Fig.1: Konishi's drawing of Jeffress idea 1993 [1a]
Fig.2: Original circuit of Jeffress paper 1947 [1b] - historical the first inter-medial interference circuit
The circuit Fig.1 has interesting properties:
Also if Konishi discussed periodical time functions: interference circuits works as better, as more the time functions have non-sinoidal character. Nerve pulses have velocities from µm/d...m/s, time-length in the range 0.1 µs...s and wave length in the range µm...cm.
A specific detail seems that the circuit don't need different media to work. It also works between two neural fields connected by some axons.
To answer te question, why nobody has found these interesting properties over 50 years, we have to remember the history of computations to the 40's. Behind Conrad Zuse (Z1 in 1938) lots of scientists tried to develope first computers. In the beginning forties the fiction of machinery computation moved the minds. Discretisation of analog signals should be the way, to make things going in a digitized world. Electrical wires imploded to potentials, to nodes.
In the beginning fifties such 'neural' nets were qualified to learn. Delay models of wires seem to be unnecessary (Perceptron, backpropagation etc.) to understand neural networks. Perhaps Jeffress model was not of interest. Even some biologists as Konishi remembered him all times to interpret acoustical experiments.
The genius of computing and neural pioneers was, to digitize Jeffress interference circuit. A time-function was interpreted in discrete values, delay portions were called states or samples. State machines, Petri-nets and digital filters were born.
While McCulloch/Pitts started the paper [2] in 1943 with the words
"The velocity along the axon varies directly with its diameter, from less than one meter per second in thin axons, which are usually short, to more than 150 meters per second in thick axons..."
they expressed neural circuits with axons and dendrites in the same minute with state sequences instead of time functions.
What a mistakes with fatal consequences!
In the physical assumptions [2] we read:
"...
2. A certain fixed number of synapses must be excited within the period of latent addition in order to excite a neurone at any time, and this number is independent of previous activity and position on the neurone.
3. The only significant delay within the nervous system is synaptic delay.
..."
Thus, it is not possible, to find any time-function in the paper. Else we find terms, characterising the new age, the age of computer technology.
Instead of terms in form of time-functions like
f1(t) = f2(t-T1) + f3(t-T2) + ...
where T1 and T2 are (float number-) delays of real wires proportional to distances, we find integer terms in the form
N3(t) :=: .N1(t-1) .v .N2(t-3)
(Please excuse the simple way to describe formulas under HTML)
The way to describe delays as integer numbers saved all the neurocomputing science to find out neural projections for 49 years [6]! Why I found out this fact in 1992? Only because I thought about slow conduction speeds of nerve pulses in dendrites: animals have to be fast to get a chance in evolution. Whats the reason, that impulses flow so slowly?
So, in one of the origins of neuro-computing we find a much simplified description of physical reality. In sight to the development of computers (state machines) this might be a very helpful way. But it covers ideas in the direction to interference, wave fields, relativity of time. The meaning of an intrinsic delay of a neural wire was lost for years. A time function in the form f(t-T) was replaced by a state- or sample sequence N(i), N(i+1), N(i+2)... .
Forthcoming models do not note incremental delays. The 'synapse' was later recognized as the learning weight of a neuron. So papers, discussing the phenomenon of non-interpretability of bio-neural networks [4], do not find out any problem.
In this sight, Jeffress imagination was a short lightning flash in the dark.
At the other hand, the models of Nobel-price candidates Hodgkin/Huxley (1952) [5] were precise enough to discover interferential projections. But, be it, as it is, these models are over-loaded, they are to complex to recognize interference projections. They are to slow, to simulate some thousands of neurones in a time analysis manner within years.
Today it is time to recognize, that quantisations can produce mistakes. At the one hand, so show PSI-simulations, time and space properties have the very impact on the attainability of interference projections. At the other hand, the re-interpretation of biological networks only is possible in case of structural equivalence. The real structure of a neural net can be interpreted better with delays and time-functions then by state machines. We have forgotten, that a state sequence is a special, discrete form of a time-function. At last, time functions develope themselves physical properties characterizing delaying spaces. So a simple divergence of a wire that's ends converge diametric produces Youngs or Huygens double split pattern.
