Tag Archives: Deep learning
#437269 DeepMind’s Newest AI Programs Itself ...
When Deep Blue defeated world chess champion Garry Kasparov in 1997, it may have seemed artificial intelligence had finally arrived. A computer had just taken down one of the top chess players of all time. But it wasn’t to be.
Though Deep Blue was meticulously programmed top-to-bottom to play chess, the approach was too labor-intensive, too dependent on clear rules and bounded possibilities to succeed at more complex games, let alone in the real world. The next revolution would take a decade and a half, when vastly more computing power and data revived machine learning, an old idea in artificial intelligence just waiting for the world to catch up.
Today, machine learning dominates, mostly by way of a family of algorithms called deep learning, while symbolic AI, the dominant approach in Deep Blue’s day, has faded into the background.
Key to deep learning’s success is the fact the algorithms basically write themselves. Given some high-level programming and a dataset, they learn from experience. No engineer anticipates every possibility in code. The algorithms just figure it.
Now, Alphabet’s DeepMind is taking this automation further by developing deep learning algorithms that can handle programming tasks which have been, to date, the sole domain of the world’s top computer scientists (and take them years to write).
In a paper recently published on the pre-print server arXiv, a database for research papers that haven’t been peer reviewed yet, the DeepMind team described a new deep reinforcement learning algorithm that was able to discover its own value function—a critical programming rule in deep reinforcement learning—from scratch.
Surprisingly, the algorithm was also effective beyond the simple environments it trained in, going on to play Atari games—a different, more complicated task—at a level that was, at times, competitive with human-designed algorithms and achieving superhuman levels of play in 14 games.
DeepMind says the approach could accelerate the development of reinforcement learning algorithms and even lead to a shift in focus, where instead of spending years writing the algorithms themselves, researchers work to perfect the environments in which they train.
Pavlov’s Digital Dog
First, a little background.
Three main deep learning approaches are supervised, unsupervised, and reinforcement learning.
The first two consume huge amounts of data (like images or articles), look for patterns in the data, and use those patterns to inform actions (like identifying an image of a cat). To us, this is a pretty alien way to learn about the world. Not only would it be mind-numbingly dull to review millions of cat images, it’d take us years or more to do what these programs do in hours or days. And of course, we can learn what a cat looks like from just a few examples. So why bother?
While supervised and unsupervised deep learning emphasize the machine in machine learning, reinforcement learning is a bit more biological. It actually is the way we learn. Confronted with several possible actions, we predict which will be most rewarding based on experience—weighing the pleasure of eating a chocolate chip cookie against avoiding a cavity and trip to the dentist.
In deep reinforcement learning, algorithms go through a similar process as they take action. In the Atari game Breakout, for instance, a player guides a paddle to bounce a ball at a ceiling of bricks, trying to break as many as possible. When playing Breakout, should an algorithm move the paddle left or right? To decide, it runs a projection—this is the value function—of which direction will maximize the total points, or rewards, it can earn.
Move by move, game by game, an algorithm combines experience and value function to learn which actions bring greater rewards and improves its play, until eventually, it becomes an uncanny Breakout player.
Learning to Learn (Very Meta)
So, a key to deep reinforcement learning is developing a good value function. And that’s difficult. According to the DeepMind team, it takes years of manual research to write the rules guiding algorithmic actions—which is why automating the process is so alluring. Their new Learned Policy Gradient (LPG) algorithm makes solid progress in that direction.
LPG trained in a number of toy environments. Most of these were “gridworlds”—literally two-dimensional grids with objects in some squares. The AI moves square to square and earns points or punishments as it encounters objects. The grids vary in size, and the distribution of objects is either set or random. The training environments offer opportunities to learn fundamental lessons for reinforcement learning algorithms.
Only in LPG’s case, it had no value function to guide that learning.
Instead, LPG has what DeepMind calls a “meta-learner.” You might think of this as an algorithm within an algorithm that, by interacting with its environment, discovers both “what to predict,” thereby forming its version of a value function, and “how to learn from it,” applying its newly discovered value function to each decision it makes in the future.
