Tag Archives: news

What distinguishes Elon Musk’s reputation as an entrepreneur is that any venture he takes on comes from a bold and inspiring vision for the future of our species. Not long ago, Musk announced a new company, Neuralink, with the goal of merging the human mind with AI. Given Musk’s track record of accomplishing the seemingly impossible, the world is bound to pay extra attention when he says he wants to connect our brains to computers.
Neuralink is registered as a medical company in California. With further details yet to be announced, it will attempt to create a “neural lace,” which is a brain-machine interface that can be implanted directly into our brains to monitor and enhance them.
In the short run, this technology has medical applications and may be used to treat paralysis or diseases like Parkinson’s. In the coming decades, it could allow us to exponentially boost our mental abilities or even digitize human consciousness. Fundamentally, it is a step towards the convergence of humans and machines and maybe a leap in human progress—one that could address various challenges we face.
Current state of research
Musk isn’t the first or only person who wants to connect brains to machines. Another tech entrepreneur, Bryan Johnson, founded startup Kernel in 2016 to similarly look into brain-machine interfaces, and the scientific community has been making strides in recent years.
Earlier this month, researchers in Switzerland announced paralyzed primates could walk again with the assistance of a neuroprosthetic system. And CNN reported a man paralyzed from the shoulders down regained use of his right hand with a brain-machine interface.
The past few years have seen remarkable developments in both the hardware and software of brain-machine interfaces. Experts are designing more intricate electrodes while programming better algorithms to interpret the neural signals. Scientists have already succeeded in enabling paralyzed patients to type with their minds, and are even allowing brains to communicate with one another purely through brainwaves. So far, most of these successful applications have been in enabling motor control or very basic communication in individuals with brain injuries.
There remain, however, many challenges to ongoing developments of BMIs.
For one, the most powerful and precise BMIs require invasive surgery. Another challenge is implementing robust algorithms that can interpret the complex interactions of the brain’s 86 billion neurons. Most progress has also been one-directional: brain to machine. We have yet to develop BMIs that can provide us with sensory information or allow us to feel the subjective experience of tactile sensations such as touch, temperature or pain. (Although there has been progress giving prosthetics-users a sense of touch via electrodes attached to nerves in their arm.)
There is also the general challenge that our understanding of the brain is in its infancy. We have a long way to go before we fully understand how and where various functions such as cognition, perception and self-awareness arise. To enhance or integrate machines with these functions, we need to understand their physical underpinnings. Designing interfaces that can communicate with individual neurons and safely integrate with existing biological networks requires a great amount of medical innovation.
However, it’s important to remember this technology is rapidly advancing.
The rise of cyborgs
Hollywood often depicts a dystopian future where machines and humans go to war. Instead, however, we are seeing hints of a future where human and machine converge.
In many ways, we are already cyborgs.
Futurists like Jason Silva point out that our devices are an abstract form of brain-machine interface. We use smartphones to store and retrieve information, perform calculations and communicate with each other. According to philosophers Andy Clark and David Chalmers’ theory of the extended mind, we use technology to expand the boundaries of the human mind beyond our skulls. We use tools like machine learning to enhance our cognitive skills or powerful telescopes to enhance our visual reach. Technology has become a part of our exoskeleton, allowing us to push beyond our limitations.
Musk has pointed out that the merger of biological and machine intelligence may also be necessary if we are to remain “economically valuable.” Brain-machine interfaces could allow us to better reap the benefits of advancing artificial intelligence. With increasing automation of jobs, this could be a way to keep up with machines that perform tasks far more efficiently than we can.
Technologist Ray Kurzweil believes that by 2030s we will connect the neocortex of our brains to the cloud via nanobots. He points out that the neocortex is the source of all “beauty, love and creativity and intelligence in the world.” Notably, due to his predictive accuracy, Kurzweil has been referred to by Bill Gates and others as the best predictor of future technologies.
Whether Kurzweil is right or things take longer than expected, our current trajectory suggests we’ll get there eventually. What might such a future look like when it arrives?
We could scale our intelligence and imagination a thousand-fold. It would radically disrupt how we think, feel and communicate. Transferring our thoughts and feelings directly to others’ brains could re-define human sociality and intimacy. Ultimately, uploading our entire selves into machines could allow us to transcend our biological skins and become digitally immortal.
The implications are truly profound, and many questions remain unanswered. What will the subjective experience of human consciousness feel like when our minds are digitized? How will we prevent our digital brains from getting hacked and overwritten with unwanted thoughts? How do we ensure access to brain-machine interfaces for all, not just the wealthy?
As Peter Diamandis says, “If this future becomes reality, connected humans are going to change everything. We need to discuss the implications in order to make the right decisions now so that we are prepared for the future.”
Image Credit: Shutterstock Continue reading →

