THE NEW ERA OF NEUROMORPHIC TECHNOLOGIES

Literally a year ago it was interesting to me, how the human brain works therefore I began with the help of artificial intelligence to compare biological elements of the brain with electronic ones. And as it turned out there are very many common elements similar by the principle of their work.

For example, neurons of the brain can be compared with electronic transistors, by the principle of work they perform the same functions as neurons in the human brain. That is, essentially this is like a processor, which consists of a billion transistors, but works in a computer as a separate part, which is installed on a motherboard, which connects it with RAM memory and data storage HDD or SSD. But I began to be more interested in what then is similar and is RAM — operational memory and SSD — memory, in the human brain. And as later turned out after careful analysis and search of data, these were — synapses. The brain — is a dense unified substance, in which neurons and synapses work in a pair, that is — these are not separate independent components, but a closely connected pair. A neuron receives signals through many synapses, each synapse has its own “strength” (conductivity), which influences how strongly the incoming signal will influence the neuron, after which the neuron integrates all these weighted signals and “decides” to fire or not. When a neuron fires, this influences back on the synapses — therefore they can strengthen or weaken.

After, a thought came to my head about that, if neurons — are like a transistor, then what unified electronic element can be similar to — synapse. I again began to dig and search for information after which I found, a unified passive electronic element. In electronic classification it is described as the fourth basic passive element (alongside with resistor, capacitor and inductance), proposed in 1971 by Leon Chua. It turned out to be — Memristor (from English memristor — memory resistor) that is: Memristor — is an element, the resistance of which depends on the history of current passage, it “remembers” its state even when power is disconnected. Essentially simultaneously replaces RAM and SSD if it is united in a pair like a neuron with a synapse, but only a transistor with a memristor. Since the memristor remembers the previous state (like flash-memory). At the given moment it is used in perspective technologies: neuromorphic systems, energy-saving memory. For the first time realized physically not so long ago, in 2008 (Hewlett-Packard Labs).

That is, when one neuromorphic chip or processor is formed, the need for operational and main memory falls away. Essentially if the box of the most powerful computer was the size of an ordinary system unit, then, when this technology realizes and begins to be implemented everywhere in electronics, the most powerful computer will be no larger by sizes than your smartphone. Precisely therefore modern developers create hybrid architectures, where: Transistor circuits imitate the integrating function of a neuron, and Memristors store synaptic weights and provide plasticity. They work in close connection, as in the biological brain. Such systems as Intel Loihi or IBM TrueNorth precisely realize this concept — each artificial neuron is connected with others through programmable synoptic connections, where the change of “strength” of connection influences the behavior of the entire network.

Memristors store and process information in one and the same place — that is, there is no separation, as in classical computers (processor ↔ RAM ↔ disk). This eliminates the narrow bottleneck of von Neumann architecture. The memristor holds the state of the process without power — replacing both RAM and SSD/HDD. Memristor matrices work simultaneously across the entire field, like the brain. Which is impossible to realize practically on classical architecture, where operations go sequentially or with limited parallelism.

The principle of neuromorphic system is realized 1T1R — one transistor + one memristor or 2T1R — two transistors + one memristor. The transistor manages current, preventing unwanted leaks and interference, The memristor stores information (analog of weight in a neural network). In connection it turns out to be a managed, stable and trainable node, suitable for neuro-like computations, analogous to those which are realized in the human brain, but without error and loss of held information. Precisely the tandem “transistor + memristor” brings us closer to the creation of an artificial brain, capable of associative thinking, memory and learning.

Therefore it became interesting to me, how to program such technology, and I had to turn with this question to artificial intelligence, after which it gave an answer about that: — That precisely here begins a revolution in programming. That is, classical code (in C+, Python, Java and so on) no longer manages directly such a processor. Instead of usual instructions a new paradigm is needed. Neuro-like programming in which code is not written in the habitual form. Instead of this: weights are trained (resistances of memristors), as in neural networks. That is management — through data and learning, and not code algorithms. Environment of the type “spiking neural networks” (SNN), where they begin to use pulse signals, as in the real brain. Management goes through “spikes” (bursts of voltage) — not by code, but by patterns of signals, with the help of hardware description (HDL). Where at the stage of design: Verilog, VHDL is used — for description of the logic of the chip. But already in the ready chip — everything works through learning, and not execution of lines of code.

