真乄科技業的頂尖投資團隊

A magnetic memory cell learns to count cars

Computers that function more like the human brain rather than conventional digital systems would be based on neuronal-like networks, as opposed to logic gates or blocks, and would require reduced power compared to digital computers. Traditional computing relies on two main types of memory. The first of these is RAM, which works very fast as long as the computer is plugged into a power supply, but this type of memory does not retain unsaved data if power is lost. The second is nonvolatile memory (for example, a flash memory key, hard drive or CD), which, for its part, does store information even when it is not powered, but is much slower than RAM.

STT-MRAM is a real alternative that researchers are working on, and it is both fast and continues to store information even when the power is switched off. This is because data is stored not as electric charges but as magnetic spins – in parallel or anti-parallel states, for example.

Stochastic switching and probabilistic programming

However, there is a problem with this new technology in that it suffers from so-called stochastic switching. "The time required to program one memory state to another can be quite long because it is a random quantity," explains team leader Damien Querlioz of the Institut d’Electronique Fondamentale at the Université Paris-Sud in Orsay. "We are thus never sure whether we have actually succeeded in storing information in the nanocomponents making up these memories – magnetic tunnel junctions (MTJs) – or not." The phenomenon is called probabilistic programming.

An MTJ is the basic structure of a magnetic memory and is composed of a fixed magnetic layer, an oxide layer and a free magnetic layer, the magnetization of which can be parallel or anti-parallel to the magnetization of the fixed layer. The anti-parallel state has a high resistance while the parallel state has a low one.

Like a binary bipolar memristor

Thanks to the spin-transfer torque (STT) effect, a positive current can switch the STT-MTJ from the anti-parallel to the parallel state, while a negative current can switch it in the opposite sense. In this respect, an STT-MTJ looks very much like a binary bipolar memristor (a device that “remembers” how much current has flowed through it).

In classical memory storage, we get around the probabilistic programming problem by applying a current to the MTJs for a sufficiently long period of time (to increase the probability of switching from one memory state to another), says Querlioz, but this strategy consumes a lot of energy.

Turning probabilistic programming into an advantage

Now, his team is suggesting that probabilistic programming might itself be turned into an advantage. The researchers say that the MTJs can in fact be used as synapses (or connections) in a neuro-inspired system that requires very little energy (less than 200 nW) to work, but which can deal with a lot of data. The probabilistic programming is a way for the system to learn what its function is – in the same way the synapses in the human brain learn.

“If we choose to use short programming current pulses, rather than long ones, there is only a certain probability that the memory will change state during the pulse,” explains Querlioz. “The memory does not change state each time. This result reminds us of models used in neuroscience, and how the brain itself works. Indeed, a synapse is not programmed to learn each time but does so progressively. We found that our MTJs can be ‘trained’ in this way too.”

Performing basic cognitive tasks

In preliminary computer simulations, the researchers have shown that their system can perform basic cognitive tasks like analyzing simple images or videos. It can also recognize certain objects in the videos – something that traditional memories are unable to do, they say.

Stanley Williams, HP Senior Fellow and the Director of the Information & Quantum Systems Lab at HP, comments that the new system is a “very interesting application” of STT devices. “It is very likely that the most important uses for all types of memristors will be in various types of neuromimetic circuits, beyond standard memories,” he says.

Making hybrid CMOS/stochastic synapse circuits?

So where next? The team, which includes scientists from CEA-List in Saclay, now indeed plans to fabricate circuits containing the STT-MTJs as synapses. These circuits will be very different from those exploiting STT-MTJs as conventional memories.

“We would also like to make hybrid CMOS/stochastic synapse circuits,” adds Querlioz. “Our work might help us find new ways in which to use memristive nanodevices too. We should not always think of unpredictability inherent in nanoscale physics as being our enemy but perhaps try to exploit it as the basis for efficient information processing using novel computing concepts.”

The research is published in IEEE Transactions on Biomedical Circuits and Systems DOI: 10.1109/TBCAS.2015.2414423.

Source:http://nanotechweb.org/cws/article/tech/60960