Last month, Google’s AI division, DeepMind, announced that its computer had defeated Europe’s Go champion in five straight games. Go, a strategy game played on a 19×19 grid, is exponentially more difficult for a computer to master than chess—there are 20 possible moves to choose from at the start of a chess game compared to 361 moves in Go—and the announcement was lauded as another landmark moment in the evolution of artificial intelligence.
Google, Facebook, and IBM have all gone all-in on brain-like computers that promise to emulate the mind of a human. The ability to learn and recognize patterns is viewed as a key next step in the evolution of AI. But Oshiorenoya Agabi believes the brain-like processors are missing one key component: actual brains.
Or, at least, living neurons. His startup, Koniku, which just completed a stint at the biotech accelerator IndieBio, touts itself as “the first and only company on the planet building chips with biological neurons.” Rather than simply mimic brain function with chips, Agabi hopes to flip the script and borrow the actual material of human brains tocreate the chips. He’s integrating lab-grown neurons onto computer chips in an effort to make them much more powerful than their standard silicon forebears.
Koniku is fundraising towards a first-round goal of $6.3 million, Agabi says. It has already landed customers in the aviation and pharmaceuticals industries, like AstraZeneca, the UK-based pharma company, Agabi says, and Boeing has signed on with a letter of intent to use the tech in chemical-detecting drones. The first batch of neuron-abetted chips are set to ship in the next few months. Agabi says that one customer, a drone company, hopes the processors will prove superior in detecting methane leaks in oil refineries. Another aims to use the processors to model the effect certain drugs will have on a human brain.
The future, Agabi believes, will run on a computer that’s much more alive.
Part of Koniku’s funding success seems to come from his genuine, even romantic vision of neuron-based chips as the future of processing. When I interviewed Agabi recently, his excitement over the future of neurotechnology was palpable.
Agabi, who was born in Nigeria, told me he first became interested in machine learning while teaching a pick-and-place robotic arm to classify objects for the Swiss robotics company, Neuronics. After eight years, he left the company to pursue his Masters in theoretical physics, focusing his thesis on the challenge of interfacing neurons with a robot. He spent the next four working to build a robotic arm that could attach to an amputee, eventually leaving to move to London to pursue his PhD in bioengineering.
Basically, he hopes to build a computer chip with living, learning processors.
Recognizing the imposing nature of his resume, the engineer paused for a moment, attempting to simplify his life’s work. “Essentially, for the last fifteen years, I have worked to understand how neurons talk to each other,” he said. “I’ve worked on how to communicate with individual neurons—how to read information from them and write information into them.”
This ability to code specific tasks into neurons, born out of Agabi’s specialized history, is the essence of what Koniku is hoping to accomplish. Through years of teaching machines to learn, and through the study of the brain’s mechanics, he believes that his team will be able to organize living neurons into circuits built to perform precise tasks—basically, he hopes to build a computer chip with living, learning processors.
“We take the radical view that you can actually compute with real, biological neurons,” he said.
Since the silicon transistor was created in 1947, the amount of transistors that can be crammed onto a chip has grown from a few thousand to more than 2 billion. Today, chip manufacturers have shrunk the size of each silicon transistor to the equivalent of three strands of DNA. Agabi said that because there is a limit to how tiny you can shrink the deep lens of a silicon transistor (IBM announced the creation of a 7 nanometers transistor in July, and a single silicon atom is 0.2 nm), silicon-based processing can only get so powerful.
“In the cycle of accelerating computing power, we’ve gone from the slate to the paper, from the paper to mechanical systems, mechanical systems to the vacuum tube, vacuum tubes to silicon,” he said. “And now we are moving to neurons.”
For a frame of reference, Dr. Laeeq Evered, a professor of neuropsychology at the Wright Institute, tells me that “a piece of brain matter the size of a grain of sand contains approximately 100,000 neurons, 2 million axons, and 1 billion synapses.”
There is, of course, a quixotic quality to the dream of actually creating an artificial chip so small and powerful, but Agabi feels he’s found the path to it. I asked Dr. Evered whether he thought it impossible to ever build a chip as powerful as the human brain.
“That’s what I think, but I think we’ve all been astounded by the progress of technology,” he said, and laughed. “So, we’ll see.”
