Tag Archives: experiment
Last year, a novelist went on a road trip across the USA. The trip was an attempt to emulate Jack Kerouac—to go out on the road and find something essential to write about in the experience. There is, however, a key difference between this writer and anyone else talking your ear off in the bar. This writer is just a microphone, a GPS, and a camera hooked up to a laptop and a whole bunch of linear algebra.
People who are optimistic that artificial intelligence and machine learning won’t put us all out of a job say that human ingenuity and creativity will be difficult to imitate. The classic argument is that, just as machines freed us from repetitive manual tasks, machine learning will free us from repetitive intellectual tasks.
This leaves us free to spend more time on the rewarding aspects of our work, pursuing creative hobbies, spending time with loved ones, and generally being human.
In this worldview, creative works like a great novel or symphony, and the emotions they evoke, cannot be reduced to lines of code. Humans retain a dimension of superiority over algorithms.
But is creativity a fundamentally human phenomenon? Or can it be learned by machines?
And if they learn to understand us better than we understand ourselves, could the great AI novel—tailored, of course, to your own predispositions in fiction—be the best you’ll ever read?
Maybe Not a Beach Read
This is the futurist’s view, of course. The reality, as the jury-rigged contraption in Ross Goodwin’s Cadillac for that road trip can attest, is some way off.
“This is very much an imperfect document, a rapid prototyping project. The output isn’t perfect. I don’t think it’s a human novel, or anywhere near it,” Goodwin said of the novel that his machine created. 1 The Road is currently marketed as the first novel written by AI.
Once the neural network has been trained, it can generate any length of text that the author desires, either at random or working from a specific seed word or phrase. Goodwin used the sights and sounds of the road trip to provide these seeds: the novel is written one sentence at a time, based on images, locations, dialogue from the microphone, and even the computer’s own internal clock.
The results are… mixed.
The novel begins suitably enough, quoting the time: “It was nine seventeen in the morning, and the house was heavy.” Descriptions of locations begin according to the Foursquare dataset fed into the algorithm, but rapidly veer off into the weeds, becoming surreal. While experimentation in literature is a wonderful thing, repeatedly quoting longitude and latitude coordinates verbatim is unlikely to win anyone the Booker Prize.
Data In, Art Out?
Neural networks as creative agents have some advantages. They excel at being trained on large datasets, identifying the patterns in those datasets, and producing output that follows those same rules. Music inspired by or written by AI has become a growing subgenre—there’s even a pop album by human-machine collaborators called the Songularity.
A neural network can “listen to” all of Bach and Mozart in hours, and train itself on the works of Shakespeare to produce passable pseudo-Bard. The idea of artificial creativity has become so widespread that there’s even a meme format about forcibly training neural network ‘bots’ on human writing samples, with hilarious consequences—although the best joke was undoubtedly human in origin.
The AI that roamed from New York to New Orleans was an LSTM (long short-term memory) neural net. By default, information contained in individual neurons is preserved, and only small parts can be “forgotten” or “learned” in an individual timestep, rather than neurons being entirely overwritten.
The LSTM architecture performs better than previous recurrent neural networks at tasks such as handwriting and speech recognition. The neural net—and its programmer—looked further in search of literary influences, ingesting 60 million words (360 MB) of raw literature according to Goodwin’s recipe: one third poetry, one third science fiction, and one third “bleak” literature.
In this way, Goodwin has some creative control over the project; the source material influences the machine’s vocabulary and sentence structuring, and hence the tone of the piece.
The Thoughts Beneath the Words
The problem with artificially intelligent novelists is the same problem with conversational artificial intelligence that computer scientists have been trying to solve from Turing’s day. The machines can understand and reproduce complex patterns increasingly better than humans can, but they have no understanding of what these patterns mean.
Goodwin’s neural network spits out sentences one letter at a time, on a tiny printer hooked up to the laptop. Statistical associations such as those tracked by neural nets can form words from letters, and sentences from words, but they know nothing of character or plot.
When talking to a chatbot, the code has no real understanding of what’s been said before, and there is no dataset large enough to train it through all of the billions of possible conversations.
Unless restricted to a predetermined set of options, it loses the thread of the conversation after a reply or two. In a similar way, the creative neural nets have no real grasp of what they’re writing, and no way to produce anything with any overarching coherence or narrative.
Goodwin’s experiment is an attempt to add some coherent backbone to the AI “novel” by repeatedly grounding it with stimuli from the cameras or microphones—the thematic links and narrative provided by the American landscape the neural network drives through.
Goodwin feels that this approach (the car itself moving through the landscape, as if a character) borrows some continuity and coherence from the journey itself. “Coherent prose is the holy grail of natural-language generation—feeling that I had somehow solved a small part of the problem was exhilarating. And I do think it makes a point about language in time that’s unexpected and interesting.”
