Information Technology Reference
In-Depth Information
7.4 A Cognitive Architecture that Supports Non-conscious
Creativity
Thus, the overarching proposal examined in the IDyOT approach is that the mecha-
nismunderlying spontaneous creativity is information-theoretically regulated predic-
tion from models built from observation of perceived input, that encode its observed
likelihood. It is important to acknowledge that there are aspects of cognition that are
not included in this model, such as affect; and also that there are aspects of mecha-
nism that are not modelled: for example, it is known that different brain components
have different learning rates, but this is not the case in IDyOT. Rather than trying
to emulate a whole mind/brain, our aim here is to push our proposal, at this purely
functional level, as far as it will go, by using it to attempt to explain as many estab-
lished phenomena as possible. Understanding where the limits are will inform us as
to what other mechanisms are needed and where and how they fit.
In this section, we unpack this dense proposal and consider the ancillary mech-
anisms required to support it. In the following section, we propose and explain the
initial testable predictions that we make of it.
7.4.1 Overview
IDyOT is a direct implementation of the Global Workspace Theory as described by
Baars [ 1 ]. A large 6 number of generators sample from a complex statistical model
of sequences, performing Markovian prediction from context [ 25 , Chap. 9]. Each
generator (which may be thought of as a simple AI agent [ 24 ]) maintains a buffer
of perceptual input, which may include mis-perceptions due to the possibility of
multiple predictions matching ambiguous input. Buffered sequences are flushed into
the Global Workspace (which may be thought of as an AI blackboard [ 9 ]) when an
information-theoretic throttling condition related to a proposal by Wiggins [ 55 ]is
met. This mechanism solves a problem within Global Workspace Theory known as
the Threshold Paradox [ 1 , 55 ]. A block diagram of the system is shown in Fig. 7.1 .
The diagram illustrates the cyclic (and hence dynamic) nature of the model. The
generators sample from statistical memory, synchronised 7 by the perceptual input,
if some exists, that it receives. If there is no input, the generators freewheel (evi-
dence for the neuroscientific validity of this position is given by Fink et al. [ 13 ]),
conditioned only by prior context, and this is where creativity is admitted; however,
for the moment, we focus on the perceptual input, for it is this that fundamentally
drives the system. Perceptual input is matched against generators' predictions, and
6 The number is not specified in Baars' theory. In IDyOT, the number of generators has a direct
bearing on the (statistical) prediction quality: as the number of generators increases, so does the
likelihood of correct predictions.
7 The initial version of IDyOT has an abstract, symbolic representation of time; however, more
developed versions will predict the real-world timing of perceptual input, as well as its content.
Search WWH ::




Custom Search