Example of “code” for such a chip:

This will be not if/else, but, for example: Connection to an array of inputs, Preparation of a training set, Launch of learning (through pulses), Check of result — as on a neural network. That is it will be necessary to program an artificial brain on memristors — not with the help of coding, but learning. You do not give instructions — but form the behavior of processes through real experience. After which it became clear to me, that in order to lay down a pure initial algorithm of learning, for such a processor or electronic brain, will be required an “artificial analyzer of perception,” as in a human, based on the perception of 6 senses. Traditionally they speak about five senses — vision, hearing, smell, taste and touch. But in reality a human has much more sensory systems! As for me and in my view there are 6 of them, the sixth is a mental sense, depending on internal experiences depending on well-being and mood that, which truly can be a separate sensory modality — internal emotional and mental perception. A sense, which we intuitively feel, but which is difficult to describe in terms of classical physiology. I had to come to this conclusion because in 2017 I was developing a concept, for a device — “Artificial Analyzer of Perception (AAP),” due to the absence of demand and financing the project was curtailed and terminated. But in reality modern science does not stop and distinguishes now about 20+ senses.

Therefore neuromorphic processors open revolutionary possibilities in various spheres thanks to their energy efficiency and ability for adaptive learning. These are the main areas of their application:

Edge Computing and IoT devices

Low energy consumption of neuromorphic architectures can help with the problem of short battery life of devices, such as smartphones and wearable devices, and their adaptability and event-oriented nature are suitable for methods of information processing of remote sensors, drones and other Internet of Things devices | IBM. Processing of big data on end devices becomes possible without the necessity of additional computational powers Neuromorphic computing — new thinking RII.

Autonomous robotics

Neuromorphic processors are already used in various areas, including sensorics, robotics, healthcare and large-scale AI applications. AI-applications in autonomous vehicles, smart homes, personal robotics and space exploration rely on fast computations Creating Futuristic Edge Systems with Neuromorphic Computing, which makes neuromorphic systems ideal for such tasks.

Medicine and healthcare

Thanks to extensive possibilities of parallel processing, neuromorphic computing can be used in machine learning applications for recognition of patterns in natural language and speech, analysis of medical images and processing of signals of images from fMRI of the brain. In the area of smart wearable devices, better battery life is achieved and more accurate health monitoring through advanced sensors Neuromorphic Computing. But on the other hand such technologies can be used for harm, when obtaining uncontrolled access to such technologies, manipulating a person and their consciousness. Even now this is possible, if illegally using, for example, such technology as Wi-Fi Sensing it is sufficient to place several sensors in a person’s apartment or use a program through home Wi-Fi. Although in perspective such technology is applied, in normal conditions for confidential supervision of elderly people.

Smart cities and transport

Neuromorphic systems contribute to innovations in electronic healthcare, science, education, transport, planning of smart cities and the metaverse Neuromorphic Computing: Cutting-Edge Advances and Future Directions. They are especially effective for real-time systems, requiring fast decision-making.

Communication technologies

Neuromorphic approaches significantly improve signal processing and network optimization in communication technologies, especially in satellite communication and IoT Neuromorphic Computing.

Intel — Loihi/Loihi 2

Intel Lab’s new chip Loihi 2 surpasses its predecessor up to 10 times and comes with an open community-oriented framework of neuromorphic computing under the name Lava On the path to an artificial brain: a new neuromorphic chip is created. The chip Loihi 2 at a high level: 128 neuromorphic cores, but now each core has 8 times more neurons and synapses. Each of these 128 cores has 192 KB of flexible memory Neuromorphic processor from India multiplies matrices hundreds of times more efficiently than graphic chips.

System Hala Point — Intel calls this new system Hala Point, and this is a major breakthrough compared to the first generation neuromorphic chip-system of the company, called Pohoiki Springs, providing 10 times more neural capacity and up to 12 times higher performance Neuromorphic computing — new thinking AI. Systems based on Loihi can execute AI-inference and solve optimization tasks, using 100 times less energy at speeds up to 50 times faster, than ordinary CPU and GPU architectures

Neuromorphic processors.

