快猫短视频

SOMETHING is going on outside your window. Rough oblongs zoom past, one of which sidles along and stops, disgorging a series of small irregular shapes in many colours. High-pitched sounds drift up as they messily assemble into a row behind a taller shape.

Microseconds pass and all becomes clear: a group of schoolchildren has pulled up in a bus and lined up behind their teacher. Your brain has taken a chaos of sensory inputs and produced a lucid experience 鈥� as it does day in, day out, throughout your life.

If only computers could do the same. We may talk about artificial intelligence learning human smarts like driving cars or playing poker, but when it comes to quickly making sense of a huge, disordered set of information, we can鈥檛 build an AI that even comes close to our brains. That鈥檚 partly down to mysteries about the workings of the human mind that we鈥檙e hard-pressed to explain. But it鈥檚 also down to a bottleneck baked into the architecture of nearly every computer for more than half a century.

Now we may be on the cusp of eliminating it thanks to a radical new computing paradigm, one that uses hardware that simultaneously stores and processes information 鈥� not unlike networks of neurons in the brain. Fulfil its promise and we could create machine minds that can parse rich streams of data in real time, spot patterns that elude us, and maybe learn without any help from humans.

Laptops, smartphones, tablets, you name it, they all adhere to an architecture that dates back to John von Neumann, one of the fathers of computing. Some 70 years ago, he proposed that computers should have separate processor and memory units. It may not sound like such a grand proposal, but it meant you no longer had to rewire a computer every time you wanted to run a fresh program. This division of labour has worked pretty well since then, allowing us to make ever-faster computers by souping up processors and memory in tandem.

But there is a catch. The von Neumann architecture means that whenever the processor needs information, it must retrieve it from the memory. That requires electrons to shuttle back and forth between the two, so the processor is often idle, waiting for data. This is one reason why your laptop probably has multiple 鈥渃ores鈥�; installing more processor units 鈥� each with their own connection to the memory 鈥� means they can request data simultaneously, speeding up the machine overall.

鈥淎 bottleneck has been baked into computer architecture for 70 years鈥�

These days the bottleneck is really beginning to constrict us. There is more data to shuttle than ever, especially with the 鈥渂ig data鈥� revolution looming. We are already glimpsing its promise: for example, there are algorithms that are who will have a heart attack than standard medical risk assessments. Designed by researchers at the University of Nottingham, UK, the algorithms achieved this prowess by digesting the electronic medical records of nearly 400,000 people, a massive data-crunching task. And with the so-called internet of things encompassing ever more everyday objects 鈥� from traffic lights to fridges 鈥� machines will have even more scope to provide insights into our lives.

Manage that correctly and it could be wonderful. But computers are already overheating under the volume of data. According to a US Department of Energy report, between 5 and 15 per cent of the world鈥檚 energy is spent on computing, much of it wasted on data trafficking. This is why the von Neumann bottleneck must be widened out or better still removed altogether.

Many attempts have been made to do that, for example, by developing programming languages that encode data more efficiently, reducing the number of electrons that need to shuttle back and forth. And in the 1980s, scientists started to consider using photons instead of electrons to encode information. Photons in optical cables travel faster than electrons in wires, so the data transfer time would be reduced. Others wanted to stick with electrons but cram more information in by encoding it into a quantum mechanical property called spin. But so far neither strategy has quite come off, mainly because they are so complex to implement that the work involved outweighs any speed gains.

In short, we have been racking our brains for an answer for decades 鈥� which is ironic because our brains are themselves a supercomputer capable of amazing feats that need no more power than a 20-watt light bulb. They have nothing like the von Neumann bottleneck, of course, because the same network of neurons both stores information and processes it.

So how to copy them? There lies the rub. By no means do we have a complete picture of how the brain works, but there are probably at least three key features needed to mimic what it does.

First, it consists of a vast network of neurons with lots of connections called synapses. Second, those connections have synaptic plasticity; that is, they can be made stronger or weaker. We know that learning manifests itself as a strengthening of the connections between sets of neurons.