I did not know Jeffress or any interference system, when I realized theoretical an electrical interference circuit in optical analogy in 1992. It had the property, to produce only mirrored projections.
In opposite to other works, calculations in this homepage are only produced using time functions and free wave propagation. Following, it don't exist a potential of a wire. Instead, we have to calculate wires with distributed parameters, potentials are connected to locations (x,y,z) in space. Moves a time-function through space, the delay term T will increase proportional to the distance from the point of excitement.
A wave moving through space therefore is interpreted as a circular wave, also, if this may be a simplification for the interpretation of neural systems.
The most simple interference circuit can be built with an two input operating element (multiplier, adder...) that's output drives one input. Also a two input element, connected with both inputs to one node via different delays can be such an interference circuit. In my book (N.I., 1993), I discussed different simple interference circuits, showing very interesting properties in space-, time- and frequency domain.
A more general form we get by connecting two fields (a generating G and a detecting D) together with some axons (A). We get a structure, that may be known from anatomical drawings. Every neurone (N, N') of both fields has a connection to every end-point of the connecting axons, whereby all wires have length-proportional delays.
Viewing any impulse, outgoing from a generator neuron (G) in all directions, we remark, that all partial impulses under circumstances can meet each other on the detecting field (D) in the place, where delays over all possible ways are the same.
These simple interference circuit produces mirrored projections and images (whenever) between generating- (G) and detecting space (D). Projections of thought.
It is able, to solve the following tasks:
Resulting excitement maps are calculable on different ways: as interference integrals or as wave fields. We calculate interference integrals point by point and movies time step by time step, but with the same algorithm, the authors H-interference transformation (HIT).
Co-ordinates of each channel source or sink can be choosen freely. Generator and detector fields can have different velocities, origins, orientations and sizes. They are designed as bitmaps with variable and independent matrix- and physical dimensions. Timing, sample rate, pulses form and intervall, pulse distances can be varied.
The help-file of PSI-Tools is downloadable from the references chapter on this homepage.
Fig.: Reconstruction (top) and projection (bottom) as interference integral from the same, synthesized channel data stream. Both fields have identical properties, velocity is the same. Channels have the same delay. Because of over-conditioning (one channel to much) we find high projection quality only in the central range.
We can recognize two conditions for projections: 1) geometrical impulse length have to be smaller then structures, 2) small channel numbers supposed, the pause between following impulses has to be large against the (time domain) field size. Time- and space-units are combined via medial velocity.
Fig.: Part of an four channel interference movie (PSI-Tools): snap shot of four waves, moving over the detecting field. The channel source points are located in the corners of the field. We observe the time step short after a four-times interference.
See also the 'S'-movie in the simulation chapter.
[1a] Konishi, M.: Die Schallortung der Schleiereule. Spektrum der Wissenschaft, Juni 1993, S. 58 ff.
[1b] Jeffress, L.A.: A place theory of sound localization. Journ. Comparative Physiol. Psychol., 41, (1948), pp.35–39
[2] McCulloch, W.S.; Pitts, W.: A logical calculus of the ideas immanent in nervous activity. Bulletin of Mathematical Biophysics 5:115-133
[3] Widrow, B., Hoff, M.E. Adaptive switching circuits. 1960 IRE WESCON Convention Record, New York: IRE, pp. 96-104
[4] Rumelhart, D.E., McClelland, J.L.: A Distributed Model of Human Learning and Memory. in: Parallel Distributed Processing. Bradford/MIT Press Cambridge, Massachusetts, vol. 2, eighth printing 1988.
[5] Hodgkin, A.L., Huxley, A.F.: A Quantitative Description of Membrane Current and Its Application to Conduction and Excitation in Nerve. Journ. Physiology, London, 117 (1952) pp. 500-544
[6] my thumb-experiment see http://www.gfai.de/www_open/perspg/g_heinz/intro/iwk_ilm.htm
© Copyright: G. Heinz, file created sept. 30, 1995
access no. since last counter reset