Abstract
All self-active living beings need to solve the motivational problem—the question of what to do at any moment of their life. For humans and non-human animals at least two distinct layers of motivational drives are known, the primary needs for survival and the emotional drives leading to a wide range of sophisticated strategies, such as explorative learning and socializing. Part of the emotional layer of drives has universal facets, being beneficial in an extended range of environmental settings. Emotions are triggered in the brain by the release of neuromodulators, which are, at the same time, are the agents for meta-learning. This intrinsic relation between emotions, meta-learning and universal action strategies suggests a central importance for emotional control for the design of artificial intelligences and synthetic cognitive systems. An implementation of this concept is proposed in terms of a dense and homogeneous associative network (dHan).
TopMotivations
In order to elucidate the general functional purposes of emotions we start by considering the motivational problem of self-determined living creatures, whether biological or artificial. We use here and in the following the general term `cognitive system’ for such an autonomous and self-determined being. The question then regards the general motivational drives for cognitive systems.
Key Terms in this Chapter
Diffusive Control: Diffusive control is intrinsically related for biological cognitive systems to the release of neuromodulators. Neuromodulators are generally released in the inter-neural medium, from where they physically diffuse, affecting a large ensemble of surrounding neurons. The neuromodulators do not affect directly the cognitive information processing, viz the dynamical state of individual neurons. They act as the prime agents for transmitting extended signals for meta learning. Diffusive control signals come in two versions, neutral and emotional. (A) Neutral diffusive control is automatically activated when certain conditions are present in the cognitive system, irrespectively of the frequency and the level of past activations of the diffusive control. (B) Emotional diffusive control has a preset preferred level of activation frequency and strength. Deviation of the preset activity-level results in negative reinforcement signals, viz the system feels `uneasy’ or `uncomfortable’.
Motivational Problem: Biological cognitive systems are `autonomous’, viz they decided by themselves what to do. Highly developed cognitive systems, like the one of mammals, regularly respond to sensory stimuli and information but are generally not driven by the incoming sensory information, i.e. the sensory information does not force them to any specific action. The motivational problem then deals with the central issue of how a highly developed cognitive system selects its actions and targets. This is the domain of instincts and emotions, even for humans. Note, that rational selection of a primary target is impossible, rational and logical reasoning being useful only for the pursue of primary targets set by the underlying emotional network. Most traditional research in artificial intelligence disregards the motivational problem, assuming internal primary goal selection is non-essential and that explicit primary target selection by supervising humans is both convenient and sufficient.
Autonomous Cognitive System: Cognitive systems are generally autonomous, i.e. self-determined, setting their own goals. This implies that they are not driven, under normal circumstances, by external sensory signals. I.e. an autonomous cognitive system is not forced to perform a specific action by a given sensory stimuli. Autonomy does not exclude the possibility to acquire information from external teachers, given that internal mechanisms allow an autonomous cognitive system to decide whether or not to focus attention on external teaching signals. In terms of a living dynamical system an autonomous cognitive system possesses a non-trivial and self-sustained dynamics, viz an ongoing autonomous dynamical activity.
Living Dynamical System: A living dynamical system is a dynamical system containing a set of variables denoted `survival variables’. The system is defined to be living as long as the value of these variables remain inside a certain preset range and defined to be dead otherwise. Cognitive systems are instances of living dynamical systems and the survival variables correspond for the case of a biological cognitive system to the heart frequency, the blood pressure, the blood sugar level and so on.
Universal Cognitive System: Simple cognitive systems are mostly ruled by preset stimuli-reaction rules. E.g. an earthworm will automatically try to meander towards darkness. Universal principles, i.e. algorithms applicable to a wide range of different environmental settings, become however predominant in highly developed cognitive systems. We humans, to give an example, are constantly, and most of the time unconsciously trying to predict the outcome of actions and movements taking place in the world around us, even if these outcomes are not directly relevant for our intentions at the given time, allowing us to extract regularities in the observed processes for possible later use. Technically this attitude corresponds to a time-series prediction-task which is quite universal in its applicability. We use it, e.g., to obtain unconsciously knowledge on the ways a soccer ball rolls and flies as well as to extract from the sentences we listen-to the underlying grammatical rules of our mother-tongue.
