Anatoly Guin: Hello, Boris. We continue recording our dialogues on the development of new artificial‑intelligence technologies. Our third series is devoted to the so‑called Progress Associative Neuromorphic Comparator (PANC™). It is an entirely new AI technology.
DeepSeek provides a definition of a neuromorphic comparator, but it contains only generic wording. Many different comparators exist in engineering, yet, as far as we know, an AI‑based comparator as a method of recognition has not been done by anyone except us. We have invented and validated a new technology that is fundamentally different from artificial neural networks (ANNs). So, Boris, the first question for you: why is our technology, which we call PANC™, not a neural network?
Boris Zlotin: Neural networks appeared more than 70 years ago and have since seemed the only way to recognize objects in a manner comparable to the brain of a living creature. Why is this important? Because, in the 1950s, the eminent cognitive psychologist James Gibson showed in The Perception of the Visual World that all mental operations—and therefore intelligence—are primarily reduced to recognition; it is the foundation of intellect.
How does a neural network recognize? In principle, it is a very beautiful, clever idea. A filter is built, much like a water filter, but it is a numerical filter that discards all random “noise” and captures patterns.
A wonderful idea, but there is one problem. Constructing this filter is astronomically costly and complex, requiring quintillions of mathematical operations. For a large, serious network the number of computational operations can exceed 10⁴⁰. Can you imagine such a number? Consequently, this “mathematical filter construction” is not only very expensive but also slow. Moreover, because of the sheer complexity and volume of operations, the probability of failures, errors, and inaccuracies is high—just a slight voltage spike, and everything breaks apart. In short, humanity accepted high costs and risks.
When we encountered neural networks during our technical consultations I was terribly dissatisfied with them at first. This was precisely because of their propensity for errors and the huge “price for benefit.” Then one of my students proposed a solution to sharply reduce time and cost. In practice, it worked. The problem was that he also didn’t know why it worked; there were no clear explanations. It was a serendipitous, hack‑like solution, which prompted us to seriously investigate how humans recognize.
The fact is that our brain cannot perform quintillions—or even millions—of calculations. The brain’s processing speed is millions of times slower than that of a micro‑chip. We concluded that the brain uses a different recognition method, not a numerical filter as in modern artificial neural networks. Yes, neuronal networks exist in the brain, but they do not build filters; they operate differently. The question was how, and how to implement it technically. After 11 years of work we discovered a very simple technical realization.
I look at you and think: even if you put on a wig, take off your glasses, shave your beard and grow a mustache, I will still recognize you. Why? Because I perceive you as a holistic image. In the 1920s–30s Gestalt psychology “Gestalt” meant a whole form, and the human mind perceives in that way. My mind contains a rich library of people. When I look at you, an internal search engine runs, looking for something similar. It will never find a 100 % match—today you look slightly different from yesterday—but it may rate one photo as 93 % similar, another as 85 %, a third as 20 %, and a completely different person as 0.2 %.
Anatoly Guin: Boris, an essential detail: PANC™ does not sequentially scan millions of stored images; it accesses the entire library simultaneously. It queries a massive set of memorized images at once—this is crucial.
Boris Zlotin: That is our pride. The idea of comparison is simple, yet we have built a comparator—a device for computerized comparison—that processes all library images in parallel, just as the human brain does. Analogy: human memory contains many “knowledge fragments,” each with “sticky ends.” New “fragments” travel along “knowledge channels” and adhere only to matching “sticky ends”, much like an atom forming a valence bond with a molecule. When a “fragment” sticks, it sends a signal: “I’m here, but this object matches me only to a certain percentage.” Consequently, the first images with 80–90 % similarity will be photographs of Anatoly Guin; somebody of a similar complexion may hit a score of 75 %. At 0.2 % you might get an octopus—some superficial similarity, but little else.
It’s important to note that our comparator requires no training. We discarded the billions of quadrillion operations needed to build a filter. Instead, the computer holds a library of images, just as my brain does (I have been assembling mine for nearly 80 years). When I see something, I compare it with my library—that is the core of our achievement. First, we replace network training with a static image library. Second, we achieve ultra‑fast parallel comparison of a single “knowledge fragment” with millions, just as in any human mind.
By the way, our software can compare any number of images simultaneously. The most important part is that recognition occurs far faster than human perception. If you appear with a mustache and wig, I may need a moment to recognize you; a machine can launch, say, 100 parallel recognition threads and obtain the result instantly.
