Dimensional misspecification was concurrently identified as a key problem in GANs by researchers at Twitter and researchers at New York University. We believe that GAN models will be widely used beyond perceptual domains such as image generation. The technique’s result shows that by making training data more variable through adding noise, one really encodes a preference that the learned machine learning system should depend in a smooth way on its input. All rights reserved. As a result, GANs summarize the distribution via a low-dimensional manifold but do not accurately capture the full training distribution. We make this problem visible in the following figure, where we visualize the training process of a normal GAN model. Originally posted on my personal blog here. Go grandmaster Lee Sedol recently announced he was retiring from the game because "there is an entity that can never be defeated": AI. Ultimately that will mean developing machines that have emotions and consciousness. Copyright © 2019 Salon.com, LLC. , Can machines be programmed to find solutions on their own, and perhaps even come up with creative solutions that humans would find difficult? For example, even a low-resolution image has tens of thousands of pixel observations and contains structure at several scales, from correlated neighboring pixels, to edges, to objects, to scene-level statistics. It didn't get to the core of the problem. Teams of machines show bonding behaviour, working together to prevent other groups from reaching a particular goal first. Google engineers are allowed to spend up to 20 percent of their time on some other Google-related project. Standing in his living room he suddenly found himself surrounded by beautiful ideas, all of them crystallising to a point. Their fix is to add random perturbations to any instance before handing it to the adversary. What about the characteristics of creativity? Mordvintsev continued to ponder how these networks functioned. In 2015 a Google engineer called Alexander Mordvintsev decided to take a crack at solving the puzzle. Rosalyn Picard, professor of media arts and sciences at MIT, works on Affective Computing, looking into how one might develop a machine with emotions. He thought he heard a noise and checked the door to the terrace of his flat. Deep learning models successfully capture such structure in their input; for example, in 2012, researchers from Toronto produced the famous AlexNet network for image recognition, and later Microsoft ResNet could reliably understand images and classify them at human-level accuracy. Be human, not machines. Stabilizing GANs Creatives. Toward this end, Microsoft Research is committed to the research to make GAN models more practical and applicable to more areas of study. Recent methods in artificial intelligence enable AI software to produce rich and creative digital artifacts such as text and images painted from scratch. Problems in training GANs Many people argue that machines cannot be creative because they aren't "out there" in the world, having emotional experiences, communing with nature, or falling in love. Machines are redefining what it is to be living, not merely … Can Machines Be Creative? Partner Researcher. Similarly, it was not the team behind AlphaGo but the machine itself that made the spectacular and totally unexpected move that trounced Go master Lee Sedol. Each of their complex network of parts is designed using Newtonian physics, characterised by causality and determinism. One of the behaviors considered to be uniquely human is our creativity. In a sense, it was not possible for machine learning models to be creative and create complex observations such as entire images. For instance, could we program machines to create high quality original music? Artists, writers and composers too are confronted by a series of problems, and the process of solving them is what fires their creativity. There are AIs that can improvise music, jam with jazz musicians, create surreal art and write bizarre screenplays, novels and poetry. The only way we can understand creativity is to examine our own human creativity. Why not recognise the machine's creativity in the same way? When a human being makes a leap forward and produces something that goes beyond the initial material, we call it creativity. Possessing a creative mind and imagination means that you have the ability to dream up new inventions and ideas that do not currently exist. How can something made up of wires and transistors be as creative as an Einstein, a Picasso, a Shakespeare or a Bach? How can a system — a person or a machine — produce results that go far beyond the material it has to work with? But that's like saying that Mozart's father, who taught Wolfgang how to compose music, should therefore be credited with his son's musical creations. One thing is certain: their creativity will be unlimited. A machine without physiological needs cannot get sick and that does not need to worry about passing on its genes to posterity, and therefore will have no reason to feel that complex emotion of 'well being' the way humans do. In research presented at NIPS last year, researchers from Microsoft generalized the above interpretation