AI Is Still No Kerouac
A coherent tone and semantic “style” might be enough to produce some vaguely-convincing teenage poetry, as Google did, and experimental fiction that uses neural networks can have intriguing results. But wading through the surreal AI prose of this era, searching for some meaning or motif beyond novelty value, can be a frustrating experience.
Maybe machines can learn the complexities of the human heart and brain, or how to write evocative or entertaining prose. But they’re a long way off, and somehow “more layers!” or a bigger corpus of data doesn’t feel like enough to bridge that gulf.
Real attempts by machines to write fiction have so far been broadly incoherent, but with flashes of poetry—dreamlike, hallucinatory ramblings.
Neural networks might not be capable of writing intricately-plotted works with charm and wit, like Dickens or Dostoevsky, but there’s still an eeriness to trying to decipher the surreal, Finnegans’ Wake mish-mash.
You might see, in the odd line, the flickering ghost of something like consciousness, a deeper understanding. Or you might just see fragments of meaning thrown into a neural network blender, full of hype and fury, obeying rules in an occasionally striking way, but ultimately signifying nothing. In that sense, at least, the RNN’s grappling with metaphor feels like a metaphor for the hype surrounding the latest AI summer as a whole.
Or, as the human author of On The Road put it: “You guys are going somewhere or just going?”
Image Credit: eurobanks / Shutterstock.com Continue reading
If a recent project using Google’s DeepMind were a recipe, you would take a pair of AI systems, images of animals, and a whole lot of computing power. Mix it all together, and you’d get a series of imagined animals dreamed up by one of the AIs. A look through the research paper about the project—or this open Google Folder of images it produced—will likely lead you to agree that the results are a mix of impressive and downright eerie.
But the eerie factor doesn’t mean the project shouldn’t be considered a success and a step forward for future uses of AI.
From GAN To BigGAN
The team behind the project consists of Andrew Brock, a PhD student at Edinburgh Center for Robotics, and DeepMind intern and researcher Jeff Donahue and Karen Simonyan.
They used a so-called Generative Adversarial Network (GAN) to generate the images. In a GAN, two AI systems collaborate in a game-like manner. One AI produces images of an object or creature. The human equivalent would be drawing pictures of, for example, a dog—without necessarily knowing what a dog exactly looks like. Those images are then shown to the second AI, which has already been fed images of dogs. The second AI then tells the first one how far off its efforts were. The first one uses this information to improve its images. The two go back and forth in an iterative process, and the goal is for the first AI to become so good at creating images of dogs that the second can’t tell the difference between its creations and actual pictures of dogs.
The team was able to draw on Google’s vast vaults of computational power to create images of a quality and life-like nature that were beyond almost anything seen before. In part, this was achieved by feeding the GAN with more images than is usually the case. According to IFLScience, the standard is to feed about 64 images per subject into the GAN. In this case, the research team fed about 2,000 images per subject into the system, leading to it being nicknamed BigGAN.
Their results showed that feeding the system with more images and using masses of raw computer power markedly increased the GAN’s precision and ability to create life-like renditions of the subjects it was trained to reproduce.
“The main thing these models need is not algorithmic improvements, but computational ones. […] When you increase model capacity and you increase the number of images you show at every step, you get this twofold combined effect,” Andrew Brock told Fast Company.
The Power Drain
The team used 512 of Google’s AI-focused Tensor Processing Units (TPU) to generate 512-pixel images. Each experiment took between 24 and 48 hours to run.
That kind of computing power needs a lot of electricity. As artist and Innovator-In-Residence at the Library of Congress Jer Thorp tongue-in-cheek put it on Twitter: “The good news is that AI can now give you a more believable image of a plate of spaghetti. The bad news is that it used roughly enough energy to power Cleveland for the afternoon.”
Thorp added that a back-of-the-envelope calculation showed that the computations to produce the images would require about 27,000 square feet of solar panels to have adequate power.
BigGAN’s images have been hailed by researchers, with Oriol Vinyals, research scientist at DeepMind, rhetorically asking if these were the ‘Best GAN samples yet?’
However, they are still not perfect. The number of legs on a given creature is one example of where the BigGAN seemed to struggle. The system was good at recognizing that something like a spider has a lot of legs, but seemed unable to settle on how many ‘a lot’ was supposed to be. The same applied to dogs, especially if the images were supposed to show said dogs in motion.
Those eerie images are contrasted by other renditions that show such lifelike qualities that a human mind has a hard time identifying them as fake. Spaniels with lolling tongues, ocean scenery, and butterflies were all rendered with what looks like perfection. The same goes for an image of a hamburger that was good enough to make me stop writing because I suddenly needed lunch.