Complex System Theory: Complex system theory deals with `complex’ dynamical systems, viz with dynamical systems containing a very large number of interacting dynamical variables. Preeminent examples of complex systems are the gen-regulation network at basis of all living, self-organizing phase transitions in physics like superconductivity and magnetism, and cognitive systems, the later being the most sophisticated and probably also the least understood of all complex dynamical systems.
Dynamical System: A dynamical system is a set of variables together with a set of rules determining the time-development of theses variables. The time might be either discrete, viz 1,2,3,... or continuous. In the latter case the dynamical system is governed by a set of differential equations. Dynamical system theory is at the heart of all natural laws, famous examples being Newton’s law of classical mechanics, the Schrödinger equation of quantum mechanics and Einstein’s geometric theory of gravity, general relativity.
Biologically Inspired Cognitive System: In principle one may attempt to develop artificial cognitive systems starting with an empty blueprint. Biological cognitive systems are at present however the only existing real-world autonomous cognitive systems we know of, and it makes sense to make good use of the general insights obtained by neurobiology for the outline of cognitive system theory. An example such a paradigmal insight is the importance of competitive neural dynamics, viz of neural ensembles competing with each other trying to form winning coalitions of brain regions, suppressing transiently the activity of other neural ensembles. Another example is the intrinsic connection between diffusive emotional control and learning mechanisms involving reinforcement signals.
Meta Learning: Meta learning and `homeostatic self-regulation’ are closely related. Both are needed for the long-term stability of the cognitive system, regulating internal thresholds, learning-rates, attention fields and so on. They do not affect directly the primary cognitive information processing, e.g. they do not change directly the firing state of individual neurons, nor do they affect the primary learning, i.e. changes of synaptic strengths. The regulation of the sensibility of the synaptic plasticities with respect to the pre- and to the post-synaptic firing state is, on the other hand, a prime task for both meta learning and homeostatic self-regulation. Homeostatic self-regulation is local, always active and present, irrespectively of any global signal. Meta learning is, on the other hand, triggered by global signals, the diffusive control signals, generated by the cognitive system itself through distinct sub-components.
Cognitive System: A cognitive system is an abstract identity, consisting of the set of equations determining the time-evolution of the internal dynamical variables. It needs a physical support unit in order to function properly, a datum also denoted as `embedded intelligence’. The primary task for a cognitive system is to retain functionality in certain environments. For this purpose it needs an operational physical support unit for acting and for obtaining sensory information about the environment. The cognitive system remains operational as long as its physical support unit, its body, survives. A cognitive system might be either biological (humans and non-human animals) or synthetic. Non-trivial cognitive systems are capable of learning and of adapting to a changing environment. High-level cognitive systems may show various degrees of intelligence.
Physical Support Unit: Also denoted `body’ for biological cognitive systems. Generally it can be subdivided into four functional distinct components. (A) The component responsible for evaluating the time-evolution equations of the cognitive system, viz the brain. (B) The actuators, viz the limbs, responsible for processing the output-signals of the cognitive system. (C) The sensory organs providing appropriate input information on both the external environment and on the current status of the physical support unit. (D) The modules responsible for keeping the other components alive, viz the internal organs. Artificial cognitive systems dispose of equivalent functional components.
Reinforcement Signal: Reinforcement signals can be either positive or negative, i.e. a form of reward or punishment. The positive or negative consequences of an action, or of a series of consecutive actions, are taken to reinforce or to suppress the likelihood of selecting the same set of actions when confronted with a similar problem-setting in the future. A reinforcement signal can be generated by a cognitive system only when a nominal target outcome is known. When this target value is given `by hand’ from the outside, viz by an external teacher, one speaks of `supervised learning’. When the target value is generated internally one speaks of `unsupervised learning’. The internal generation of meaningful target values constitutes the core of the motivational problem.