of GANs to a broader class of games, providing a deeper theoretical understanding of the GAN learning objective and enabling GANs to apply to other machine learning problems such as probabilistic inference. The quality of samples is clearly improved through the addition of regularization. Highly creative work often operates in a distinctly Darwinian environment. The solution to his problem bubbled up into his consciousness. These four researchers collaborate on a project called “Tractable by Design” as part of the Microsoft Research Swiss Joint Research Centre (Swiss JRC in short), with Kevin Roth’s PhD studies being supported through the Swiss JRC. In the work presented today at NIPS 2017, the problem of noise variance is overcome, as presented in a paper entitled Stabilizing Training of Generative Adversarial Networks through Regularization by Kevin Roth, Aurelien Lucchi, Sebastian Nowozin and Thomas Hofmann. Sebastian Nowozin Humans are the ones programming them, they are just slaves to the instructions given and carry out tasks accordingly. It can scan disciplines that may only touch on the area under study and detect similarities which scientists have overlooked and thus discover a new and more relevant problem to research. Machines might be able to beat humans in creativity, but when an entity does not have a consciousness, it then lacks emotions, planning, abstract reasoning, and other higher cognitive functions. Is it possible for machines to also be unpredictable? Researchers often use a huge data set of images. To assert that machines will be eternally incapable of creativity for the simple reason that they are not human is a blinkered way of looking at progress, especially in a field that goes beyond science and technology and touches on our everyday lives. Machines such as AlphaGo are unquestionably displaying clear glimmerings of creativity. These activities, often characterized as knowledge work, can be as varied as coding software, creating menus, or writing promotional … Reproduction of material from any Salon pages without written permission is strictly prohibited. While this argument, in theory, sounds plausible, computers are not “creative,” do not “learn” and cannot “predict”. Associated Press articles: Copyright © 2016 The Associated Press. If humans did the same, it is likely you would. But how exactly does it do this? Meet 9 AI 'Artists' By Mindy Weisberger 01 June 2018. Some scholars have argued that the creative moment is not at the end of a deliberate computation. So any job that requires creating something aesthetically pleasing, from clothing lines to interiors to websites, is best left to actual human designers. On the left, the network is trained without regularization, and on the right, the network is trained with regularization. Where machines could replace humans—and where they can’t (yet) ... (9 percent automation potential) or that apply expertise to decision making, planning, or creative work (18 percent). I develop novel algorithms and models for artificial intelligence and machine learning applications. Although computers have advanced dramatically in many respects over the last 50 years, they still do not possess the basic conceptual intelligence or perceptual capabilities that most humans take for granted. Artificial neural networks are loosely inspired by the way the human brain is wired, how its neurons are connected. Many people argue that machines cannot be creative because they aren't "out there" in the world, having emotional experiences, communing with nature, or falling in love. By That is, that it should be robust to small variations. Lectures from Microsoft researchers with live Q&A and on-demand viewing. But how? Allowing for flexible probabilistic models is important in order to capture rich phenomena present in complex data. The generator network transforms a set of random numbers through a neural network into an observation, such as an image. They're aware of the problem they're working on and of their own wiring. This seemed too complicated. 2. creativity is accomplished by problem solving. Poincaré's four stage cycle in action: The creation of DeepDream. and computers aren’t capable of can be a fool’s errand. The second equation shows the form of the regularization term, incorporating the variation of the discriminator function, weighted by a function depending on the particular f-divergence being used. Jobs that rely heavily on the right side of the brain—from writers to … The extraordinary thing about artificial neural networks, the most creative AI machines, of which AlphaGo is an example, is that we know they work but we don't fully understand how. Imagination and inspiration are essential elements, as is unpredictability — the ability to make an unexpected leap. How generative adversarial networks work GANs learn to generate content by playing a two-player game between a generator network and an adversary network. All this can lead hopefully to an illumination, as the solution to the problem bubbles up into consciousness. Mordvintsev, however, was dissatisfied. Since the design of machines is one of these intellectual activities, an ultraintelligent machine could design even better