Prior work in the area has had some success, but according to DeepMind, LPG is the first algorithm to discover reinforcement learning rules from scratch and to generalize beyond training. The latter was particularly surprising because Atari games are so different from the simple worlds LPG trained in—that is, it had never seen anything like an Atari game.
Time to Hand Over the Reins? Not Just Yet
LPG is still behind advanced human-designed algorithms, the researchers said. But it outperformed a human-designed benchmark in training and even some Atari games, which suggests it isn’t strictly worse, just that it specializes in some environments.
This is where there’s room for improvement and more research.
The more environments LPG saw, the more it could successfully generalize. Intriguingly, the researchers speculate that with enough well-designed training environments, the approach might yield a general-purpose reinforcement learning algorithm.
At the least, though, they say further automation of algorithm discovery—that is, algorithms learning to learn—will accelerate the field. In the near term, it can help researchers more quickly develop hand-designed algorithms. Further out, as self-discovered algorithms like LPG improve, engineers may shift from manually developing the algorithms themselves to building the environments where they learn.
Deep learning long ago left Deep Blue in the dust at games. Perhaps algorithms learning to learn will be a winning strategy in the real world too.
Image credit: Mike Szczepanski / Unsplash Continue reading
Posted in Human Robots
#437236 Why We Need Mass Automation to ...
The scale of goods moving around the planet at any moment is staggering. Raw materials are dug up in one country, spun into parts and pieces in another, and assembled into products in a third. Crossing oceans and continents, they find their way to a local store or direct to your door.
Magically, a roll of toilet paper, power tool, or tube of toothpaste is there just when you need it.
Even more staggering is that this whole system, the global supply chain, works so well that it’s effectively invisible most of the time. Until now, that is. The pandemic has thrown a floodlight on the inner workings of this modern wonder—and it’s exposed massive vulnerabilities.
The e-commerce supply chain is an instructive example. As the world went into lockdown, and everything non-essential went online, demand for digital fulfillment skyrocketed.
Even under “normal” conditions, most e-commerce warehouses were struggling to meet demand. But Covid-19 has further strained the ability to cope with shifting supply, an unprecedented tidal wave of orders, and labor shortages. Local stores are running out of key products. Online grocers and e-commerce platforms are suspending some home deliveries, restricting online purchases of certain items, and limiting new customers. The whole system is being severely tested.
Why? Despite an abundance of 21st century technology, we’re stuck in the 20th century.
Today’s supply chain consists of fleets of ships, trucks, warehouses, and importantly, people scattered around the world. While there are some notable instances of advanced automation, the overwhelming majority of work is still manual, resembling a sort of human-powered bucket brigade, with people wandering around warehouses or standing alongside conveyor belts. Each package of diapers or bottle of detergent ordered by an online customer might be touched dozens of times by warehouse workers before finding its way into a box delivered to a home.
The pandemic has proven the critical need for innovation due to increased demand, concerns about the health and safety of workers, and traceability and safety of products and services.
At the 2020 World Economic Forum, there was much discussion about the ongoing societal transformation in which humans and machines work in tandem, automating and augmenting the way we get things done. At the time, pre-pandemic, debate trended toward skepticism and fear of job losses, with some even questioning the ethics and need for these technologies.
Now, we see things differently. To make the global supply chain more resilient to shocks like Covid-19, we must look to technology.
Perfecting the Global Supply Chain: The Massive ‘Matter Router’
Technology has faced and overcome similar challenges in the past.
World War II, for example, drove innovation in techniques for rapid production of many products on a large scale, including penicillin. We went from the availability of one dose of the drug in 1941, to four million sterile packages of the drug every month four years later.
Similarly, today’s companies, big and small, are looking to automation, robotics, and AI to meet the pandemic head on. These technologies are crucial to scaling the infrastructure that will fulfill most of the world’s e-commerce and food distribution needs.
You can think of this new infrastructure as a rapidly evolving “matter router” that will employ increasingly complex robotic systems to move products more freely and efficiently.
Robots powered by specialized AI software, for example, are already learning to adapt to changes in the environment, using the most recent advances in industrial robotics and machine learning. When customers suddenly need to order dramatically new items, these robots don’t need to stop or be reprogrammed. They can perform new tasks by learning from experience using low-cost camera systems and deep learning for visual and image recognition.