The third feature is called spike-time-dependent plasticity. This idea, less well understood than other features of the brain, says that a synapse only strengthens if the two neurons fire at similar times; if they fire out of sync, then it weakens. Over the long term this process builds strong connections between neurons that are working together to pass messages around, and weakens connections that don鈥檛 seem to be important. It is thought that this is essential to the way our brains manage to learn independently. Imagine you see a green traffic light; you immediately know it means 鈥済o鈥� because the sequence of neurons involved in that thought have developed strong connections over the years.

鈥淯ntil the memristor came along, we鈥檇 never been able to mimic a synapse鈥�

In truth we have been trying to ape the way the brain computes for a long time. And this field, today known as neuromorphic computing, has seen some neat progress.

One of our earliest efforts was the gloriously named Mark 1 Perceptron, unveiled in 1958 by a researcher named Frank Rosenblatt. This wardrobe-sized array of electronics was organised in a network reminiscent of neurons. Rosenblatt showed cards bearing circles or triangles to the machine鈥檚 camera for it to name the shape, and he would then correct its mistakes. Within 50 tries, the perceptron had learned to output one signal for circles and another for triangles.

The perceptron was limited by the electrical engineering of the day, so its neural network wasn鈥檛 particularly extensive, nor were its abilities that exciting. But things have moved on considerably. Today, Google DeepMind鈥檚 neural networks can pull off impressive feats, such as when its AlphaGo program defeated the best human player of the game Go last year.

The thing about DeepMind鈥檚 neural network, however, is that it is entirely simulated in software, and runs on standard silicon electronics. So although it learns in a similar way to a network of neurons, it does not get around the von Neumann bottleneck.

滨叠惭鈥檚 , which appeared in 2014, goes further. It boasts 5.5 billion silicon transistors arranged into a brain-like architecture of 1 million interconnected 鈥渘eurons鈥�. It could recognise in real time objects like cars and bicycles in videos, using about as much power as a smartphone does in sleep mode. That sounds impressive, but if the chip were scaled up to match the 100 billion neurons of the human brain, it would consume 10,000 times more energy than the brain. 鈥淚t is actually a wasteful approach,鈥� says , a neuromorphic engineer at the University of Zurich, Switzerland.

In short, while we鈥檝e managed to mimic some of the features of the brain, we鈥檝e never been able to combine all three in a physical system. The TrueNorth chip, for example, has a lot of highly connected 鈥渘eurons鈥� but it can鈥檛 adjust the strength of the connections between them, except with software.

That failing is down to the fact that conventional electronics hasn鈥檛 delivered a device that truly mimics a synapse. But we have a way out of that impasse, thanks to an idea that surfaced almost half a century ago.

In 1971 Leon Chua, an electrical engineer at the University of California, Berkeley, was looking at the equations connecting the basic circuit components students learn about 鈥� the resistor, capacitor and inductor 鈥� when he noticed that there was another way the terms could be arranged. This produced an equation for a fourth component whose resistance would vary depending on the current. Chua called it a 鈥渕别尘谤颈蝉迟辞谤鈥� because its resistance seemed to display a memory. But with no material or device known to behave in this way, his idea was largely forgotten.

Then, about a decade ago, a team at Hewlett Packard led by Stan Williams was working on a new type of memory that would, unlike a desktop computer, retain its data when the power was switched off. The researchers were investigating devices based on thin films of titanium when they noticed that their resistance was changing in strange ways depending on the current passing through them. Eventually, they realised it wasn鈥檛 just electrons that were moving within the films, but also atoms, which subtly and reversibly changed the material鈥檚 structure and so its resistance. In other words, the team had (Nature, vol 453, p 80).

Williams鈥檚 work helped explain why memristance had never been seen before; it only manifests itself on tiny scales. But with that unearthed, a range of other materials have now been found to act as memristors, including .

The advent of real life memristors has animated researchers for several reasons, including opening up the possibility of computing in more sophisticated, efficient languages than 0s and 1s (see 鈥�Beyond binary鈥�).

But soon real action was happening on the neuromorphic computing scene. Shortly after Williams鈥檚 discovery, , an engineer at the University of Michigan, took the crucial step and showed that memristors can act as plastic synapses. He used a device made of several thin layers of silicon, one of them with a smattering of silver ions, and showed this can mimic that second feature of the brain. Lu later showed that memristors can simulate the third ingredient too; the memristor synapse could be strengthened or weakened depending on the exact timing of applied electrical spikes.