machines; there would then unquestionably be an ‘intelligence explosion,’ and the … In principle, the regularization technique is not new; it was proposed originally in 1995 by professor Chris Bishop, now lab director at Microsoft Research Cambridge. We start by consciously working on a problem but eventually may hit a block and take a break. As readers likely remember, an artificial intelligence known as AlphaGo defeated Lee in 2016. Then connections between apparently unconnected concepts can suddenly emerge. People have been grappling with the question of artificial creativity -- alongside the question of artificial intelligence -- for over 170 years. Two marks of true genius are 1) the ability to home in on the real problem which no one else has noticed, and 2) the ability to spot connections between concepts that at first glance have nothing in common. In the middle of the night on 18 May 2015, he awoke with a start. Machine learning models trained on images need to capture this multileveled structure to effectively reason about image content. Immediately he wrote the code for his new algorithm – DeepDream — and then explored what it could actually do. Then again, we humans are merely an amalgam of nerves, arteries, bones and cells — yet we manage to be creative. But when all these are assembled into a new entity it can lead to chaotic behaviour: unpredictability. Competition, cooperation, problem-discovery and finding connections between concepts. But was the machine truly creative? From there it will be only a short step to Artificial Superintelligence, in which machines evolve emotions different from ours, more intrinsic to their own physiology, whatever that might turn out to be. Generative adversarial networks are a recent breakthrough in machine learning. Without a problem there is no creativity. The technical contribution of the work is to derive an analytic approximation to the addition of noise and showing that this corresponds to a particular form of regularization of the variation of functions. Each generated instance is then checked by the adversary, which makes a decision as to whether the sample is “real” or “fake.” The adversary is able to distinguish real samples from fake samples because it is also provided with a reference data set of real samples. The Danger is not Machines Becoming Humans, but Humans Becoming Machines The extent to which human beings are willing to be duped by computers is already very large. SALON ® is registered in the U.S. Patent and Trademark Office as a trademark of Salon.com, LLC. Machines actually have unpredictability built in. There are three types of work that humans do really well but computers cannot (yet): 1) Unstructured problem-solving: solving for problems in which the rules do not currently exist. One must fight for recognition of one's ideas. Computers are certainly more adept at solving quandaries that benefit from their unique skillset, but humans hold the edge on tasks that machines simply can’t perform. Another key factor is awareness. The work also establishes connections to different approaches aimed at improving GAN models, such as gradient regularization in Wasserstein GANs and to a numerical stability analysis, the latter work also involving researchers from Microsoft. Over two thousand years ago, Plato, in the Meno, pondered the origins of new knowledge. Because they can process information much faster than humans can, they can experiment with new combinations of data in a fraction of the time. Will we develop machines that are yet more creative? GANs learn to generate content by playing a two-player game between a generator network and an adversary network. There is a human element, especially in dire situations, that can best be delivered by a person. The Ascent. Most AIs take the Hitchcockian approach to creative work. One technique used in creating these artifacts are generative adversarial networks (GANs). When we show an image of a cat to a machine trained on a database that contains cats, it will most likely recognise that the image is a cat. Arthur I. Miller is the author of "The Artist in the Machine: The World of AI-Powered Creativity" (MIT Press). Creativity is the production of new knowledge from already existing knowledge; and. The stumbling block is that the machine didn't know that it had made a brilliant move. In the not too distant future machines will have the ability to read a language fluently. Just as we train our brain, so we feed data into an artificial neural network, allowing it to react to what it sees and hears. Machines already have a sense of low-level awareness. he had written up a detailed report, thus completing the verification phase. The machine went beyond its training set to create images that no one had ever dreamt of before. How generative adversarial networks work When playing this game over time, both players learn and, by the end, the generator network can create realistic instances similar to the reference data. One of the oddities of collaboration is that tightly knit teams are not the most creative. Human creative achievement, because of the way it is socially embedded, will not succumb to advances in artificial intelligence. Humans are born to