These more flexible robots can work around the clock, helping make facilities less sensitive to sudden changes in workforce and customer demand and strengthening the supply chain.
Today, e-commerce is roughly 12% of retail sales in the US and is expected to rise well beyond 25% within the decade, fueled by changes in buying habits. However, analysts have begun to consider whether the current crisis might cause permanent jumps in those numbers, as it has in the past (for instance with the SARS epidemic in China in 2003). Whatever happens, the larger supply chain will benefit from greater, more flexible automation, especially during global crises.
We must create what Hamza Mudassire of the University of Cambridge calls a “resilient ecosystem that links multiple buyers with multiple vendors, across a mesh of supply chains.” This ecosystem must be backed by robust, efficient, and scalable automation that uses robotics, autonomous vehicles, and the Internet of Things to help track the flow of goods through the supply chain.
The good news? We can accomplish this with technologies we have today.
Image credit: Guillaume Bolduc / Unsplash Continue reading
Posted in Human Robots
#437222 China and AI: What the World Can Learn ...
China announced in 2017 its ambition to become the world leader in artificial intelligence (AI) by 2030. While the US still leads in absolute terms, China appears to be making more rapid progress than either the US or the EU, and central and local government spending on AI in China is estimated to be in the tens of billions of dollars.
The move has led—at least in the West—to warnings of a global AI arms race and concerns about the growing reach of China’s authoritarian surveillance state. But treating China as a “villain” in this way is both overly simplistic and potentially costly. While there are undoubtedly aspects of the Chinese government’s approach to AI that are highly concerning and rightly should be condemned, it’s important that this does not cloud all analysis of China’s AI innovation.
The world needs to engage seriously with China’s AI development and take a closer look at what’s really going on. The story is complex and it’s important to highlight where China is making promising advances in useful AI applications and to challenge common misconceptions, as well as to caution against problematic uses.
Nesta has explored the broad spectrum of AI activity in China—the good, the bad, and the unexpected.
The Good
China’s approach to AI development and implementation is fast-paced and pragmatic, oriented towards finding applications which can help solve real-world problems. Rapid progress is being made in the field of healthcare, for example, as China grapples with providing easy access to affordable and high-quality services for its aging population.
Applications include “AI doctor” chatbots, which help to connect communities in remote areas with experienced consultants via telemedicine; machine learning to speed up pharmaceutical research; and the use of deep learning for medical image processing, which can help with the early detection of cancer and other diseases.
Since the outbreak of Covid-19, medical AI applications have surged as Chinese researchers and tech companies have rushed to try and combat the virus by speeding up screening, diagnosis, and new drug development. AI tools used in Wuhan, China, to tackle Covid-19 by helping accelerate CT scan diagnosis are now being used in Italy and have been also offered to the NHS in the UK.
The Bad
But there are also elements of China’s use of AI that are seriously concerning. Positive advances in practical AI applications that are benefiting citizens and society don’t detract from the fact that China’s authoritarian government is also using AI and citizens’ data in ways that violate privacy and civil liberties.
Most disturbingly, reports and leaked documents have revealed the government’s use of facial recognition technologies to enable the surveillance and detention of Muslim ethnic minorities in China’s Xinjiang province.
The emergence of opaque social governance systems that lack accountability mechanisms are also a cause for concern.
In Shanghai’s “smart court” system, for example, AI-generated assessments are used to help with sentencing decisions. But it is difficult for defendants to assess the tool’s potential biases, the quality of the data, and the soundness of the algorithm, making it hard for them to challenge the decisions made.
China’s experience reminds us of the need for transparency and accountability when it comes to AI in public services. Systems must be designed and implemented in ways that are inclusive and protect citizens’ digital rights.
The Unexpected
Commentators have often interpreted the State Council’s 2017 Artificial Intelligence Development Plan as an indication that China’s AI mobilization is a top-down, centrally planned strategy.
But a closer look at the dynamics of China’s AI development reveals the importance of local government in implementing innovation policy. Municipal and provincial governments across China are establishing cross-sector partnerships with research institutions and tech companies to create local AI innovation ecosystems and drive rapid research and development.
