1. What is a theory of consciousness?
A theory of consciousness is a theory of the nature of experience. Bernard J. Baars (1993) believes that "conscious experience is notoriously the great, confusing and nub of psychological science". Consciousness is not something we can observe directly other than in ourselves. Historically, psychologists have neither addressed nor evaded consciousness successfully. There are two psychological metatheories, introspectionism and behaviorism. These metatheories are inherently and deeply flawed.
What is introspectionism? According to Rui Miguel Costa (2020), introspection is the method of studying psychological processes relying on systematic self-observation of thoughts, perceptions, feelings and bodily sensations. Behaviorist critics have introduced “introspectionism” to subsume all the diverse views and meanings of introspection under a common umbrella. What is behavioralism? According to John Staddon, behaviorism is in the middle of the mental-biological or physiological, neither mentalistic nor, at its core, physiological. For John Watson, a Johns Hopkins psychologist, it is a redefinition of psychology as an objective study of behavior.
19th-century psychologists like Wilhelm Wundt and William James suggested that consciousness was the fundamental constitutive problem for psychology. The purpose is not to interpret the marvellous historical literature but develop a theory to simplify our understanding of conscious experience. William James believed that one of the most puzzling phenomena in psychology, conscious experience, was the foundation of scientific psychology.
However, building on a confusing and misunderstood foundation is a recipe for disaster. The new cognitive metatheory has overcome the problems with behaviorism and instropectionism. It encourages psychologists to move beyond raw observations to infer explanatory entities if the evidence for them is compelling. Both conscious and unconscious processes consist of inferences from publicly observable data.
2. Four common theories of consciousness and Coding Consciousness in AI
There are four main theories of consciousness: higher-order theories, global workspace theories, re-entry and predictive processing theories and integrated information theory.
Rocco J. Gennaro (2018) asserts that the key to higher-order theories is their hierarchical or iterative structure, so they have been called “double-tiered” theories. Higher-order theories do not reduce consciousness directly to neurophysiological states. They don’t try to explain consciousness in physicalistic terms. Higher-order theories attempt to explain consciousness in mentalistic phrases, i.e. “thoughts” and “awareness”.
Global Workspace Theory (GWT) began with this question:” How does a serial, integrated and very narrow stream of consciousness emerge from a nervous system that is predominantly unconscious, distributed, parallel and of enormous capacity?”. GWT is a widely used model for the role of conscious and unconscious events in the functioning of the brain, a set of explicit assumptions which we can test, as many of them have been over several decades. According to Bernard J. Baars (2005), GW theory generates understandable predictions for conscious aspects of perception, emotion, motivation, learning, working memory, voluntary control, and self-systems in the brain. It is somewhat similar to biological theories such as Neural Darwinism and dynamic theories of brain functioning.
Tomáš Marvan and Marek Havlík (2020) suggest that predictive processing is presently one of the most debatable theories of brain function. The theory captures the brain as a hypothesis-testing machine matching perceptual hypothesis produced by an internal hierarchical model with input coming via sensory channels. Learning and “hard-wired” evolutionary constraints construct hypotheses of the internal model. If Predictive Processing indeed aspires to be a universal theory of the brain and to explain “perfection and action and everything mental in between”, Predictive Processing needs to be able to explain consciousness. Some reports propose a close link between some of the principles postulated by Predictive Processing theories and the actual contents of conscious perception.
G. Tononi (2012) claims that Integrated Information Theory (IIT) consists of an approach to explain the neural correlates of consciousness, why there is generated experience in the cortex and not in the cerebellum, why it fades in certain stages of sleep, why some cortical areas contribute color and other sounds and so forth. Three thought experiments lie at the heart of IIT: the photodiode thought experiment, the camera thought experiment, and the internet thought experiment. Based on the intuition from these three thought experiments, the main pillars of IIT may be a set of phenomenological axioms, ontological postulates and identities.
To program consciousness in AI, we need to have an AI-consciousness theory. According to Adam Safron (2020), Integrated World Modelling Theory (IWMT) portrays a synthetic approach to understanding minds revealing compatibility leading theories of consciousness, thus enabling inferential synergy. IWMT is a synthetic approach to understanding consciousness, using the Free Energy Principle and Active Inference Framework (FEP-AI) to combine multiple theories into a unified perspective. IWMT focuses on Integrated Information Theory (IIT) and Global Neuronal Workspace Theory (GNWT) as the most well-known theories of consciousness. IWMT suggests such coherence is only likely attainable for embodied agentic systems with controllers capable of supporting complexes of high degrees of integrated information.
FEP-AI (The Free Energy Principle) states that persisting systems must entail predictive models to resist entropic mixing, prevent destruction and maintain their forms. Systems must adaptively respond to various events and be able to model these events in some capacity. With FEP-AI, we can view neural dynamics as implementing approximate Bayesian inference where activation cascades contain a message-passing regime. Theoretically, differential rates of message passing may automatically find maximally connected subnetworks and, thus, converge on processes of variable elimination.
IWMT believes that many processes and systems underlying consciousness can be transferrable regarding computational principles from machine learning. We can derive computational principles from machine learning. Some architectures may correspond closely with the neural processes contributing to consciousness via coherent world modelling.
3. Future of Conscious AI
By understanding and utilising theories of consciousness, we might be able to understand and develop consciousness in AI. Combining some of these theories may result in better comprehension of consciousness. Coding consciousness in AI may be similar to how we code emotions in AI. Future Human-level AI might require consciousness. Future articles will discuss theories of consciousness for AI and how to program it.
References
Kommentarer