control the machines. The great question: What makes us creative? Sentient cognition transcends the limits of formal computation, it is not equivalent to Turing Machine, it is much more powerful than that.We humans are not formal systems, we are not Turing Machines. The work presented today is a milestone because it addresses a key open problem in a line of research on generalizing GANs with a simple and principled solution. Machines don’t have the ability to see things in the same way humans do. In this case both human and machine showed creativity. Critical Thinking. This material may not be published, broadcast, rewritten or redistributed. Today it is within our hands to invent the future in infinitely different and rich ways. Jobs that Cannot Be Automated Designer. To do this we must all enhance our own creativity and learn to live with the creativity of machines too, to lead us all into a brighter future. However, one key difficulty in training GANs is the issue of dimensional misspecification: a low-dimensional input is mapped continuously to a high-dimensional space. A machine can survey an area in physics at lightning speed and spot that it is riddled with redundancies and inconsistencies, revealing that researchers are focussing on the wrong problem. “Let an ultraintelligent machine be defined as a machine that can far surpass all the intellectual activities of any man however clever. Michael Graziano, professor of psychology and neuroscience at Princeton University, studies consciousness. One group of researchers had tried tinkering with the connections between the machine's mathematically-simulated neurons so as to generate something that resembled a cat at each layer of neurons. The answer is a definite yes! By 2 AM. Given a new problem, it was difficult to predict whether a GAN would work and typical failure modes included the collapse of the entire generator to a trivial solution of producing the same output over and over — demonstrating no creativity at all! To many, the notion of machine "creativity" is an oxymoron. Shares. People may deny that the machine showed creativity and argue that Mordvintsev programmed in the DeepDream algorithm, therefore the creativity was all his. In the perhaps not too distant future machines will have evolved emotions, consciousness and creativity that duplicate ours. The first equation represents the f-GAN training objective with an additional regularization term. He argues that, in general, machines can only do what they’ve been programmed to do—we can program a machine to create a picture, and in that sense it is creative; but we still have to tell it how to create by giving it … Scientists obviously solve problems. Predicting what A.I. Both the generator and the adversary play the same game but with different goals: the generator tries to fool the adversary, and the adversary tries to remain accurate in identifying samples from the generator. Despite the excitement around GANs, back at NIPS 2016 the situation looked bleak: GANs remained notoriously difficult to train and the reasons for these difficulties were not fully clear. He looks into how one may in future be able to program consciousness into a computer akin to the way that we evolve consciousness. While a machine can perform a given task, often more efficiently than we can, what it lacks is the artistry in the activity, that uniquely human ability to cater to the needs of the individual. Astonishingly, they have the potential to do so. But could machines achieve such breakthroughs? They will acquire more knowledge by scouring the internet than we can gain in a lifetime, and by experiencing emotions vicariously will be able to convince themselves and us that they have actually had these human experiences. Yes, robots are able to recognise and analyse existing data and matter, and at a certain level computers can produce art, music, food, or writing. Only Humans, not machines, can learn. Computers can only be tasked with making inductive predictions based on past experiences. They will have attained Artificial General Intelligence: they will be as intelligent as us. In 1908 the French mathematician, philosopher and scientist Henri Poincaré suggested a four-stage cycle: conscious thinking, unconscious thinking, illumination and verification. Ideas never emerge fully formed and perfect. They are robots — a term that came from the Czech word robota, which literally translates … Even though we're no longer consciously thinking about the problem, the passionate desire to solve it keeps it alive in the unconscious where it can be mulled over freely and uninhibited in ways not always possible with conscious thought. However, while these systems are good at understanding image content, before GANs arrived, it was not possible to produce images or generate similarly rich outputs. Instead of trying to reconstruct the image that was input at a certain depth into the machine, as everyone else had, Mordvintsev let the machine generate what it saw at that particular place in its innards. They include introspection, the need to focus and home in on our strengths, and the need to collaborate, compete, and occasionally even steal ideas. The work also extends