Why Christmas 1992 became the most beautiful holiday
First doubts about the so-called "Neural Networks"
As a microelectronics engineer, I worked on a prestigious project in 1991, the largest integrated circuit (IC) in Germany (or the world?) at the time. We installed 860,000 transistors with many thousands of logic gates on a silicon surface of 14x14 mm² (Fig.). Since the clock frequency of this IC was very high and fixed at 154 MHz, a key problem was to arrange the gates and components in such a way that no clock signal arrived too late, even if it had to run across the entire chip. A reduction in the clock rate would have been the end of the design.
You have to know that our entire PC world, all our digital knowledge, is based on Boolean circuits (CMOS gates) and state machines (latches, flip-flops, ROM, RAM, PLA). Here, all data flow is synchronized in hardware with latches. Their synchronous signals (clocks) are derived from a central clock. However, this clock often travels very far and takes too long to reach remote corners of the chip. The signals, in turn, travel longer and longer the smaller the structures become, the bigger the die is, and the faster the clock frequency becomes.
Simultaneity therefore becomes more and more of a problem as the degree of integration increases. Even larger and faster ICs require separate clocks for each building block.
Fig.: CMOS IC with 0.7 µm gate length, 860,000 transistors, three conductor levels (red, green, blue), 14x14 mm. The individual conductors, transistors or gates are too tiny and can no longer be recognized. Photo: G. Heinz.
Every digital IC, whether microprocessor or memory or peripheral, works in a similar way. If just one signal travels too long on long conductors, an IC no longer works correctly. An input signal must be present at the next latch until the latch receives the clock edge so that the information can be accepted.
The clock is no longer simultaneous anywhere in large, fast ICs. Simultaneity becomes impossible with increasing size, number of gates and increasing clock frequency. Other philosophies are then required. Regions of simultaneity must be decoupled from one another using asynchronous interfaces, as is common in computer networks.
In contrast, simultaneity is unthinkable in the nervous network. Neurons deliver razor-sharp spikes with geometric lengths in the millimeter or micrometer range. Latches and clocks are probably rather unknown in the nervous system. Conductivity speeds are also a million times lower than on ICs.
If we assume a very rough conductivity speed of ten million meters per second (10,000 km/s) for ICs, then for nerve fibers it is just a few centimeters to meters per second. The conduction paths of an IC therefore conduct millions of times faster - and yet the limits of conventional circuit technology were reached with this IC. Either latches at the top left or bottom right could clock correctly, but simultaneity could only be achieved with a clock introduced six times from all sides.
The question arose, as to how a network that is millions of times slower - whose simultaneity problems are millions of times more difficult to solve - can handle things that are millions of times more complex than an IC.
Even today, even with the largest computers, we are nowhere near able to control the movement of a fly in real time. This has just three hundred thousand neurons.
In addition, the size of our brain is around ten times larger in the x and y directions than this IC, and even thousands of times larger in the z direction (all gates of an IC lie in a 2-dimensional plane). If we multiply x by y and z, our cortex may be roughly 100,000 times larger than this IC.
If we assume roughly comparable conductive path thicknesses, we find in the cortex that the conductive paths are on average ten times longer and the conduction speed is 10 million times lower. But there is apparently no signal waiting for a clock pulse.
The computer science problems scale in the most unfortunate direction in all parameters towards the nervous system. Ultimately, we have a synchronization problem in the nervous system that is around 10,000,000 times 10 times 100,000 ~ 10,000 billion times harder than in our IC.
In 1992, I asked myself, how this can go. For a microelectronics engineer, the task of having to develop an IC with conductive paths that are millions of times slower, ten thousand times larger and can do things that are a thousand times more complex would be pure fantasy.
So what is the secret of neural networks? How do they work? Apparently they work completely different?
The conclusion is obvious that nature must have invented a somewhat different computer science - how else can it process data in such complex networks properly?
In 1992 there were large budgets for research into what was called "Neural Networks - NN" (now "Artificial Neural Nets" - ANN). But to our astonishment these are also clocked! ANNs are also based on the postulate of simultaneity, they are all state machines. They need clocks! ANNs cannot help us with the question of avoiding clocks.
Even worse: our entire mathematical methodology and deduction is based on i and i+1 and i+2, i.e. on states and their simultaneity measured on a time axis.
The final discovery was: simultaneity can also be created tactlessly, without clocks - through correlation of a pulse-shaped signal with itself or with its relatives: the predecessors or successors. I called the former self-interference, the latter cross interference (Link).
Unfortunately, all the sources that I checked 1992 contradicted this assumption. EEG devices had a cutoff frequency of a few kHz; they were still very slow at that time. The assumption, that pulses can be geometrically very short, was confirmed later.
So it was up to me to find some evidence that such clock-free simultaneity can actually take place somewhere in the nervous system.
By coincidence I met the neurologist Dr. Torsten Griepentrog at a good friend's birthday party on March 8, 1992. I asked him whether it was possible to observe several nerve fibers simultaneously in a non-invasive way anywhere on the body. He pointed to a suitable place for this: two nerves can be easily recorded on the wrist: the nervus radialis and the nervus medianus.
Now it was time to study literature and reflect. The ideas for clock-free nerve networks gradually took shape. Later they would lead to the renaming of clocked "Neural networks" into "Artificial Neural Networks" (ANN). And at some point I had an idea for an experiment. We agreed to a test date shortly before Christmas in Torsten's practice in Teupitz.
We met in the Teupitz clinic late in the afternoon on December 16, 1992 to try a thumb experiment. It was already getting dark. And it was snowing. As I walked across the clinic grounds, there was complete silence. Strings of lights shone everywhere - a prophecy?
The experiment was successful: depending on the position of the thumb, we saw different delays between the observed two nerve fibers. I was so overwhelmed that I was shaking with joy. I could have cried. The theoretical foundations that later led to the book "Neural Interference" were largely already fixed in sketches.
It was the best Christmas of my life.
In 1993, under time pressure, the manuscript "Neural Interferences" was written day and night in three months.
The real nerve- or neural networks, the Interference Networks (IN) were born. Later, these gave rise to the interference integrals, the interference images (I. projections and I. reconstructions) and the acoustic photography and cinematography (Acoustic Camera).
At least it was a beginning.
Read
here
more about how McCulloch/Pitt's idea changed the world in 1943 and how relative simultaneity according to Lloyd A. Jeffress went into a deep sleep for almost fifty years in 1948...
File created around 2010
Smartphone layout 2022/10/27
Some adds and Google-translation December 2024
Many thanks to the freaks at Google Translate. You made it possible to translate important pages very easy, so that they can be accessed by scientists all over the world. Thank you very much!
2023, a year ago, I had to check the meaning of every subject-specific term, but now the tool works (mostly) so well that I only have to skim the text. So I could still imagine how incredibly complicated the development of a translation program is. Today, nobody takes notice of this gigantic achievement! Continued success wishes Gerd Heinz!
Top http://www.gheinz.de/index.html
Mail to info@gheinz.de
Visitors since 2021/12/06: