Neuromorphic engineering

Neuromorphic engineering, also known as neuromorphic computing, is a concept developed by Carver Mead, in the late 1980s, who spent over 40 years developing analysis systems designed to mimic the human senses and processing mechanisms of our bodies, such as touch, sight, hearing, and thought. Neuromorphic computing is a branch of neuromorphic engineering that focuses primarily on thinking and processing in systems that mimic human behavior.

Neuromorphic engineering describes the use of Very Large Scale Integration (VLSI) systems containing analog electronic circuits to mimic biological-neural architectures of the human nervous system. The term neuromorphic is used to describe the digital analog or a mixed digital and analog VLSI system, and software systems that implement neural system models for perception, motor control, and multisensory integration. Implementation of neuromorphic computing at the hardware level can be accomplished with memristor, spintronic memories, threshold value switches, and transistors.

Neuromorphic chips

Neuromorphic chips are computational units that attempt to emulate, from a computer science perspective, the behavior of the human brain, with particular reference to its network of neurons. The goal of neuromorphic chips is to be continuously inspired by the extraordinary efficiency of the human brain, to reach levels of performance unimaginable until now in order to perform increasingly complex calculations with low power consumption.

The biological brain represents one of the most efficient computational organs known so far, given the amount of things it can do at the same time and the amount of information it can acquire from any context in which it is located.

To make the idea clear, Intel presented an effective example that compares the biological brain of a cute parrot with the “brain” of a small autonomous drone, which we can assume constitutes the current state of the art of such technology.

Needless to say, the result, by comparison, is merciless, to say the least. The bird’s brain weighs just over 2 grams and consumes 50 mW, flies at 35 km/h without risking any collisions, can memorize and repeat many words, as well as manipulate simple objects. The drone has a SoC (System-on-Chip) with integrated CPU, GPU and memory that weighs 40 grams and consumes 18000 mW to travel at 10 km/h without learning anything in flight: everything it does, comes from information previously loaded on the system.

The biological brain, and especially the human one, is endowed with a particular characteristic: neuroplasticity, that is the ability to specialize, organize and improve progressively, adapting to the context thanks to its progressive experience. The brain is not born with a set of instructions and a data set to refer to in order to make certain decisions, but it learns from the first day it is in the world to learn from the context thanks to the information it receives from the senses.

The mechanism is decidedly complex and neuroscience has not yet been able to define many of the aspects that regulate the functioning of the brain, starting with how neurons communicate so efficiently with each other. The desire to emulate the neuroplasticity of a biological brain and the efficiency with which this is made operational represent, therefore, one of the foundations of neuromorphic computing.

Being able to translate into computer terms the biological concepts that characterize the human brain is an extremely complex operation, both from the computational point of view and from the purely functional point of view. To simulate these properties it is possible to start from two alternatives

  • develop hardware similar to the human brain
  • exploit current technologies to simulate the brain, due to the increasing computational capacity

The neuromorphic chips base their operation on some compromise solutions, aimed at overcoming the limits of classical computing based on the binary system, to access to computational forms of purely analogical inspiration, just like in biological processes.

While a digital brain proves to be very efficient in performing the operations for which it has been programmed, the biological brain continuously acquires a large amount of information and weighs it according to its importance. Neurons take in information from the senses and begin to generate some signals, communicating with each other in various ways.

To give a very simple example, when we watch television, most of our attention is focused on the content of the broadcast, but in the meantime our senses, even unconsciously, acquire other information, such as the time shown on the clock in the same room, the temperature of the room, the fact that there is a drink on the table, the phone ringing, the sound of an ambulance in the street, and so on.

Thanks to a phenomenon defined as “temporal correlation”, the television broadcast is able to generate the most intense neuronal activity in a given period, ensuring the highest level of priority among those available. Everything else is placed in the background. This guarantees the full comprehensibility of what is happening on the TV screen, without however eliminating the perception of what is happening around. If, for example, I noticed that the room was too bright, I could decide to stand up to darken it and get a better view, just as I would take the remote control to turn up the TV volume if the noise coming from outside tended to disturb the audio of the broadcast I was following.

The ability to weigh in real time the signals received from the environment is, in the first instance, an analog process, as it consists of a plurality of intermediate nuances, very complex to describe with a classical computer, based on digital logics that provide only two states: 0 and 1, on or off.

If the various branches of Artificial Intelligence were born to emulate the learning capabilities of the human mind, neuromorphic chips deal more specifically with reproducing in the best possible way the functioning of the brain, with an architecture model that, on the hardware side, reflects what Deep Learning tries to reproduce thanks to deep neural networks.

To produce neuromorphic chips, it is first necessary to discard the Von Neumann architecture model on which classical computing is based, which performs binary computation thanks to units that provide for physical and logical separation between the CPU and the memory of the system. In contrast, neuromorphic chips provide for a large number of artificial neurons, consisting of complete computational units, each equipped with CPU and memory, avoiding unnecessary waste of resources otherwise due to continuous data transfer.

The fundamental components of a traditional chip are the transistors, which are activated and deactivated according to the current flow, defining the 0 and 1 states of the binary computing system. In the case of neuromorphic chips, the equivalent is the memristor which, in addition to switching on and off, is able to remember the various intermediate stages thanks to the level of intensity of the current that has passed through it.

Without further trivializing this concept, it is enough to know that the possibility of defining many states allows the neuromorphic chip to describe the analogical aspects typical of the functioning of a biological brain. In addition to hardware, it is necessary to have software based on neuromorphic computing, usually Artificial Intelligence applications capable of fully exploiting the properties of neuromorphic chips.


Notify of

Inline Feedbacks
View all comments
Scroll to Top