the applicability of GAN models to larger deep learning architectures. Lady Ada Lovelace, the world’s first computer programmer raised … Certainly AlphaGo showed creativity. To do so, we need to understand precisely what we mean by creativity, in terms that we can apply to machines as much as to people. We have to check and refine and edit our solution, and deduce the consequences. The grandmaster later commented that AlphaGo had displayed "human intuition.". Gil Weinberg investigates this creative … In a sense, it was not possible for machine learning models to be creative and create complex observations such as entire images. Through unconscious thought, he had found a way to use computer vision to explore how artificial neural networks worked. For a machine well being may exist but in a much more simplified form. Similar to humans, emotions plus unpredictability can be explosive and can trigger creativity. So, the end product of creativity is an idea or an object or a piece of music that has never existed before, and the process by which it is achieved is problem-solving. Awareness: The key difference between human and machine creativity. Not yet, anyway. Explaining and fixing the difficulties of training GANs was one of the main problems discussed at the NIPS 2016 workshop on adversarial training, and solutions such as unrolling the GAN game, additional stability objectives as in CVAE-GANs and minibatch discrimination, label smoothing and other heuristics have been put forth by the leading researchers. Initially proposed by Ian Goodfellow and colleagues at the University of Montreal at NIPS 2014, the GAN approach enables the specification and training of rich probabilistic deep learning models using standard deep learning technology. But they can still acquire such knowledge vicariously. The aim is to develop a machine that will work with and empathize with us rather than compete and supersede us. Adding this regularizer immediately stabilizes the training of GAN models as shown in the figure below. Michael Wilber, a PhD candidate at the SE(3) Computer Vision Group at Cornell Tech, is doubtful. One of the soft skills that humans have that AI does not … As yet artificial neural networks can't appreciate the art, literature or music they make but these are areas that scientists are currently working on. David Gelernter. The situation is illustrated in the following figure, where a one-dimensional input is mapped onto a two-dimensional distribution. One key element of creativity is unpredictability, going beyond logic, often the result of unconscious thought. In the following figure, we show a comparison of a network generating face images using the ResNet architecture. Mordvintsev's bold idea was to use the code he'd created to ask the machine to reveal what it actually saw at the level of a certain layer of neurons inside it. These are the qualities in a human being that make it likely they'll be creative and they are also qualities that we ordinary mortals can cultivate in order to be more creative. So how are problems solved? But is it possible for machines to display these same character traits and so learn to be more creative? Mordvintsev's day job was researching how to prevent spam from infecting search results, but previously he'd worked on artificial neural networks. Many people balk at the very thought. The next step is when the boundary between 'artificial' machine intelligence and 'natural' human intelligence disappears. My current research interest is in the following areas: Probabilistic, Programming languages & software engineering, Ian Goodfellow and colleagues at the University of Montreal, NIPS 2016 workshop on adversarial training, additional stability objectives as in CVAE-GANs, minibatch discrimination, label smoothing and other heuristics, Stabilizing Training of Generative Adversarial Networks through Regularization, Microsoft Research Swiss Joint Research Centre, gradient regularization in Wasserstein GANs, Next Swiss Joint Research Centre Workshop, Swiss Joint Research Center Workshop 2019, Swiss Joint Research Center Workshop 2018, Newly discovered principle reveals how adversarial training can perform robust deep learning, A deep generative model trifecta: Three advances that work towards harnessing large-scale power, From blank canvas unfolds a scene: GAN-based model generates and modifies images based on continual linguistic instruction, A picture from a dozen words – A drawing bot for realizing everyday scenes—and even stories. A particularly advanced set of machines could replace humans at literally all jobs. The usual argument against computers being creative goes that they can only do what they are told to do. True machine creativity cannot be derived from a system that solely takes input, performs mathematical functions, and presents an output to the eager programmer that created it. How does the machine see the world? She uses the analogy of autistic children who, like computers, have to be taught to read other people's social and emotional clues to interpret the meaning of facial expressions. Yet if a machine was to compose original music as good as Beethoven or paint as well as Picasso, would you call it creative?