
The Cortical Conductor Theory: Towards Addressing Consciousness in AI Models Joscha Bach ([email protected]) Harvard Program for Evolutionary Dynamics, One Brattle Square #6 Cambridge, MA 02139 USA Abstract The absence of an understanding of substrate independent machines lead Leibniz to the rejection of mechanist AI models of the mind rarely discuss the so called “hard problem” of consciousness. Here, I will sketch informally a philosophy: “Perception, and what depends on it i.e., possible functional explanation for phenomenal cognition], is inexplicable in a mechanical way, that is, consciousness: the conductor theory of consciousness (CTC). using figures and motions. Suppose there would be a Unlike IIT, CTC is a functionalist model of consciousness, machine, so arranged as to bring forth thoughts, with similarity to other functionalist approaches, such as the experiences and perceptions; it would then certainly be ones suggested by Dennett and Graziano. possible to imagine it to be proportionally enlarged, in such Keywords: phenomenal consciousness, cortical conductor a way as to allow entering it, like into a mill. This theory, attention, executive function, binding, IIT presupposed, one will not find anything upon its examination besides individual parts, pushing each other— and never anything by which a perception could be “No computer has ever been designed that is ever aware of explained.” (1714). Conversely, its inkling prompted Julien what it's doing; but most of the time, we aren't either.” Offray de LaMettrie’s (1748) remark that while minds are – Marvin Minsky machines, these had to be thought of as “immortal” and Introduction: Artificial Intelligence as a “transcendental”. Computation is sometimes seen in opposition to computational science of the mind dynamical systems (see van Gelder 1998), and we can Understanding the nature of our minds and their relationship distinguish different classes of computational systems to to the universe has always been one of the most significant account for that, based on the classes of functions they can questions philosophy sought to address. compute effectively (in the unlimited case) and efficiently For centuries, scientists and philosophers emphasized the (with reasonably bounded resources), starting from role of mathematics in this quest. The discovery of the idea (deterministic or probabilistic) finite state machines over of computation and its formalization in the 1920ies by Alan Turing Machines to different classes of hyper-computers Turing and Alonzo Church paved the way to a new way of capable of continuous state change or even a-causal thinking about thinking, and replaced the old intuitions of computers that may allow a transition function to use mechanistic philosophy with more precise ones of information from a future state of the machine. We find that computationalism, and opened up the way of building the while dynamical systems often cannot be effectively new family of theories of computational systems. computed on a finite state machine (such as a von Neumann The relationship between mathematics and computation is computer), they can often be efficiently approximated. (The not trivial, and even though computation is defined metaphysical implications of whether our universe can only mathematically (and constructive mathematics is arguably realize finite state machines or hyper-computation are computational), it makes sense to understand them as profound and sometimes of concern in the philosophy of separate realms. Mathematics is the domain of all formal mind, but outside the scope of this discussion.) languages, and allows the expression of arbitrary statements The formation of a new, computational study of the mind (most of which are uncomputable). Computation may be was fraught with difficulty from the start. By the 1950ies, understood in terms of computational systems, for instance the influence of positivism had lead to the emergence and via defining states (which are sets of discernible differences, entrenchment of behaviorism in psychology, which stifled i.e. bits), and transition functions that let us derive new theoretical psychology and made it evidently impossible for states. Whereas mathematics is the realm of specification, psychologists to formulate comprehensive theories of the computation is the realm of implementation; it captures all mind, so a new discipline was established: Artificial those systems that can actually be realized. Intelligence was the attempt of thinkers like Marvin Computational systems are machines that can be Minsky, John McCarthy and others to treat the mind as a described apriori and systematically, and implemented on computational system, and thereby open its study to every substrate that elicits the causal properties that are experimental exploration by building computational necessary to capture the respective states and transition machines that would attempt to replicate the functionality of functions. minds. Artificial Intelligence soon formed two camps: one that systems that maximize Φ by maximally distributing was dedicated to the study of intelligence, and one that information (for instance via highly interconnected XOR focused on the automation of tasks that required human gates). Should we assign consciousness to processing intelligence. While both camps developed applications and circuits that are incapable of exhibiting any of the theories and often worked on similar systems, the rift interesting behaviors of systems that we usually suspect to between “cognitive AI “ and “narrow AI” widened, partly be conscious, such as humans and other higher animals? because large factions of the cognitive AI camp championed From a computationalist perspective, IIT is problematic, symbolic approaches and rejected neural learning as because it suggests that two systems that compute the same simplistic. The failure to deliver on some of the early, function by undergoing a functionally identical sequence of optimistic promises of machine intelligence, as well as states might have different degrees of consciousness based cultural opposition, lead to cuts in funding for cognitive AI, on the arrangement of the computational elements that and eventually the start of the new discipline of Cognitive realize the causal structure of the system. A computational Science. However, Cognitive Science did not develop a system might turn out to be conscious or unconscious cohesive methodology and theoretical outlook, and became regardless of its behavior (including all its utterances an umbrella term for neuroscience, AI, cognitive professing its phenomenal experience) depending on the psychology, linguistics and philosophy of mind. physical layout of its substrate, or the introduction of a In the last five years, AI research has been dominated by distributed virtual machine layer. the success of deep learning, which was fueled by A more practical criticism stems from observing theoretical insights into the training of neural networks with conscious and unconscious people: a somnambulist (who is a large number of hidden layers, advances in computer generally not regarded as conscious) can often answer hardware, and partially by the availability of large amounts questions, navigate a house, open doors etc., and hence of training data. The rapid advances of learning machines should have cortical activity that is distributed in a similar have lead to a renewed interest in the original goals of AI, as way as it is in an awake, conscious person (Zadra et al. well as the dissemination and development of ideas on the 2013). In this sense, there is probably only a low nature of learning, perception, and mental representation. quantitative difference in Φ, but a large qualitative However, the recent progress was arguably driven by difference in consciousness. This qualitative difference can successes in the narrow AI camp, and AI as a field is not probably be explained by the absence of very particular, very much concerned with the study of minds any more. local functionality in the brain of the somnambulist: while Progress on this front will likely require a better her cortex still produces the usual content, i.e. processes understanding of our mental architecture, reasoning, sensory data and generates dynamic experiences of sounds, language, reflection, self model and consciousness. patterns, objects, spaces etc. from them, the part that normally attends to that experience and integrates it into a Consciousness in cognitive science protocol is offline. This integrated experience is not the While AI offers a large body of work on agency, autonomy, same as information integration in IIT. Rather, it is better motivation and affect, cognitive architectures and cognitive understood as a particular local protocol by one of the many modeling, there is little agreement on how to address what is members of the “cortical orchestra” : its conductor. usually called “the hard problem” of consciousness. How is it possible that a system can take a first person perspective, In this contribution, I will sketch how a computational and have phenomenal experience? model can account for the phenomenology and functionality One of the better known recent attempts to address of consciousness, based on my earlier work in the area of phenomenal consciousness is Guilio Tononi’s Integrated cognitive architectures (Bach 2009); we might call this Information Theory (IIT) (2012, 2016), which has been approach the “conductor theory of consciousness” (CTC). championed by the
Details
-
File Typepdf
-
Upload Time-
-
Content LanguagesEnglish
-
Upload UserAnonymous/Not logged-in
-
File Pages8 Page
-
File Size-