players’ cognitive skills.
In October 2016, we released a game that, for two weeks, guided how we created Bose-Einstein condensates in our laboratory. By manipulating simple curves in the game interface, players generated experimental sequences for us to use in producing these condensates—and they did so without needing to know anything about the underlying physics. A player would generate such a solution, and a few minutes later we would run the sequence in our laboratory. The number of ultracold atoms in the resulting Bose-Einstein condensate was measured and fed back to the player as a score. Players could then decide either to try to improve their previous solution or to copy and modify other players’ solutions. About 600 people from all over the world participated, submitting 7,577 solutions in total. Many of them yielded bigger condensates than we had previously produced in the lab.
So this exercise succeeded in achieving our primary goal, but it also allowed us to learn something about human behavior. We learned, for example, that players behave differently based on where they sit on the leaderboard. High-performing players make small changes to their successful solutions (exploitation), while poorly performing players are willing to make more dramatic changes (exploration). As a collective, the players nicely balance exploration and exploitation. How they do so provides valuable inspiration to researchers trying to understand human problem solving in social science as well as to those designing new AI algorithms.
How could mere amateurs outperform experienced experimental physicists? The players certainly weren’t better at physics than the experts—but they could do better because of the way in which the problem was posed. By turning the research challenge into a game, we gave players the chance to explore solutions that had previously required complex programming to study. Indeed, even expert experimentalists improved their solutions dramatically by using this interface.
Insight into why that’s possible can probably be found in the words of the late economics Nobel laureate Herbert A. Simon: “Solving a problem simply means representing it so as to make the solution transparent [PDF].” Apparently, that’s what our games can do with their novel user interfaces. We believe that such interfaces might be a key to using human creativity to solve other complex research problems.
Eventually, we’d like to get a better understanding of why this kind of gamification works as well as it does. A first step would be to collect more data on what the players do while they are playing. But even with massive amounts of data, detecting the subtle patterns underlying human intuition is an overwhelming challenge. To advance, we need a deeper insight into the cognition of the individual players.
As a step forward toward this goal, ScienceAtHome created Skill Lab: Science Detective, a suite of minigames exploring visuospatial reasoning, response inhibition, reaction times, and other basic cognitive skills. Then we compare players’ performance in the games with how well these same people did on established psychological tests of those abilities. The point is to allow players to assess their own cognitive strengths and weaknesses while donating their data for further public research.
In the fall of 2018 we launched a prototype of this large-scale profiling in collaboration with the Danish Broadcasting Corp. Since then more than 20,000 people have participated, and in part because of the publicity granted by the public-service channel, participation has been very evenly distributed across ages and by gender. Such broad appeal is rare in social science, where the test population is typically drawn from a very narrow demographic, such as college students.
Never before has such a large academic experiment in human cognition been conducted. We expect to gain new insights into many things, among them how combinations of cognitive abilities sharpen or decline with age, what characteristics may be used to prescreen for mental illnesses, and how to optimize the building of teams in our work lives.
And so what started as a fun exercise in the weird world of quantum mechanics has now become an exercise in understanding the nuances of what makes us human. While we still seek to understand atoms, we can now aspire to understand people’s minds as well.
This article appears in the November 2019 print issue as “A Man-Machine Mind Meld for Quantum Computing.”
About the Authors
Ottó Elíasson, Carrie Weidner, Janet Rafner, and Shaeema Zaman Ahmed work with the ScienceAtHome project at Aarhus University in Denmark. Continue reading
Posted in Human Robots