Beyond the thriving major cities of Beijing, Shanghai, and Shenzhen, efforts to develop successful innovation hubs are also underway in other regions. A promising example is the city of Hangzhou, in Zhejiang Province, which has established an “AI Town,” clustering together the tech company Alibaba, Zhejiang University, and local businesses to work collaboratively on AI development. China’s local ecosystem approach could offer interesting insights to policymakers in the UK aiming to boost research and innovation outside the capital and tackle longstanding regional economic imbalances.
China’s accelerating AI innovation deserves the world’s full attention, but it is unhelpful to reduce all the many developments into a simplistic narrative about China as a threat or a villain. Observers outside China need to engage seriously with the debate and make more of an effort to understand—and learn from—the nuances of what’s really happening.
This article is republished from The Conversation under a Creative Commons license. Read the original article.
Image Credit: Dominik Vanyi on Unsplash Continue reading
Posted in Human Robots
#437202 Scientists Used Dopamine to Seamlessly ...
In just half a decade, neuromorphic devices—or brain-inspired computing—already seem quaint. The current darling? Artificial-biological hybrid computing, uniting both man-made computer chips and biological neurons seamlessly into semi-living circuits.
It sounds crazy, but a new study in Nature Materials shows that it’s possible to get an artificial neuron to communicate directly with a biological one using not just electricity, but dopamine—a chemical the brain naturally uses to change how neural circuits behave, most known for signaling reward.
Because these chemicals, known as “neurotransmitters,” are how biological neurons functionally link up in the brain, the study is a dramatic demonstration that it’s possible to connect artificial components with biological brain cells into a functional circuit.
The team isn’t the first to pursue hybrid neural circuits. Previously, a different team hooked up two silicon-based artificial neurons with a biological one into a circuit using electrical protocols alone. Although a powerful demonstration of hybrid computing, the study relied on only one-half of the brain’s computational ability: electrical computing.
The new study now tackles the other half: chemical computing. It adds a layer of compatibility that lays the groundwork not just for brain-inspired computers, but also for brain-machine interfaces and—perhaps—a sort of “cyborg” future. After all, if your brain can’t tell the difference between an artificial neuron and your own, could you? And even if you did, would you care?
Of course, that scenario is far in the future—if ever. For now, the team, led by Dr. Alberto Salleo, professor of materials science and engineering at Stanford University, collectively breathed a sigh of relief that the hybrid circuit worked.
“It’s a demonstration that this communication melding chemistry and electricity is possible,” said Salleo. “You could say it’s a first step toward a brain-machine interface, but it’s a tiny, tiny very first step.”
Neuromorphic Computing
The study grew from years of work into neuromorphic computing, or data processing inspired by the brain.
The blue-sky idea was inspired by the brain’s massive parallel computing capabilities, along with vast energy savings. By mimicking these properties, scientists reasoned, we could potentially turbo-charge computing. Neuromorphic devices basically embody artificial neural networks in physical form—wouldn’t hardware that mimics how the brain processes information be even more efficient and powerful?
These explorations led to novel neuromorphic chips, or artificial neurons that “fire” like biological ones. Additional work found that it’s possible to link these chips up into powerful circuits that run deep learning with ease, with bioengineered communication nodes called artificial synapses.
As a potential computing hardware replacement, these systems have proven to be incredibly promising. Yet scientists soon wondered: given their similarity to biological brains, can we use them as “replacement parts” for brains that suffer from traumatic injuries, aging, or degeneration? Can we hook up neuromorphic components to the brain to restore its capabilities?
Buzz & Chemistry
Theoretically, the answer’s yes.
But there’s a huge problem: current brain-machine interfaces only use electrical signals to mimic neural computation. The brain, in contrast, has two tricks up its sleeve: electricity and chemicals, or electrochemical.
Within a neuron, electricity travels up its incoming branches, through the bulbous body, then down the output branches. When electrical signals reach the neuron’s outgoing “piers,” dotted along the output branch, however, they hit a snag. A small gap exists between neurons, so to get to the other side, the electrical signals generally need to be converted into little bubble ships, packed with chemicals, and set sail to the other neuronal shore.