In a move reflective of HBO's hit show "Westworld," the entertainment company has filed a patent for humanoid robot characters. Continue reading →

Much of the groundbreaking AI research of recent decades originated in Canada, but it’s largely Silicon Valley that’s brought it into the real world. Now Canada is looking to take back its lead with the launch of a new research hub dedicated to the technology.
The non-profit Vector Institute, launched last week, will be based in Toronto and is designed to accelerate research and commercialization of AI and machine learning technology. The federal and provincial governments have pledged 150 million Canadian dollars (about $110 million), and a group of 31 corporate donors will also support the hub’s work over the next 10 years.
The federal government is putting forward C$40 million as part of a C$125 million countrywide artificial intelligence strategy, which will see similar institutes being established in Montreal and Edmonton.
The deep learning approach that is at the heart of most cutting-edge AI research had its genesis in Canadian universities, in particular the University of Toronto, thanks to the work of godfathers of the field like Geoffrey Hinton, Yann LeCun, and Yoshua Bengio.
But both Hinton and LeCun were lured south of the border when Silicon Valley started paying attention to the field, moving to Google and Facebook, respectively. Now Canada wants to stem the tide by attracting and retaining the field’s top global talent.
"Silicon Valley and England and other places really jumped on it, so we kind of lost the lead a little bit. I think the Canadian federal government has now realized that," Dr. Alan Bernstein, CEO and president of the Canadian Institute for Advanced Research, the organization tasked with administering the federal program, told CBC News.

"Machine learning is seen by executives as a dark art with few acolytes, and so companies like Google, Facebook, Apple and Microsoft have been hoarding talent."

They face a major challenge, though. The AI brain drain is well documented, with technology giants snapping up academics before they’ve even finished their PhDs and start-ups before they’ve even released a product.
Machine learning is seen by executives as a dark art with few acolytes, and so Silicon Valley companies like Google, Facebook, Apple and Seattle's Microsoft have been hoarding talent. Even more traditional engineering behemoths like GE and Samsung are jumping on the bandwagon, scared of being left behind.
Competing with these companies will take more than matching salary offers. Writing for The Globe and Mail back in January, while plans for the Vector Institute were nearing fruition, Hinton, who will be the institute's chief scientific adviser, its chair Ed Clark, and several other AI experts said when they asked AI researchers why they jumped ship to California, it was rarely the money.
Instead, it was the resources these companies could put at their disposal, and the chance to solve meaningful problems. To compete on these terms, they said, it will be necessary to create a critical mass of scientists, engineers, computer resources and data. That is the aim of the institute, and it will require boosting the number of machine learning graduates, forging industrial partnerships to get access to data, and acquiring the significant funding needed to support these activities.
In their article, Hinton et al talked about trying to “lure investment from foreign data-rich companies,” and they’ve already had some success. Google is helping fund the institute and has announced its intention to open an AI lab of its own in Toronto. Last November it also invested C$4.5 million in the University of Montreal's Montreal Institute for Learning Algorithms.
Thomson Reuters and General Motors both recently moved AI labs to Toronto as well, and the Royal Bank of Canada has launched a new Research in Machine Learning lab at the University of Toronto. Foteini Agrafioti, who heads that lab, told the BBC she’s hopeful these kinds of moves can stem the tide.
"I would hate to see one more professor moving south," she says. "Really, I hope that five years from now we look back and say we almost lost it but we caught it in time and reversed it."
The institute has a couple of other carrots too. Speaking to Motherboard, Hinton said, "We can offer people the chance to do any mix of basic research and applications that they want, and they're going to have lots of data, particularly from hospitals.”
Companies like Google have traditionally given their researchers a long leash, so how tempting this would be remains to be seen. But the institute does have another trump card — the political climate south of the border might make Canada a more tempting destination than before. “I think Trump might help there,” Hinton told the Toronto Star.
Whether the gambit will pay off remains to be seen, but will rely heavily on being able to convince the government and industry to invest more in the coming years. The Institute’s finances look in good shape right now, but CIFAR’s Bernstein doesn’t sugar coat it, telling CBC "it's not enough money."
"My estimate of the right amount of money to make a difference is half a billion or so, and I think we will get there," he added.
Image Credit: Shutterstock Continue reading →

