Original ArticleApplying global workspace theory to the frame problem
Introduction
The frame problem was originally couched as a difficulty within classical Artificial Intelligence: How can we build a program capable of inferring the effects of an action without reasoning explicitly about all its obvious non-effects? But many philosophers saw the frame problem as symptomatic of a wider difficulty, namely how to account for cognitive processes capable of drawing on information from arbitrary domains of knowledge or expertise. So-called “informationally unencapsulated” processes of this sort, exemplified by analogical reasoning, are especially troublesome for theories of mind that rely on some sort of modular organisation to render them computationally feasible.
However, one thing is clear. If the frame problem is a genuine puzzle, the human brain incorporates a solution to it. In global workspace theory, we find clues to how this solution might work. Global workspace theory posits a functional role for consciousness, which is to facilitate information exchange among multiple, special-purpose, unconscious brain processes (Baars 1997, 1998). These compete for access to a global workspace, which allows selected information to be broadcast back to the whole system. Such an architecture accommodates high-speed, domain-specific processes (or “modules”) while facilitating just the sort of crossing of domain boundaries required to address the philosophers’ frame problem.
The paper is organised as follows. In 2 The frame problem, 3 The computational theory of mind, the philosophers' conception of the frame problem is presented. Section 4 challenges the premise that informationally unencapsulated cognitive processes are, in principle, computationally infeasible. In Section 5, global workspace theory is outlined. Arguments and evidence in favour of the theory are reviewed, and the global workspace architecture is commended as a model of combined serial and parallel information flow capable of overcoming the frame problem.
Section 6 concerns analogical reasoning, the epitome of informational unencapsulation, and demonstrates that the most successful of contemporary computational models of analogical reasoning are strikingly compatible with global workspace theory. The concluding discussion addresses a variety of topics including modularity, conscious information processing, and the relationship between parallel and serial computation in a generic account of cognitive function.
Section snippets
The frame problem
The frame problem, in its original form, was to address the following question (McCarthy and Hayes, 1969, Baars, 1997, Dennett, 1978, Newell, 1962). How is it possible to write a collection of axioms in mathematical logic that captures the effects of actions, without being obliged to include an overwhelming number of axioms that describe the trivial non-effects of those actions? In everyday discourse, we can describe the effect of, say, painting an object simply by detailing how its colour
The computational theory of mind
The concern of this paper is the frame problem in the wide sense intended by Fodor.4
Complexity and informational encapsulation
There is no doubt, of course, that some tasks are computationally intractable, in a sense that has been made mathematically precise (Garey & Johnson, 1979). To sharpen the discussion, it is worth reviewing the basic computer science. Consider a function F. Suppose it can be proved that an algorithm exists that, for any input string x of length n, can compute F(x) in less than or equal to T(n) steps. So T sets an upper bound on how long the computation will take, in the general case. The rate of
Global workspace theory
The discussion of the previous section suggests that a convincing case for the computational infeasibility of informationally unencapsulated cognitive processes has not been made. Proponents of the infeasibility thesis are insufficiently rigorous in their treatment of algorithmic complexity and are unsuccessful in demonstrating that computational problems follow from the nature of the cognitive processes in question. So it is legitimate to regard the existence of such processes as a problem
Analogical reasoning
Fodor says little about the computational model behind his claim that informationally unencapsulated cognitive processes are computationally infeasible. Yet there are strong hints of a commitment to a centralised, serial process that somehow has all the requisite information at its disposal, and then has the responsibility of choosing what information to access and when to access it. Although parallel peripheral processes are part of the picture, they are passive sources of information that
Discussion
Let's review the argument so far. We set out by undermining the in-principle claim that informationally unencapsulated cognitive processes are computationally infeasible. It turned out that the case put forward by Fodor and others is too weak to sustain such a conclusion. The way the biological brain handles such processes is thereby demoted from an out-and-out mystery to a scientific challenge. The global workspace architecture, with its blend of parallel and serial computation, was then
Acknowledgements
Thanks to Igor Aleksander, Ron Chrisley, Stan Franklin, and Mercedes Lahnstein. Thanks also to the paper's three anonymous reviewers.
References (55)
The conscious access hypothesis: Origins and recent evidence
Trends in Cognitive Science
(2002)- et al.
How conscious experience and working memory interact
Trends in Cognitive Science
(2003) - et al.
Brain, consciousness and the observing self
Trends in Neurosciences
(2003) - et al.
Towards a cognitive neuroscience of consciousness: Basic evidence and a workspace framework
Cognition
(2001) - et al.
Connectionism and cognitive architecture: A critique
Cognition
(1988) - et al.
A software agent model of consciousness
Consciousness and Cognition
(1999) Structure-mapping: A theoretical framework for analogy
Cognitive Science
(1983)A blackboard architecture for control
Artificial Intelligence
(1985)- et al.
Analogical mapping by constraint satisfaction
Cognitive Science
(1989) - et al.
Constraints on analogical mapping: A comparison of three models
Cognitive Science
(1994)
Analog retrieval by constraint satisfaction
Artificial Intelligence
Complexity and coherency: Integrating information in the brain
Trends in Cognitive Science
A cognitive theory of consciousness
In the theater of consciousness: The workspace of the mind
How brain reveals mind: Neuroimaging supports the central role of conscious experience
Journal of Consciousness Studies
Perceptual symbol systems
Behavioral and Brain Sciences
The cognitive functions of language
Behavioral and Brain Sciences
On Fodor's problem
Mind and Language
Minimal rationality
Cerebral mechanisms of word masking and unconscious repetition priming
Nature Neuroscience
A neuronal network model linking subjective reports and objective physiological data during conscious perception
Proceedings of the National Academy of Science
Brainstorms
Cognitive wheels: The frame problem in artificial intelligence
Consciousness explained
The structure-mapping engine: Algorithm and examples
Artificial Intelligence
The language of thought
The modularity of mind
Cited by (109)
The cognitive reality monitoring network and theories of consciousness
2024, Neuroscience ResearchCOVID-19 Symptoms app analysis to foresee healthcare impacts: Evidence from Northern Ireland
2022, Applied Soft ComputingCitation Excerpt :The context is described by producing an edge list created by concatenating the string name with its value as described by the following algorithm: The result is a network of nodes with colours representing communities and lines (thicker lines represent a more relevant interaction between nodes) that can be interpreted by non-experts in mathematics or statistics (as happens in more complex representations) and support a near real-time visualisation and attention structure [37] for COVID-19 analysis. Information entropy in complex adaptive systems [35,38] can measure the system complexity and is an important measurement to describe the structural change of a complex network.
Robot navigation as hierarchical active inference
2021, Neural NetworksUnconscious perception of one's own name modulates amplitude of the P3B ERP component
2020, NeuropsychologiaKnowledge acquisition through introspection in Human-Robot Cooperation
2018, Biologically Inspired Cognitive ArchitecturesAHA: A general cognitive architecture for Darwinian agents
2018, Biologically Inspired Cognitive ArchitecturesCitation Excerpt :Mapping unconstrained stimuli to a specific response does not scale up to realistic environments and quickly reaches an almost intractable computational complexity, thereby making strictly bottom-up “behaviorist” approach untenable (Tsotsos, 1995). More generally, the frame problem of the computational theory of mind illustrates how non-modular, unencapsulated processes are computationally infeasible (Shanahan & Baars, 2005). A surprisingly similar emphasis on the importance of internal architecture and phenotypic constraints can be found in the study of behavioral ecology and evolution beyond the “phenotypic gambit” (Andersen et al., 2016; Giske et al., 2013, 2014; Rubin, 2016; Springer, Crespi, & Swanson, 2011).