The Future Use Cases
GAN networks were first introduced in 2014, and given their relative youth, researchers and companies are still busy trying out possible use cases.
One possible use is image correction—making pixillated images clearer. Not only does this help your future holiday snaps, but it could be applied in industries such as space exploration. A team from the University of Michigan and the Max Planck Institute have developed a method for GAN networks to create images from text descriptions. At Berkeley, a research group has used GAN to create an interface that lets users change the shape, size, and design of objects, including a handbag.
For anyone who has seen a film like Wag the Dog or read 1984, the possibilities are also starkly alarming. GANs could, in other words, make fake news look more real than ever before.
For now, it seems that while not all GANs require the computational and electrical power of the BigGAN, there is still some way to reach these potential use cases. However, if there’s one lesson from Moore’s Law and exponential technology, it is that today’s technical roadblock quickly becomes tomorrow’s minor issue as technology progresses.
Image Credit: Ondrej Prosicky/Shutterstock Continue reading
Making sure artificial intelligence does what we want and behaves in predictable ways will be crucial as the technology becomes increasingly ubiquitous. It’s an area frequently neglected in the race to develop products, but DeepMind has now outlined its research agenda to tackle the problem.
AI safety, as the field is known, has been gaining prominence in recent years. That’s probably at least partly down to the overzealous warnings of a coming AI apocalypse from well-meaning, but underqualified pundits like Elon Musk and Stephen Hawking. But it’s also recognition of the fact that AI technology is quickly pervading all aspects of our lives, making decisions on everything from what movies we watch to whether we get a mortgage.
That’s why DeepMind hired a bevy of researchers who specialize in foreseeing the unforeseen consequences of the way we built AI back in 2016. And now the team has spelled out the three key domains they think require research if we’re going to build autonomous machines that do what we want.
In a new blog designed to provide updates on the team’s work, they introduce the ideas of specification, robustness, and assurance, which they say will act as the cornerstones of their future research. Specification involves making sure AI systems do what their operator intends; robustness means a system can cope with changes to its environment and attempts to throw it off course; and assurance involves our ability to understand what systems are doing and how to control them.
A classic thought experiment designed to illustrate how we could lose control of an AI system can help illustrate the problem of specification. Philosopher Nick Bostrom’s posited a hypothetical machine charged with making as many paperclips as possible. Because the creators fail to add what they might assume are obvious additional goals like not harming people, the AI wipes out humanity so we can’t switch it off before turning all matter in the universe into paperclips.
Obviously the example is extreme, but it shows how a poorly-specified goal can lead to unexpected and disastrous outcomes. Properly codifying the desires of the designer is no easy feat, though; often there are not neat ways to encompass both the explicit and implicit goals in ways that are understandable to the machine and don’t leave room for ambiguities, meaning we often rely on incomplete approximations.
The researchers note recent research by OpenAI in which an AI was trained to play a boat-racing game called CoastRunners. The game rewards players for hitting targets laid out along the race route. The AI worked out that it could get a higher score by repeatedly knocking over regenerating targets rather than actually completing the course. The blog post includes a link to a spreadsheet detailing scores of such examples.
Another key concern for AI designers is making their creation robust to the unpredictability of the real world. Despite their superhuman abilities on certain tasks, most cutting-edge AI systems are remarkably brittle. They tend to be trained on highly-curated datasets and so can fail when faced with unfamiliar input. This can happen by accident or by design—researchers have come up with numerous ways to trick image recognition algorithms into misclassifying things, including thinking a 3D printed tortoise was actually a gun.
Building systems that can deal with every possible encounter may not be feasible, so a big part of making AIs more robust may be getting them to avoid risks and ensuring they can recover from errors, or that they have failsafes to ensure errors don’t lead to catastrophic failure.
And finally, we need to have ways to make sure we can tell whether an AI is performing the way we expect it to. A key part of assurance is being able to effectively monitor systems and interpret what they’re doing—if we’re basing medical treatments or sentencing decisions on the output of an AI, we’d like to see the reasoning. That’s a major outstanding problem for popular deep learning approaches, which are largely indecipherable black boxes.
The other half of assurance is the ability to intervene if a machine isn’t behaving the way we’d like. But designing a reliable off switch is tough, because most learning systems have a strong incentive to prevent anyone from interfering with their goals.
The authors don’t pretend to have all the answers, but they hope the framework they’ve come up with can help guide others working on AI safety. While it may be some time before AI is truly in a position to do us harm, hopefully early efforts like these will mean it’s built on a solid foundation that ensures it is aligned with our goals.
Image Credit: cono0430 / Shutterstock.com Continue reading