In other words, without chemical signals, the brain can’t function normally. These neurotransmitters don’t just passively carry information. Dopamine, for example, can dramatically change how a neural circuit functions. For an artificial-biological hybrid neural system, the absence of chemistry is like nixing international cargo vessels and only sticking with land-based trains and highways.
“To emulate biological synaptic behavior, the connectivity of the neuromorphic device must be dynamically regulated by the local neurotransmitter activity,” the team said.
Let’s Get Electro-Chemical
The new study started with two neurons: the upstream, an immortalized biological cell that releases dopamine; and the downstream, an artificial neuron that the team previously introduced in 2017, made of a mix of biocompatible and electrical-conducting materials.
Rather than the classic neuron shape, picture more of a sandwich with a chunk bitten out in the middle (yup, I’m totally serious). Each of the remaining parts of the sandwich is a soft electrode, made of biological polymers. The “bitten out” part has a conductive solution that can pass on electrical signals.
The biological cell sits close to the first electrode. When activated, it dumps out boats of dopamine, which drift to the electrode and chemically react with it—mimicking the process of dopamine docking onto a biological neuron. This, in turn, generates a current that’s passed on to the second electrode through the conductive solution channel. When this current reaches the second electrode, it changes the electrode’s conductance—that is, how well it can pass on electrical information. This second step is analogous to docked dopamine “ships” changing how likely it is that a biological neuron will fire in the future.
In other words, dopamine release from the biological neuron interacts with the artificial one, so that the chemicals change how the downstream neuron behaves in a somewhat lasting way—a loose mimic of what happens inside the brain during learning.
But that’s not all. Chemical signaling is especially powerful in the brain because it’s flexible. Dopamine, for example, only grabs onto the downstream neurons for a bit before it returns back to its upstream neuron—that is, recycled or destroyed. This means that its effect is temporary, giving the neural circuit breathing room to readjust its activity.
The Stanford team also tried reconstructing this quirk in their hybrid circuit. They crafted a microfluidic channel that shuttles both dopamine and its byproduct away from the artificial neurons after they’ve done their job for recycling.
Putting It All Together
After confirming that biological cells can survive happily on top of the artificial one, the team performed a few tests to see if the hybrid circuit could “learn.”
They used electrical methods to first activate the biological dopamine neuron, and watched the artificial one. Before the experiment, the team wasn’t quite sure what to expect. Theoretically, it made sense that dopamine would change the artificial neuron’s conductance, similar to learning. But “it was hard to know whether we’d achieve the outcome we predicted on paper until we saw it happen in the lab,” said study author Scott Keene.
On the first try, however, the team found that the burst of chemical signaling was able to change the artificial neuron’s conductance long-term, similar to the neuroscience dogma “neurons that fire together, wire together.” Activating the upstream biological neuron with chemicals also changed the artificial neuron’s conductance in a way that mimicked learning.
“That’s when we realized the potential this has for emulating the long-term learning process of a synapse,” said Keene.
Visualizing under an electron microscope, the team found that, similar to its biological counterpart, the hybrid synapse was able to efficiently recycle dopamine with timescales similar to the brain after some calibration. By playing with how much dopamine accumulates at the artificial neuron, the team found that they loosely mimic a learning rule called spike learning—a darling of machine learning inspired by the brain’s computation.
A Hybrid Future?
Unfortunately for cyborg enthusiasts, the work is still in its infancy.
For one, the artificial neurons are still rather bulky compared to biological ones. This means that they can’t capture and translate information from a single “boat” of dopamine. It’s also unclear if, and how, a hybrid synapse can work inside a living brain. Given the billions of synapses firing away in our heads, it’ll be a challenge to find-and-replace those that need replacement, and be able to control our memories and behaviors similar to natural ones.
That said, we’re inching ever closer to full-capability artificial-biological hybrid circuits.
“The neurotransmitter-mediated neuromorphic device presented in this work constitutes a fundamental building block for artificial neural networks that can be directly modulated based on biological feedback from live neurons,” the authors concluded. “[It] is a crucial first step in realizing next-generation adaptive biohybrid interfaces.”
Image Credit: Gerd Altmann from Pixabay Continue reading
Posted in Human Robots