Red Line of ChatGPT

Boris Kravtsov, PhD
6 min readSep 7, 2023

--

Photo by Mariia Shalabaieva on Unsplash

Introduction

Task. Two kinds of geometrical figures are placed on the plane randomly. When moving them, make sure that there are only figures of one kind on each side of the red line.

Any five-year-old child can solve the “red line problem”, even if it is the first time that he sees the figures in the picture. But the newfangled ChatGPT technology, alas, is unable to do so. Devoid of creativity, it cannot come up with a method of comparing geometric figures of different kinds and after significant transformations. Creativity is a hallmark of intelligence, and ChatGPT technology is simply a complex and often an uncontrolled compilation of already known facts (to the author, it reminds him of his old buddy who likes to tell jokes, but unfortunately, the latter always has old jokes).

Enormous computing resources and volumes of information — which ChatGPT operates with — do not save the situation either. The accumulation of data and the generation of new ideas based on them are fundamentally different things. We should also note that even excessive information can be harmful. Einstein, for example, avoided reading scientific journals while thinking about his next article.

Creating Artificial Intelligence

Of all the manifestations of human intelligence, we are interested only in its ability to recognize. “Creating artificial intelligence,” in this case means modeling the work of the occipital part of the brain: its visual cortex.

There are two directions, which we will conditionally refer to as the “Programmers’ Approach” and the “Physicists’ Approach”:

1. “The Programmers’ Approach” and deep learning technology. To recognize cats, for example, we are suggested to start with setting up or training the system, i.e., loading hundreds of thousands of cat photos into it, and then… Stop, enough! It doesn’t matter what happens next since our visual cortex doesn’t need all this to accomplish the task; it can compare objects it sees for the first time. In other words, the programmers’ approach, which requires pre-tuning of the system, is inadequate for the task of modeling and should be immediately excluded. Catchy terminology like “training of neural networks” essentially changes nothing, as there is no relation to real connections of brain neurons with such an approach. “War is too serious a matter to be entrusted to generals.”

2. “The Physicists’ Approach.” Physicists start cognition of any natural phenomenon with observations, experiments, and formulation of hypotheses. This is how a physical model is created (let us pause here and sympathize with poor Galileo, who had to climb the three hundred steps of the Tower of Pisa many times until he concluded that “all bodies fall equally”). In the future, the model will be clarified, or it will be rejected entirely. Physicists, unlike programmers, strictly monitor the adequacy of their models to all known experimental facts.

Physicists view the visual cortex as another natural phenomenon. This approach seems natural and logical to the author, a physicist by specialty. Laboratory experiments with the brain, particularly with its visual cortex, have become commonplace, with thousands of publications devoted to model building. We are particularly interested in those starting with Campbell and Robson, 1968 Pollen and Lee, 1971 Maffei and Fiorentiny, 1973; Glezer et al., 1989, who consider the visual cortex as a frequency analyzer. It is well known that almost all image processing operations can be performed in the image domain or frequency domain. The authors assume that the visual cortex has chosen the second path. By consistently adhering to this hypothesis, we are trying to track what practical results this can lead to (see publications below).

Amazing AI Tables — Feb 16, 2023

1. In the frequency domain, the concept of similarity of geometric figures is introduced, and its effectiveness is shown on sizeable statistical material.

2. Using similarity, we successfully solved the “red line problem,” a complicated version of which we called “AI Tables.”

3. Handwritten number recognition demonstrated without a pre-training procedure.

Image Processing in the Visual Cortex — Feb 18, 2022

The transition to the frequency domain made it possible to explain the well-known optical illusions (Müller-Lyer illusion, Vertical-horizontal illusion).

New High-Quality Edge Detector — Feb 2, 2022

One of the most notable achievements in experimental brain research (Nobel Prize 1981) is the discovery of orientation-selective neurons. In a very simplified form, it is about the following: the authors of the discovery, D. Hubel and T. Wiesel, showed the experimental animal (a domestic cat) slides with single lines depicted on them. When such a line appeared on the screen, the cat showed activity of one of the cortical neurons. This activity depended on the angle of inclination of the line, and it was possible to establish the single direction at which the neuron’s activity reached its maximum.

Further, by changing slides and imitating the slow rotation of the stimulus line, the scientists observed how the decrease in the activity of this neuron was replaced by the appearance and growth of the action of another (second) neuron. The decrease in the activity of the second neuron was followed by the appearance and increase in the activity of the third neuron, etc. The activity maxima of cortical neurons are observed with a periodicity of 12 degrees: 0, 12, 24, and so on. Although more than 60 years have passed since the discovery of orientation-selective neurons, their role in human and animal vision remains unclear. This article attempts to answer that question.

In the language of image processing experts, D. Hübel and T. Wiesel found a built-in set of edge detectors in the visual cortex. Therefore, let’s look at one of them: the Sobel Edge Detector (Sobel’s filter with Gy kernel). The figure below shows the result of filtering.

As we can see, this filter detected horizontal lines and edges in the image, but as it moves away from the horizontal, its sensitivity decreased. As a result, it could not detect vertical lines (see the right side of the figure). However, there is a second type of Sobel filter with Gx kernel, which is just “stitched” to detect vertical lines. But accordingly, it will not be able to detect horizontal edges and lines of the image. Here, it is recommended to use both kinds of filter together — as the sensitivity (activity) of one of them decreases, the sensitivity (activity) of the other one increases. Let us again recall the poor cat to note the complete similarity in the behavior of Sobel filters and the way a pair of “neighboring” orientation-selective cortical neurons behaves.

We have shown how a transition to frequency domain can change the direction of maximum sensitivity of the Sobel filter. Now this direction can be any angle — not only 0 degrees (Gy kernel) or 90 degrees (Gx kernel). This allows us to model any orientation-selective neuron and see the result of the collective action of all orientation-selective cortical neurons. For this purpose, we summarize the performance of the entire set of edge detectors (Sobel filters) tuned to the following directions: 0, 12, 24, etc. degrees and finally obtain a high-quality contour image.

Finding contours is the first step of all known methods of recognition. As we can see, the role of orientation-selective neurons is to create a contour image of the field of view.

Experimental confirmation of this position and finding out how and where the contours are stored in the visual cortex will be a serious bid for the next Nobel Prize.

--

--

Boris Kravtsov, PhD

I'm trying to share some of my old thoughts and new perspectives.