“You don’t see something
until you have the right metaphor to let you
perceive it”
Robert
Stetson Shaw
The study of higher
cognitive functions includes the study of intelligent systems which exhibit a
peculiar behavior: choosing courses of action that are relevant to achieving
their goals. Cognitive science, defined as the study of intelligence and its
computational processes, has been interested in the study of higher cognitive
functions for more than a century. Since at least 1950, a branch of computer
science called “artificial intelligence” has been studying the intelligence
exhibited by machines, thereby implicitly bridging the gap between computations
and behaviors exhibited by biological and non-biological systems. Although not
everyone accepts the interchangeable usage of intelligence as applied both to
animals and computers, the approach of this work follows the general framework
of regarding intelligence as partially independent of the nature of the
physical system in which it is implemented. This choice allows the experimenter
interested in biological intelligence to simulate the system target of its
study, therefore achieving a level of understanding which is beyond the
capability of traditional, non-synthetic approaches. The objective of this work
is to develop a computational model of higher cognitive functions using the
connectionist paradigm. The psychological and neuropsychological literature is
rich in models of higher cognitive functions and goal-oriented behavior
(Posner, 1989), but a lack of explicit, mathematically defined, and
biologically-plausible models is an obstacle both to the understanding of the
underlying processes and to the maturation of cognitive science as a “hard”
discipline. In this thesis, an emphasis is placed on the development of a biologically-inspired
model of higher cognitive functions. In particular, great attention is devoted
to a subset of neurophysiological and anatomical data which have pointed, in
the recent past, to the involvement of specific cortical and subcortical areas,
as well as neurotransmitters, in the control of higher cognitive functions and
goal oriented behavior. In particular, the role of the neurotransmitter
dopamine (DA) in the maintenance of short-term memory in prefrontal cortex
(Pfc) neurons has been a subject of interest for the physiological,
neuropsychological, neuropathological, as well as for the modeling community
(Lidow et al., 1998; Goldman-Rakic et al., 2000). Such
a convergence of interests is justified by the importance of Pfc and DA
functioning in various high-level tasks, which are usually grouped under the
broader category of “cognitive control”, as well as by the fact that these
tasks are the ones more often impaired in various neurological and
psychopathological conditions (schizophrenia and Parkinson disease, among
others). A crucial role in the physiological functioning of Pfc is played by
DA. Numerous human and animal studies have demonstrated that the integrity of
the main DAergic pathway (Ventral Tegmental Area - VTA - to Pfc) is crucial for
the normal functioning of Pfc (Goldman-Rakic et al., 2000). The first part of
this thesis reviews some of the main theories and models regarding the presumed
role of DA in Pfc functioning and cognitive control. Tasks of different
complexities are simulated in the attempt to understand how a
biologically-plausible model can control the complex interactions typical of
the behavior which require cognitive control. The main emphasis of the work is
placed in how these properties can spontaneously arise in a system that is not
designed “ad-hoc” to cope with the task. In order to fit this requirement, the
model is developed step by step, with increasingly difficult tasks and with a
parallel growth of model complexity. This process allows the reader to
understand the need of progressive degree of complexity that the model should
incorporate in order to cope with progressively increasing “cognitive
functions”, and to solve otherwise unsolvable paradoxes that arises when a
simple model is stretched trying to cope with complicated behavioral problems.
An analogy is drawn between the rather abstract concept of cognitive control
and simpler, well-known classical conditioning paradigms, which are suggested
to be the prototype of these higher, articulated function. The biological
grounds of the model are then presented, and a surprisingly large dataset of
known anatomical and physiological data is shown to fit with the model, despite
the fact that the design of the system was guided by theoretical reasons, not
by the primary target of fitting neurobiological data. The model presented in
the first part of this work will show how a field of Pfc neurons can store a
given Short Term Memory (STM) representation and at the same time face the
presentation of several distractors interleaved between the to-be-stored cue
and the release of the action. The first set of simulations, however, leaves
open the question of how does the organism learn what is relevant in
obtaining the reward, namely how to selectively attend to the relevant stimuli
linked to the reward and to appropriately time the release of the behavior
which is functional to its obtainment. Furthermore, how does the brain
solve the paradox of being unable to store Pfc representations unless these are
followed by a DAergic response, and at the same time learn contingencies
between stimuli that were presented (and probably decayed) long before
the reward was generated? The model developed in the second part addresses
the above issues. In particular, the focus is on how an organism can learn a)
the delayed consequences of its behavior and b) to adaptively time the
release of its behavior in order to obtain a reward. These two properties of a
system are the foundations of what in cognitive science is known as of
“cognitive control”, and are features of higher cognitive functions.
Summarizing, this work is an attempt to attack the
problem of how the functions described in a) and b) can autonomously develop as
an emergent property of a neural network model. The system developed in this
work is designed to cope with a simplified task that embodies a delay between a
stimulus, a reward and a response and a delay between a cue and the release of
the appropriate action. It is shown how the performance of an apparently simple
task, carried out in a apparently effortless fashion even by non-primates,
requires an unexpectedly complex neural machinery. The work will show how the
interplay of computational, behavioral, and neurophysiological constraints can
lead to the development of a coherent and ecologically plausible model of how
higher cognitive functions can develop from a biologically-plausible neural
model. These characteristics of the model are a truly unique feature of this
work, and differentiate it with respect to the major models published so far.