Cariani P. (1997) Emergence of new signal-primitives in neural systems. Intellectica 25: 95–143. Fulltext at https://cepa.info/4361
Emergence of new signal-primitives in neural systems.
Intellectica 25: 95–143.
Fulltext at https://cepa.info/4361
Emergence is the process by which new structures and functions come into being. There are two fundamental, but complementary, conceptions of emergence: combinatoric emergence, wherein novelty arises by new combinations of pre-existing elements, and creative emergence, wherein novelty arises by de novo creation of new kinds of elements. Combinatoric emergence is exemplified by new strings constructed from existing alphabetic letters, whereas creative emergence is exemplified by the addition of new kinds of letters to an alphabet. The two conceptions are complementary, providing two modes for describing and understanding change: as the unfolding consequences of a fixed set of rules or as new processes and interactions that come into play over time. Within an observer-centered, operational framework, the two kinds of emergent novelty can be distinguished by what an external observer must do in order to successfully predict the behavior of an evolving system. Combinatoric and creative emergence can be operationally distinguished by changes in apparent effective dimensionality. Whenever a new independent observable is added to a model, its dimensionality increases by one. A system that only recombines requires no new observables, and does not expand in effective dimension. In contrast, a system that creates new primitives requires new observables for its description, such that its apparent dimensionality increases over time. Dimensional analysis can be applied to signaling systems. Signals have two basic functional properties: signal-type (category, variable, type) and signal-value (state, value, token). These properties can be conveyed by a variety of means: by the signal’s physical channel, by the internal form of the signal (waveform, Fourier spectrum), by its time of arrival, and by its magnitude (average power). Neural coding schemes can similarly be based on which neurons fire, which temporal patterns of spikes are produced, when volleys of spikes arrive, or how many spikes are produced. Traditional connectionist networks are discussed in terms of their assumptions about signal-roles and neural codes. For the most part, connectionist networks are conceptualized in terms of new linkage combinations rather than in terms of new types of signals being created. Neural networks that increase their effective dimensionalities can be envisioned. Some kinds of neural codes, such as temporal pattern and time-of-arrival codes, permit encoding and transmission of multidimensional information by the same elements (multiplexing). We outline how synchronous time-division and asynchronous code-division multiplexing might be realized in neural pulse codes. Multidimensional temporal codes permit different kinds of information to be encoded in different time patterns. Broadcast-based coordination strategies that obviate the need for precise, specified point-to-point connections are then made possible. In such systems new signal types arise from temporal interactions between time-coded signals, without necessarily forming new connections. Pitches of complex tones are given as examples of temporally-coded, emergent Gestalts that can be seen either as the sums of constituent micro-patterns (combinatoric emergence) or as the creation of new ones. Within these temporally-coded systems, interacting sets of neural assemblies might ramify existing, circulating signals to construct new kinds of signal primitives in an apparently open-ended manner.