Signal Theory of Intelligence

3          The emergence of intelligence and consciousness

Intelligence and consciousness are an inevitable result of the evolution of nervous systems in segmented animals. It all began with single-celled organisms. Through colony formation and the onset of differentiation, the first multicellular organisms emerged.

“The idea that the first multicellular organisms arose from colonial single-celled organisms was first systematically developed by Ernst Haeckel in his Anthropogeny (1874). His colony theory describes the emergence of the Metazoa as a consequence of increasing differentiation within such cell aggregates.”

Bibliography:

Ernst Haeckel (1874): Anthropogeny or the History of Human Development

There are several theories regarding the origin of segmented animals. The classical locomotion theory explains metamerism as an adaptation for more efficient locomotion. Modern developmental biological approaches, however, postulate a segmented proto-bilaterian as the common ancestor of the Bilateria. Other authors, such as R. B. Clark, emphasise that metamerism may have arisen independently on multiple occasions.

In addition to classical functional and developmental biological theories on the origin of segmentation, there is a neuroarchitectural approach that explains the emergence of segmented animals primarily in terms of the organisation of the nervous system. This theory was developed by Malczan (2019) and describes segmentation as an emergent structure of a modularly organised neural system, which ultimately led to the organisation of vertebrates.

References:

·  Malczan, A. (2019): Brain Theory of Vertebrates. ISBN 978-3-00-068559-0.

·  Online publication: https://www.andreas-malczan.de/Gehirntheorie_Kapitel_3.html

3.1        The transition to segmented nervous systems according to Malczan

The transition to segmented nervous systems arises from fundamental functional necessities. As soon as the body of an early metazoan was divided into repeated segments, an intensive exchange of substances and signals arose between these sections. The neural coordination required for this could only be ensured by a similarly segmented nervous system ( ), which differentiated from the unsegmented nervous system of the precursor organisms.

Even in these early forms, there was a fundamental separation between sensory and motor signals. This functional dichotomy was incorporated into the segmental structure and led to the ladder-like nervous system of segmented animals. Ascending signals from a segment arrived at the segment boundary, where interneurons formed, transmitting the excitation to neurons in the neighbouring segment. Corresponding connections also developed on the motor side.

The input neurons, interneurons and output neurons became nodes of an emerging neural network. The interneurons took on the function of intermediate neurons, so that neural networks with an intermediate layer emerged for the first time at the junctions between the segments. This intermediate layer mediated the transmission of sensory signals to the next segment and simultaneously coordinated the motor feedback. Consequently, a modular network node formed at each segment boundary, laying the foundation for the later complex network architecture of vertebrates.

Signal divergence enabled the distribution of information across multiple target neurons. Hebbian learning led to the adjustment of synaptic strengths and thus to the formation of functional weights within the network. Inhibitory interneurons created a break in symmetry in line with the mechanisms described by Oja and Sanger.

This early segmented nervous system thus already possessed all the fundamental properties that later enabled the emergence of complex neural networks and, consequently, of intelligence. For the evolving vertebrates, this architecture represented an unparalleled evolutionary advantage.”

Let us first consider a single interneuron. It receives signals from several input neurons, combines these into a scalar value, and distributes this across several output neurons of the subsequent segment according to the strength of its efferent synapses. The strength of each synapse determines how much of an incoming signal is passed on. In the language of artificial intelligence, this is referred to as ‘weights’, but biologically it is simply a matter of synaptic transmissions of varying strength.

The key feature of this convergence–divergence circuit is that the intermediate neuron can generate an output that is not fully contained within the current input. The scalar product of the input and the synaptic strengths forms a compressed representation of the input. This scalar is then distributed along the divergent outputs. If individual components are missing from the input, the learned output is still generated because the synaptic strengths compensate for the missing parts. The intermediate neuron thus completes the output – an elementary form of intelligence.

Mathematically, this structure can be described as the dyadic product of an input vector and an output vector. A single hidden neuron thus generates a rectangular matrix of synaptic weights that fully determines the transformation from input to output. This dyadic product has the remarkable property of replacing missing input components with learned patterns, thereby generating a complete output.

In a neural network, however, several hidden neurons are usually active simultaneously. Each of them generates its own dyadic product. The superposition of these products results in a dyadic sum. This sum is a linear system. Yet it is precisely here that the intelligence-generating property of the individual dyadic product is lost: the linear superposition smooths out the individual contributions of the interneurons and prevents the independent generation of missing output components. The system loses its ability to complete signal patterns.

Nature solved this problem through neural competition. Interneurons inhibited one another, and the interneuron that was most strongly excited inhibited its neighbours most strongly. This created a non-linearity that led to a break in symmetry. This break in symmetry meant that not all interneurons were active simultaneously, but only the one whose pattern best matched the input. This restored the intelligence-generating property of the individual dyadic product.

In artificial neural networks, this symmetry breaking is modelled by activation functions such as ReLU or related methods. The biological solution – lateral inhibition and neural competition – is functionally equivalent and represents the evolutionary basis for the emergence of intelligent neural systems.

Below, we will explore these insights in more depth using concrete examples