AI research has a long history of repurposing old ideas that have gone out of style. Now researchers at Elon Musk’s open source AI project have revisited “neuroevolution,” a field that has been around since the 1980s, and achieved state-of-the-art results.
The group, led by OpenAI’s research director Ilya Sutskever, has been exploring the use of a subset of algorithms from this field, called “evolution strategies,” which are aimed at solving optimization problems.
Despite the name, the approach is only loosely linked to biological evolution, the researchers say in a blog post announcing their results. On an abstract level, it relies on allowing successful individuals to pass on their characteristics to future generations. The researchers have taken these algorithms and reworked them to work better with deep neural networks and run on large-scale distributed computing systems.

"To validate their effectiveness, they then set them to work on a series of challenges seen as benchmarks for reinforcement learning."

To validate their effectiveness, they then set them to work on a series of challenges seen as benchmarks for reinforcement learning, the technique behind many of Google DeepMind’s most impressive feats, including beating a champion Go player last year.
One of these challenges is to train the algorithm to play a variety of computer games developed by Atari. DeepMind made the news in 2013 when it showed it could use Deep Q-Learning—a combination of reinforcement learning and convolutional neural networks—to successfully tackle seven such games. The other is to get an algorithm to learn how to control a virtual humanoid walker in a physics engine.
To do this, the algorithm starts with a random policy—the set of rules that govern how the system should behave to get a high score in an Atari game, for example. It then creates several hundred copies of the policy—with some random variation—and these are tested on the game.
These policies are then mixed back together again, but with greater weight given to the policies that got the highest score in the game. The process repeats until the system comes up with a policy that can play the game well.

"In one hour training on the Atari challenge, the algorithm reached a level of mastery that took a [DeepMind] reinforcement-learning system…a whole day to learn."

In one hour training on the Atari challenge, the algorithm reached a level of mastery that took a reinforcement-learning system published by DeepMind last year a whole day to learn. On the walking problem the system took 10 minutes, compared to 10 hours for Google’s approach.
One of the keys to this dramatic performance was the fact that the approach is highly “parallelizable.” To solve the walking simulation, they spread computations over 1,440 CPU cores, while in the Atari challenge they used 720.
This is possible because it requires limited communication between the various “worker” algorithms testing the candidate policies. Scaling reinforcement algorithms like the one from DeepMind in the same way is challenging because there needs to be much more communication, the researchers say.
The approach also doesn’t require backpropagation, a common technique in neural network-based approaches, including deep reinforcement learning. This effectively compares the network’s output with the desired output and then feeds the resulting information back into the network to help optimize it.
The researchers say this makes the code shorter and the algorithm between two and three times faster in practice. They also suggest it will be particularly suited to longer challenges and situations where actions have long-lasting effects that may not become apparent until many steps down the line.
The approach does have its limitations, though. These kinds of algorithms are usually compared based on their data efficiency—the number of iterations required to achieve a specific score in a game, for example. On this metric, the OpenAI approach does worse than reinforcement learning approaches, although this is offset by the fact that it is highly parallelizable and so can carry out iterations more quickly.
For supervised learning problems like image classification and speech recognition, which currently have the most real-world applications, the approach can also be as much as 1,000 times slower than other approaches that use backpropagation.
Nevertheless, the work demonstrates promising new applications for out-of-style evolutionary approaches, and OpenAI is not the only group investigating them. Google has been experimenting on using similar strategies to devise better image recognition algorithms. Whether this represents the next evolution in AI we will have to wait and see.
Image Credit: Shutterstock Continue reading →