IBM — TrueNorth

IBM decided to try to emulate the brain with TrueNorth, a 4096-core chip, packing 1 million neurons and 256 million synapses Creating Futuristic Edge Systems with Neuromorphic Computing. The newest IBM neurosynaptic computer chip, called TrueNorth, consists of 1 million programmable neurons and 256 million programmable synapses, transmitting signals between digital neurons.

TrueNorth — is a 5.4 billion transistor, 4096-core, 1M neuron, 256M synaptic neurosynaptic chip, realized in 28nm technology. Thanks to a mixed asynchronously-synchronous design and custom instrumental flow, it achieves 58GSOPS and 400GSOPS/W efficiency when operating neural networks in real-time mode on 65mW Neuromorphic computing: The future of AI and beyond — Atos.

SpiNNaker 1 (Great Britain, University of Manchester)

Part of Human Brain Project. Both architectures BrainScaleS and SpiNNaker will be discussed during a webinar on March 22 which took place in 2023. Together the systems, located in Heidelberg and Manchester, compose the “Platform of neuromorphic computing” of the project Human Brain Project Neuromorphic Computing and Engineering with AI. And this was available already in 2024 Optimizing embedded edge AI with neuromorphic computing — Embedded. SpiNNaker 2 — is already a new project, which is being developed at Dresden Technical University, Germany. This is not simply a continuation of the British project, but its separate development.

BrainScaleS (Germany, Heidelberg)

Also part of European Human Brain Project, works in tandem with SpiNNaker2 as a unified neuromorphic platform for research.

Chinese contribution

Tianjic is based on a 156-core architecture with localized memory and simplified data flow, which can be used for simulation of 40,000 neurons and 10 million synapses. The chip supports both artificial neurons and spiking neurons.

Key American achievements

Intel Loihi 2 — the most advanced commercial neuromorphic processor with open framework Lava and system Hala Point for large-scale research.

IBM TrueNorth — pioneer in the area of digital neuromorphic architectures with impressive energy efficiency.

European projects — focus on large-scale modeling of the brain and fundamental research of neuromorphic principles.

Russian developments

Research of neuromorphic systems in Russia by 2024 is at a high level, comparable with international Neuromorphic processors. The company “Motiv NT” from Novosibirsk develops systems of technical vision and hardware solutions for their work, having laid out open-source code for creation, training and use of pulse neural networks — Russian neuromorphic processor.

Future perspectives

In everyday computing neuromorphic computing can be used together with AI for automation of simple tasks, such as combining PDF-documents, creating resumes, keeping notes of meetings and much more — all this with lightning speed without impact on the computational power of the device Optimizing embedded edge AI with neuromorphic computing — Embedded.

The key advantage of neuromorphic processors — their ability to learn and adapt in real time with minimal energy consumption, opens the path to creation of reality of intelligent autonomous systems.

Speaking in simple language if your smartphone with active use with a battery of 5000 mAh is enough for 1–2 days, then after implementation of neuromorphic technologies in your smartphone will extend this active mode of work up to 20 days. Reduction of internet traffic will substantially decrease, due to accounting for your needs and their forecasting. At the same time in your phone there will be a full-fledged and autonomous artificial intelligence in the form of a personal assistant, at the level of real communication as with a person, significantly exceeding modern language models of AI. The technology will reach commercial maturity (approximately in the year 2030). Now one can only imagine what human-like robotized machines will become. But the result is more than obvious, later they will begin to implement accordingly laws of treatment with human-like androids, since the attitude toward them will have to be equal to the level of a human, otherwise understanding of aggression and reality will go against humanity or individual persons. Since they will no longer simply play with you according to learned patterns, but truly “plan” strategy, evaluating positions, making decisions in real time.

RAI — Real Artificial Intelligence (real artificial intelligence). This is an unofficial term, which usually is used for emphasizing, that the matter concerns not a simulation or voice model, but an independently thinking, learning and adapting system, maximally brought close to human thinking.

© Arthur KEHLER, 19/10/2025 FR