This work shows that it is 鈥渞eally an exciting time for neuromorphic engineering鈥�, says Indiveri. Beatriz Noheda, a physicist at the University of Groningen in the Netherlands, agrees. 鈥淚t鈥檚 time to give up on silicon transistors,鈥� she says, and focus on developing full-blown memristor-based neural networks.

鈥淢emristor networks pull off all three of the brain鈥檚 most crucial functions鈥�

It might seem that would be simply a case of scaling up Lu鈥檚 work. Although his efforts involved only a single synapse with an input and output neuron, it showed that memristors could pull off all three crucial functions of the brain. The way forward would be to build networks with more and more layers of networked memristor neurons; with each added layer, the network can 鈥渢hink鈥� in more sophisticated concepts (see diagram).

Not so fast, says at 滨叠惭鈥檚 Almaden research lab in California. He says the spike-time-dependent plasticity Lu has demonstrated is all very well on a small scale, but neuroscientists aren鈥檛 sure how this feature plays a role in learning on a large scale in the brain. 鈥淚t must happen somehow,鈥� he says. 鈥淏ut we aren鈥檛 even close to figuring it out.鈥� That means that implementing it in a large artificial neural network is no guarantee of getting closer to brain-like computing.

Burr prefers to stick with networks that don鈥檛 have spike-time-dependent plasticity baked in. The one he uses is like those that power Google DeepMind鈥檚 neural networks, which have plastic synapses controlled by software. But by running them on memristors rather than transistors he can potentially use thousands of times less energy.

In 2014, Burr constructed with almost 165,000 synapses. Training it on a database of handwritten letters, he then showed that it could accurately recognise them. Burr鈥檚 memristors were made from a chalcogenide glass, a material that can switch between phases where its atoms are more or less ordered, altering its conductivity. Such phase-change memristors are becoming so reliable that chip manufacturer Intel this year began selling based on them.

Others think memristors could lead to machines that can learn entirely on their own. That includes , a nanoelectronics researcher at the University of Southampton, UK. Starting small, last year he built a network of four input and two output neurons connected by memristor synapses. He could feed it electrical signals such as 鈥�1001鈥� or 鈥�0110鈥� 鈥� akin to showing circles or triangles to the 1950s perceptron. But unlike that machine, which required a human to tell it whether it had guessed the right shape, Prodromakis鈥檚 network had spike-time-dependent plasticity, and its own to fire off one output neuron when it saw 1001 and the other for 0110. This worked even with noisy input signals, an important win given that real-life data is messy.

Finally, we seem to be recreating with memristors the essence of what your brain does when you look out of a window. This is independent learning with no bottlenecks.

Suitably scaled up, such self-teaching systems could screen data in real time, for example, monitoring the behaviour of self-driving cars or the integrity of bridges or nuclear plants. This could reduce the need for more sprawling data storage centres, like the ones that store data for social networks. These are sometimes built near the Arctic because they require so much cooling. But if our data is parsed in real time by networks of memristors then maybe we don鈥檛 need to keep it.

Computers made from memristors have one more potential benefit: because they work akin to our brains, they may be easier to interface with them. There are already silicon-based devices out there that pick up brain activity and relay it to things in the physical world, enabling paralysed people to control exoskeletons, for instance, or let someone control a computer while dreaming.

But many challenges remain. The behaviour of neurons in the brain is complex in the extreme, and existing neural interfaces find all that information hard to handle. 鈥淭he electronics to process this very rich, high-bandwidth data, becomes unbearable,鈥� says Prodromakis. Memristors could be the ideal solution because they only record signals that spike significantly, ignoring the noisy background. This excites Prodromakis, who has recently started developing memristor-based neural interfaces with , a UK company formed last year in a 拢540 million partnership between GlaxoSmithKline and a Google subsidiary.

One of the biggest questions hanging over memristor networks is whether we could ever manufacture them efficiently. Silicon circuits are pumped out by well-oiled factories; would anything like that be possible for memristors? The first step to answering that question is to properly scope out the best materials to make them from, and Noheda is now setting up a research centre in Groningen to do just that. If she and the other memristor champions are successful then the computers of the future could be made from a component that, for 40 years, we thought didn鈥檛 even exist.

This article appeared in print under the headline 鈥淭hink on this鈥