Anatoly Guin: How does this work? Our comparator uses a different kind of artificial neuron. In classic ANNs, no single neuron stores an entire image; the whole network collectively memorizes images by adjusting each neuron’s weight. All weights are interdependent, so tweaking one requires adjusting all others, demanding massive computation. Consequently, training must be performed on the entire dataset, and any modification after training requires costly retraining.
In our system, each neuron stores a single image, independent of other neurons. We can modify, delete, or add images at will, providing operational flexibility absent in classic networks. Post-training, re-training, recognizing by class, by cluster, and so on.
When I visited Boris in the United States in 2014, I knew nothing about neural networks. I asked whether a basic model existed that worked reliably. Boris described the simplest network—a perceptron. I realized that if a simple perceptron works, then PANN™ (Progress Artificial Neural Network) should also work, because the concept is as simple as a wheel. Once you understand it, without sparing any time, you discover that it is indeed very simple. Moreover, if, to this day, deep‑learning networks (of which there are already hundreds of architectures) lack an adequate mathematical model that precisely describes how they function—because the task is overly complex—then we can construct a very simple mathematical model. In fact, that model is already known and has been built.
At that stage it was not yet a comparator; it was a new type of neural network: the Progress Artificial Neural Network (PANN™), featuring very fast training and other unique qualities. Several more years had to pass before we progressed from PANN™ to PANC™, from an artificial neural network to a comparator.
Boris Zlotin: I will add a few points.
First. There exist theories and extremely intricate mathematical models of conventional neural networks, but these theories are of limited practical use; they do not allow prediction of the behavior of a specific network. In practice, networks are developed mainly by trial and error. This is one reason for their high cost and sub‑optimality.
Second. ANNs are specialized: one network performs one task, another a different task, a third yet another, and so on. Each is expensive and must be trained separately. By contrast, PANC™ is universal and works with universal libraries, enabling many functions to be performed on a single system, simultaneously.
Example. On my computer I have a library of photographs of fifteen people, eleven images each, covering various facial expressions and lighting conditions. For a neural network it is very good if, with, say, 95 % confidence, it recognizes “this is Pete, this is John, this is Mary…”. Our system, however, can also say: “this is Pete—he is a white, middle-aged man, with a mustache and glasses.” In principle, PANC™ can deliver recognition across ten, a hundred, or more parameters. You simply supply a person’s photograph and it returns a full dossier.
Third. A conventional neural network examines an image and says, “This is a cat, not a dog,” providing a single most‑preferred match. It does not indicate how accurate that match is or how large the differences are—perhaps the cat is almost indistinguishable from a dog, like a cheetah, for instance.
PANC™ returns the entire set of analogues, ranked by similarity. Thus, I can see that here is Anatoly Guin, here is his lookalike, and over there are unknown Johns, Michaels or Jacobs who resemble Guin because they also have a bald head and an elongated face.
Fourth: Because there is no clear theory, ANN operation is completely opaque, leading to serious problems. For example, a very expensive, sophisticated network was built to recognize a wide variety of animals and performed excellently—until it turned out that it was classifying snow or ice as a polar bear, even when no bear was present in the picture.
The error is, of course, fixable, but it requires a costly, full retraining of the entire system, even though all other animals are recognized perfectly. Moreover, it is uncertain whether the new training will degrade recognition of, say, white hares.
A similar “background‑instead‑of‑object” error could occur in PANC™, but our system is transparent—errors are easy to locate and correct.
The finer points aside, the essential message is that PANC™ is not a neural network; it employs a different recognition principle. This does not imply that neural networks are fundamentally flawed or unusable. There will continue to be many tasks that require conventional neural networks. Conversely, a huge number of problems can be solved more simply with a comparative method.
TRIZ stresses the important principle of hybridization. When two (or more) alternative systems exist—e.g., a piston engine and a jet engine for an aircraft—the next logical step is to combine them. Today such hybrids are turboprop and turbofan engines.
I believe the next major stage in software development is hybridization with PANC™. Powerful ANNs and other software products (operating systems, the Internet, cloud platforms, various applications, etc.) already exist and work well. Adding modest PANC™ modules to them will make them perform far better, because PANC™ is, roughly speaking, an “intelligence block” that can be grafted onto existing software. For instance, you could take an accounting package and give it a brain.
Anatoly Guin: Very well, Boris. Shall we conclude this dialogue here? In the next session we’ll discuss the energy efficiency of PANC™—a topic that is both fascinating and timely.