[CogSci Summaries home | UP | email]

Sun, R., Coward, L. A., & Zenzen, M. J. (2005). On levels of cognitive modelling. Philosophical Psychology, 18, 613-637.

  author =       {Sun, Ron and Coward, L. Andrew and Zenzen, Michael J.},
  title =        {On Levels of Cognitive Modelling},
  journal =      {Philosophical Psychology},
  year =         {2005},
  volume =       {18},
  pages =        {613--637}

Author of the summary: Matt Martin, 2012, mmarti10@connect.carleton.ca

Cite this paper for:

This paper proposes a levelled, hierarchical framework for cognitive modelling that ranges from sociological to neurological levels. (p. 613, 614)

Computational modelling appears to be more promising than mathematical or verbal modelling of cognition within cognitive science. (p. 613)

The proposed hierarchical model contains four semi-discrete levels of explanation: Social/anthropological models of group behaviour, behavioural models of individual agency, cognitive models involving representations and processes, and biological/physiological models of neural processes etc. (p. 614)

A rigorous set of theories in cognitive science would involve a hierarchy of descriptions/explanations that can map causal relationships between levels. (p. 614)

Understanding the human mind/brain strictly from behavioural observation is untenable, there needs to be an a priori theoretical underpinning of how the system works for the mind to be properly understood. (p. 614)

Computational models are required to unify behavioural data, they provide specificity and conceptual clarity. (p. 615)

Computational models are better than pure mathematical models for understanding the human mind, they provide more flexibility and are better at describing the complex nature of the mind. (p. 615)

Cognitive architectures provide a concrete framework for a detailed understanding of cognitive phenomena; they are “broad, generic theories of cognition.” (p. 616)

“The main task of computational cognitive modelling is to develop models of cognitive agents and their interactions.” (p. 616, 617)

Marr's theory: the computational theory level determines the proper computational strategy to be performed, the representation and algorithm level carries out the computational strategy of the first level, and the hardware implementation level is the physical realization of the algorithms of the second level. These three levels are loosely connected, and some phenomena can only be explained by reference to one or two of them, not all of them simultaneously. (p. 617)

Newell and Simon's (1976) three level theory: the knowledge level explains an agent's actions by reference to its goals and knowledge, the symbol level encodes the knowledge and goals of an agent, the physical level realizes the symbols in some physical form. Cognition is best explained from a top-down approach beginning at the knowledge level and moving down to the physical level in this view. (p. 618)

Marr's theory and Newell and Simon's three-levelled approach make useless and fuzzy distinctions that have not lead to an increase in understanding. These approaches focus too much on the tools of modelling rather than examining the phenomena cognitive scientists are trying to study. (p. 618)

The authors propose a four-levelled approach to understanding cognitive phenomena. The four layers consist of the sociological level, the psychological level, the componential level, and the physiological level. (p. 619, 620)

The sociological level consists of inter-agent processes, socio-cultural processes and the interaction between agents and their (social) environments. (p. 619)

The psychological level consists of behaviours, beliefs and skills at an individual level. (p. 619)

The componential level consists of intra-agent processes and mechanisms. (p. 619)

The physiological level consists of the implementation of computation, ie. at the neural level (p.620)

The four layers interact with one another, and each layer may not be fully understood in isolation from the other layers. Multiple layers should be pursued simultaneously in order to fully understand cognition as a whole. (p. 620, 621)

In science, deeper more detailed theories can unify seemingly independent domains at higher levels. Deeper theories do not necessarily replace higher-level theories, they merely explain the concepts and causal relationships between higher levels. (p. 622)

Humans can only process a limited amount of information, so higher-level theories are needed to think about broader phenomena. For example, it does not make sense to talk about the interactions of billions of atoms in order to understand simple chemical equations. (p. 623)

Causal relationships between lower levels and higher levels must exist for a particular cognitive theory to be considered valid. (p. 623)

A scientific theory of cognition using the four-levelled approach should be able to describe causal relationships for sociocultural processes at its own high level as well as be able to explain those high relationships by reference to causal relationships in the next lower level and so on to the deepest level. (p. 624)

An effective theory of cognition should be mechanistic in nature, it would treat human cognitive systems in a way analogous to that of man-made technology. (p. 624)

A theory of cognition should establish a set of entities or conditions that explain causal relationships at a high level in conjunction with a smaller set of deeper level entities that comprise the entities in the higher levels. Descriptions at a higher levels are less detailed and contain less densely packed information than lower-levelled descriptions. (p. 625)

Whereas Marr's theory held that the three layers were only loosely coupled, the four-levelled approach holds that causal relationships at one level correspond to causal relationships at another level. The main difference between the levels is the amount of detail and the kinds of entities being explained. (p. 626)

Implementing cognitive architectures involve a two stage process. First, the essential structures and operations are created before simulation. Second, the details of the cognitive processes are filled in within the bounds of the constraints given by the essential structures of the first process. (p. 627)

Cognitive modelling that involves cross-level analysis has been intellectually enlightening, and may be essential to the future of computational cognitive modelling. Although most of science deals with singular levels of explanation, it is useful to understand the mapping between levels when considering higher phenomena. (p. 627, 628)

Mixed-level analysis may be necessary to explain certain phenomena; certain processes may be well explained at a certain level, but sometimes require reference to a lower level in order to explain certain exceptional events, eg. classical vs quantum physics. This speaks to the need of the ability to integrate various levels of explanation. (p. 628)

An example of mixed analysis involves social cognition, certain phenomena need to be explained in terms of the social level, the psychological level and the componential level. (p. 629)

A shift away from substance ontology and towards a process ontology of causality is needed within cognitive science. (p. 629)

The discovery of a causal nexus of cognition may result from a four-levelled approach; cognition at all levels could be explained by certain essential causal structures that map to all domains and levels. Certain cognitive architectures like ACT-R and CLARION may represent the implementation of such a causal nexus. (p. 630)

Some criticisms of computational modelling approaches are that the models can fit any range of possible human results, and that the proposed models only superficially fit real (noisy) human data. (p. 630)

The criticisms of the computational modelling approach do not invalidate the enterprise of cognitive modelling in itself, and some of the current problems may be overcome through use of better statistical techniques. (p. 630)

Some of the objections to computational modelling may be based out of outdated models of science such as inductionism and hypothetico-deductionism. (p. 631)

Computational cognitive modelling represents a paradigm in the sense of Thomas Kuhn's (1970) idea; initial assumptions of an architecture can be tested and revised as the data mandates, until the paradigm shifts again. (p. 631)

In order for a truly deep understanding of cognition to be possible, all levels of explanation must be taken into account, including causal correspondence between levels. Each level is important unto itself, and one should not 'dominate' the others, nor are others ultimately irrelevant (such as the physiological level). (p. 632)

Cognitive architectures such as ACT-R and CLARION may represent the kind of cognitive modelling that can bridge several different levels. These models may have the potential to become scientific theories in their own right, analogous to various theories in the physical sciences. (p. 633, 634)

A multi-level framework is necessary in order to deal with varying levels of complexity. (p. 634)

Causal descriptions must exist between the levels. Higher-level phenomena must bear some kind of consistent causal relationship between lower-levelled phenomena. (p. 634)

For a theory of cognition to be viable, not only must there be a consistent description within a singular level, but the entities and processes at one level must be able to hold causal connections with deeper levels. (p. 634)