top

FAQs about Neural Networks and Interference Networks


Q: You are talking about nerve nets, why didn't you talk about Neural Networks?

A: In the first time (1993...1997) I tried to do so. Permanent misunderstandings demanded a different name. The terminus "Neural Net" or Neuronal Network (NN) was occopied and burned by a different behaviour. So called "Neural Networks" are digital, technical constructions, requiring a clock synchronization.

Nerve nets show a totally different behaviour, because they have no clock. After some years, the NN-community changed the name to "Artificial Neural Networks". But within the last ten years this knowledge was lost again. Now we find again the wrong terminus "Neuronal Network" in peer reviewed papers.


Q: What is the problem with Neural Nets?

A: The greatest challenge is, that any smallest change of the delay structure of a net destroys its function or behaviour completely (think about FIR-filters). Most of neural nets were 1993 "artificial", delays played no rule in the pioneering works. With a mismatched delay structure a model has nothing to do with the origin.

So lots of interesting works disappeared in the gigantic mass of ANN-papers, ignoring the delay structures. If you go into detail, you will find the remark "delays win over weights" in interference networks in my pages. Like optical projections, clockless networks can only address any information via their delay structure.


Q: What was the reason to introduce the name "Interference Network"?

A: An average pyramidal neuron in human cortex has 7400 synapses (input terminals), and our brain has 40 billion pyramidal neurons (the numbers vary in different publications). The permanent state for a living nerve cell is, that all thousands synapses get unsynchronous fire, if a neuron sleeps.

To bring a neuron to excitement (spike delivery), we need clean delays and synchronous arrival at thousands of synapses. Spikes are very short - partially in the submillisecond range, while smallest nerve fibres are very, very slow.

The construction of such circuits and the finding of conditions, restrictions and cases opened the field of interference network research.

If a neuron in nerve system gets permanent fire from all directions, the situation is comparable with radio broadcast interferences until the 60th of the last century. Driving over land on Medium Wave (MW) the people heared sometimes only noises, whisper or whistling in the car-radio.

Two places in a nerve net (for example a position of skin with a place in brain) can only communicate, if they overlay the "interferencial noise", coming from other communications. How to do this?

Thinking about synchronous fire and noise, we find optic-like communication principles (projections). We have to deal with cross- and self-interferences. Interference network research will help to understand such processes better. And we need to understand interference networks to call the right questions to nerve systems.


Q: What have Interference Networks to do with Acoustic Cameras, with Radar, Sonar or optical lense systems?

A: Experts in optics speak a very different language compared with experts for Radar, Sonar or acoustics. They all can not understand the neuro-surgeon. But the knowledge of all fields is the same. They all talk about probability waves, interference integrals and interference circuits.

The Acoustic Camera was the first and simplest demonstration how interference nets work, how to transform time series into images and vice versa.


Q: Wave field theories are Fourier-driven. Why can't I find Fourier approaches in your papers?

A: To overlay two waves in space-time, in the one-dimensional case we use convolution or cross-correlation. If we have very long time functions (lets say with thousands of samples per picture) the calculation is very time consuming. So for very long series, a shorter way is, to transform the time series into frequency series using the Fourier transformation (FFT).

Finally you have to add the Fourier coefficients instead to add the samples. The history of field theory is married with Fourier-transformation.

And be it as it is - with Fourier analysis it is nearly impossible, to avoid strong delay mistakes. A main problem of the FFT-approach are delays longer a single wavelenght of the interesting frequency, for example if the sound source comes from far left.

A complex phasor turns up to 360 degrees, then it starts with zero. If a single delay is higher then 360°, you need wave numbers to keep the correct delays. Otherwize you get additional aliasing figures (side-beams) in your reconstruction.

In case of interference networks correct delays play the most important rule. For nerves, the spike-like nature of time functions is the worst case for Fourier analysis (the spectrum of a spike is white noise). And the first software was called "Bio-Interface". Our hope was to get smart spikes from EEG or ECoG. But it should not happen, so we changed the sensors. Instead of electrodes we used microphones. So first acoustical images were made with the EEG-software.

Last not least practical aims inspired me, to avoid the frequency domain and to stay consequently in time domain. If you calculate high speed movies, the sample pieces for each picture are very short, the calculation in time domain is here faster then in frequency domain.

Things were complicate enough. So we started with the simplest case. It is, as Albert Einstein said: "Any intelligent fool can make things bigger, more complex, and more violent!"

Thanks to the Daniel Herfert group at GFaI for this short interview (2015?).



Impressum

Smartphone-layout enabled 2022/10/27
Top http://www.gheinz.de/index.html
Mail to info@gheinz.de
Visitors since 2021/12/06:
All rights reserved, © Copyrights Gerd Heinz