Neuromorphic systems: computers inspired by the human brain
We tell you who and why is involved in them.
/ photo Paweł Czerwiński Unsplash
What is a neuromorphic chip
This is a processor whose operation is based on the principles of the action of the human brain. Such devices simulate the work of neurons and their processes - axons and dendrites - responsible for the transmission and perception of data. Connections between neurons are formed due to synapses - special contacts through which electrical signals are transmitted.
One of the tasks of neuromorphic devices is to accelerate the training of convolutional neural networks for image recognition. Artificial intelligence systems based on this technology do not need to access a massive repository with training data over the network - all information is constantly contained in artificial neurons. This approach makes it possible to implement machine learning algorithms locally. Therefore, it is expected that neuromorphic chips will find application in mobile devices, IoT gadgets, as well as data centers.
Why engineers are inspired by the human brain
Engineers involved in the development of neuromorphic chips (we will talk more about them later), first of all, note the high computing abilities of the human brain.
According to a number of studies, our brain has the potential performance of one exaflops. Traditional supercomputers of such computing power are still being developed - the first machines are expected no earlier than 2021 .
At the same time, the brain has extremely high energy efficiency, which is the second important factor for those involved in the development of neuromorphic systems.
/ photo jesse orrico Unsplash
Of course, artificial systems have significant limitations. Neural networks differ from their biological counterparts in their inability to “remember” past skills when learning a new task. An algorithm trained to recognize dogs cannot distinguish between people. But experts from this field hope that neuromorphic chips will open up new opportunities for training multi-tasking neural networks and solve similar problems.
The first neuromorphic chips
The first attempts to create artificial neurons were made back in the 60s of the last century. Then one of the future inventors of the microprocessor Ted Hoff (Ted Hoff), together with Stanford professor Bernard Widrow (Bernard Widrow) created a single-level neural network based on memistors - electrochemical resistors with memory function. It is believed that this development laid the foundation for neuromorphic engineering.
In the 80s, engineer Carver Mead(Carver Mead) suggested using transistors as analog components, rather than digital switches. In the 90s, a team led by Mead introduced an artificial synapse capable of storing information for a long time, and a neuromorphic processor based on floating-gate transistors .
At the same time, US President George W. Bush announced the beginning of the “ Decade of the Brain ” and called for sponsorship of programs aimed at studying this organ. All this gave an impetus to the development of neuroinformatics and computational neurobiology and led to the creation of infrastructure for further study of the topic.
Over the past ten years, human knowledge about the work of the brain has reached new heights. Since 2013, Switzerland has been developing the Human Brain Project (HBP). In the same year, America launched the BRAIN Initiative . These initiatives have had a major impact on the field of artificial intelligence systems and have led to the emergence of new neuromorphic technologies.
What is being developed today
Today, neurochips are being created at IBM. Back in 2008, the company's engineers, with the support of DARPA, took part in the SyNAPSE program, which developed computer architectures other than von Neumann's. For three years, IBM managed to develop a core with 256 artificial neurons (each of them had 256 synapses). Three years later, the company introduced the TrueNorth processor , consisting of 4096 such cores - more than a million neurons. And it is already used in gesture recognition and speech recognition tasks . The company's developers say that TrueNorth-based computing systems will be able to successfully simulate the cat's brain. However, a number of experts consider such statements a clear exaggeration.
Another major IT company developing neuromorphic computing systems is Intel. Last year they introduced the Loihi chip. It contains 128 neuromorphic nuclei, each of which simulates 1024 neurons. You can program the processor using an API written in Python. The first copies of these devices have already been sent to the data centers of several leading universities to conduct tests on real tasks.
Speaking of universities, engineers from Manchester University are also working on neuromorphic chips. Last year, they introduced the SpiNNaker architecture , which consists of a million cores capable of emulating the operation of a hundred million neurons. This installation consumes 100 kW. You can program a computer using the PyNN language . To date, the machine is used to simulate the processes occurring in the mouse brain.
Despite the progress of recent years, we can say that neuromorphic iron is in the early stages of its development. The tasks that AI systems pose on its basis are mainly limited to object recognition. Nevertheless, representatives of the IT industry are convinced that in the future, neuromorphic hardware will allow for full-fledged simulations and open up completely new computing capabilities.
What we write about on our blog:
- Why vGPUs are not inferior in performance to hardware solutions
- VMware EMPOWER 2019: share experiences and talk about the most interesting
- Prospects for data centers: technologies that will improve server performance
- Processors for servers: discussing the latest
- Data Center Development: Technology Trends
- How to increase the energy efficiency of the data center