3. Introducing Cognitive Structures

In the previous post, I have remarked that the existing forms of SR do not use the full capacity of their logical frameworks to account for a substantial relation between the structure of the scientific theories and reality. If we regiment the structure of scientific theories into formal frameworks that have a propensity for being grounded in the world, the problem of representation will be dissolved. Cognitive structures, i.e., informational structures embodied in cognitive mechanisms, possess such a precious propensity. Thus, Cognitive SR (CSR) replaces purely abstract structures with embodied cognitive structures that have a propensity for being entwined with the causal structure of the world.  By the late 1980s, R. Giere and some colleagues such as Paul Churchland recognized the role of modelling tools offered by AI and neuroscience (e.g., neural networks and connectionist webs) in developing (biologically) realistic models of advanced forms of cognitive rationality (such as theories). In chapter five of the book, I argue that the same connections networks that are employed in Churchland’s account of theories could play the role of the underlying structures of scientific models. Given the neural networks’ propensity for representing the causal structure of the world by the configuration of the vector spaces, the scientific structures (when regimented into neural networks) could be grounded in the external world easily. It is true that Churchland’s theory has been presented in intimate connection with Cognitive Models of Science Approach (CMSA) which is a foil to SR. However, in chapter five, I argue that the cognitive models that are introduced by Churchland could be incorporated into structuralist strategies that address some notorious problems that target realism—these are problems of Pessimistic Meta-Induction and Metaphysical Underdetermination. Pessimistic meta-induction builds upon cases of useful though false theories in the history of science to challenge the assumption of the cumulative trajectory of knowledge and thereby question the relation between the empirical success of theories and their truth.

It is possible to reconstruct Churchland’s connectionist account for the cumulative history of science on the basis of assimilation of less precise representational maps by more comprehensive and precise maps. ‘Representational maps’ can be construed in terms of cognitive structures, or the activation spaces of the neural networks, when such spaces represent the states of the external world. Thus, it could be argued that the veridical core of less correct theories would be retained and developed by the more advanced theories, and theoretical continuity underpins shifts and revolutions in the cumulative history of science. This provides a basis for defending a meaningful relation between the empirical success of theories and their verisimilitude. As the orthodox structural realist states it, continuity endures at the level of the structure of theories, not their content, but here the structure is cognitive and it is specified in terms of the configuration of neural networks (the problem of underdetermination could be dissolved in a similar manner). Be that as may, CSR accomplishes the goals that any other version of SR seeks to fulfil. It systematises the structures of scientific theories precisely enough, and it relies on structuralist strategies to address the problems of pessimistic induction and underdetermination. But it has the edge over the orthodox version of OSR because owing to its reliance on cognitive structures, it accounts for the theories’ relationship to the causal structures of the world in naturalistically plausible terms (on the basis of evolutionary theory and computational neuroscience). It is possible to set a foundation for CSR by reconstructing Churchland’s realist account of scientific theories along the lines of SR (by emphasising its capacity for facing pessimistic induction and underdetermination). But neuroscience has been developed extensively since the 1980s and it could be contended that artificial neural networks and their learning algorithms are not biologically realistic enough. To consolidate the plausibility of CSR, we need to draw on recent breakthroughs in computational neuroscience. Also, we have to provide further details about how CSR’s account of scientific representation could boost the realist core of the theory. In the final post, I will take care of these issues. That is to say, to consolidate the realist core of CSR, I will draw on resources of recent computational neuroscience and Predictive Processing Theory (PPT) to furnish the requisite details regarding the capacity of the biological brains for forging reliable models of their environment. Thus, CSR builds its account of scientific representation on the capacity of the brains for increasing the accuracy of their models and minimising the discrepancy between their and causal structure of the world.

Back to Top