Agabi told me he believes any hesitations around neuron-based chips will vanish when Koniku can successfully and publicly exhibit the chip’s practical application. “You want to build ideas that people will say, ‘That’s so obvious.’ Today, it’s not that way because no one has demonstrated this yet,” he said. “But I feel very confident that in two years when we demonstrate it, it will become like, ‘Ah, this is obvious.’”
For a third opinion, I turned to Sherif Eid, a systems engineer behind the deep-learning program DRIVE PX that some believe could be the key to the self-driving car. He said he was intrigued by the idea of neuron-based processors, but he said the technology was still based on a lot of unknowns.
“It is just that there are so many secrets we haven’t unlocked yet in the brain,” he said. “The neuron-based chips could unlock something in the future, but it takes investors with so much faith or very deep pockets willing to throw away money to see what comes out of it.”
Eid thinks it will be a few decades yet before the neuron-based processors were adopted, if they were ever adopted at all. In Agabi’s view, however, the technology is inevitable—and on the horizon today. He told me he believes his chips will be powering robotics around the world within five years. Which raises the question: What happens if he actually pulls it off?
When I first heard of Koniku, I was a little spooked. I’ve kept a close eye on the race toward true artificial intelligence and have been most persuaded by philosopher Nick Bostrom’s calls for caution. To me, Koniku felt like a potential Skynet moment—Agabi, after all, seemed to be planning to give the machines human brains.
“Carbon is a material like any other material. So for us the premise that we start from is that neurons are a material.”
Naturally, I mentioned that infamous malignant AI to Agabi, and asked if he was burdened by the effect theTerminatorfilms have had on his research. “Yes, yes, yes,” he said, letting out a wearied laugh. He told me the idea that his company was putting human parts into machines was just a simple case of anthropomorphizing. Neurons are present in many animal brains aside from humans, and Agabi reminded me that Koniku’s neurons are grown in a lab. “Carbon is a material like any other material,” he said. “So for us the premise that we start from is that neurons are a material.”
For Agabi, what he calls the “AI drama” is much less interesting than the simple question of efficiency. He notes thatthe Tianhe-2, the most powerful supercomputer built to date, demands 24 megawatts of power, while the human brain runs on just 10 watts. In other words, he says, the most powerful computer on earth burns 2.4 million times the energy of the human brain. “It’s not a matter of luxury, or just because we can do it. It’s a matter of urgency,” he said. “We have to find a way to build much more with less if we as a species are going to survive.”
Dr. Evered agrees that much of the brain’s tremendous efficiency stems from its ability to learn to recognize and reinforce the optimal connections between neurons. Though we are born with 100 billion neurons, we lose 100,000 per day—and it is the ability for the remaining neurons to form connections with beneficial counterparts that determines the power of the brain.
“It’s not a question of nature or nurture. It’s nature and nurture. We’re going to have a certain number of neurons and neuronal connections that have genetic determination,” he said. “But then, interacting with our environment is at least as critical, if not more so. It’s those connections, through learning and development, that are going make a very strong brain.”
Thus, much of the challenge in creating brain-like processors will be the pursuit of programing adaptability into computers. In a lab setting, Agabi says Koniku has proven its chips are capable of deep learning—the ability to recognize patterns and retain that knowledge—by demonstrating a process called spike timing-dependent plasticity, the idea that neurons build circuits with beneficial neurons.
Agabi believes his neuronic chips will be better at learning than traditional silicon processors, because they can more closely mirror human brain function.
Near the end of our conversation, I asked Agabi if he thought his neuron-powered chips could be the key to powering humanity past Moore’s Law—the rule that holds that the processing power of computers will double every two years. There has been concern among some experts that Moore’s Law has plateaued and that the future of AI will rest on engineers’ ability to find faster and more efficient computation than is currently available. Agabi points out that Moore’s Law only applies to increases in computational power through the addition of more silicon processors on a chip over the last 70 years—and argues that it may just be part of a larger trend; a second law, one that describes the long-tail improvement of computing tools over the course of human history.
It will, he says, take moving away from silicon itself to allow Silicon Valley to continue to innovate.
“The fact that we’re constantly getting increases in computational power, tha tlaw of computing holds for the last 2,000 to 5,000 years. Moore’s Law is a little patch, it’s only a little piece of that law,” he said. “One of these two laws is going to have to give in, and I suspect that Moore’s Law is the one that’s going to have to give in. But our power to calculate faster and faster—that law is here to stay.”