An expert in neuromorphic technologies at "Kaspersky Lab," Andrey Lavrentyev, discussed the future of AI. An innovative neuromorphic chip has been created, and its capabilities open up a new direction in the development of smart technology. This chip is already ready, implemented in hardware, and was created by the company "Motiv-NT" in partnership with specialists from "Kaspersky Lab." Connected to this chip, a neuromorphic camera, using pulse neural networks, can see and analyze millions of events per second, such as counting falling grains of sand in an hourglass. Moreover, it does this while consuming thousands of times less energy. Andrey Lavrentyev, Head of Technology Development at "Kaspersky Lab," spoke to "Gazeta.Ru" about the prospects for the development of such devices and the neuromorphic AI of the future. A Unique Chip
— What is the new neuromorphic chip?
— Neuromorphic means biologically inspired. The artificial intelligence that everyone has heard of, such as ChatGPT, works based on what are known as artificial neural networks. They take a number as input and produce a number as output. They work by summing these numbers in the neurons of such a network, multiplied by the weights of the connections, and then calculating activation, which gives the number at the neuron's output.
There are no numbers in the biological brain. There are electrical impulses, or spikes. There's a signal – an impulse goes. No signal – no impulse.
Our chip operates on the same principle. It works with spiking neural networks. We call it a neuromorphic platform.
— What sets your chip apart from regular ones?
— Its key difference is energy efficiency. There's an event – there's work. No event – no work. A regular neural network, on the other hand, is engaged in computations constantly, which is why it requires a lot of energy.
— Are you the only ones in the world who have created a spike-based neuromorphic chip?
— The story began in 2014 when IBM made the TrueNorth chip, similar to a crossbar – a rectangular grid with inputs horizontally and signal outputs vertically. Artificial spiking neurons are located at the nodes of this grid. They can only do simple things: sum spikes. Such a hardware solution's circuitry allows it to be very energy-efficient.
After IBM, Intel made a similar chip called Loihi. Then the Chinese made the Tianjic chip. Now, it's us.
— What are IBM and Intel using them for?
— Research is still ongoing because this is an entirely new approach. There are both high hopes and skepticism because it's not yet entirely clear how to train and effectively operate such networks.
Neuromorphic chips are also being attempted to be built on new electronics based on memristors, which "remember" the current passing through them. Our chip is built on a traditional transistor electronic basis. For this, chip manufacturing technology of at least 28 nm is required. We don't have such technology in our country yet.
Superfast Cameras
— What devices will work on your neuromorphic chips, and what will they be able to do?
— First and foremost, our chips are excellent because they are highly energy-efficient. This means that a smart device could run on a watch battery for a long time. Without wires, without a power outlet. And, above all, these are cameras.
There are still relatively few examples worldwide of commercial applications of neuromorphic devices for ordinary users. For example, there is a motion-detecting spiking camera. This camera is lightweight and inexpensive, it can be mounted on a wall, and it registers the movements of everyone in the apartment. It can differentiate between a person and a dog, and if an elderly person falls, it sends an alert signal. The idea here is that such a camera does not transmit the full picture, preserving the privacy of the home environment. If an incident occurs, only certain object contours will be visible at the response center.
There are many more examples for industry. For instance, IBM uses such cameras in computer vision systems on aircraft: they analyze a high-resolution image at high speed and identify the objects of interest. However, this is still in the research stage.
Intel uses such cameras for scientific purposes. Projects where scientists attempt to replicate the capabilities of biological organisms are particularly interesting. For example, a European group, studying an ant's navigation, modeled an ant's spiking neural network for drone navigation. With the help of a camera, the network memorizes the path from one place to another and can then, without GPS, autonomously guide the drone back to the starting point.
