Muhammad Ali Khalidi’s book Cognitive Ontology: Taxonomic Practices in the Mind-Brain Sciences offers a compelling non-reductionist approach to understanding the furniture of the mind: the ‘real kinds’ of cognitive science. This book is a thoughtful and timely addition to current debates over the ontology of the mind-brain sciences. In these debates, some researchers propose that we should individuate cognitive kinds in virtue of their informational features: the format of internal representations and the ways they are manipulated and transformed. Such an approach focuses on what Marr (1982) called the algorithmic level of analysis: what kind of information storage and processing explains the transition from our sensory inputs to our motor outputs? Other researchers, meanwhile, propose that cognitive kinds should be individuated in terms of the neural states, processes, and activities which the brain uses, described at Marr’s implementation level of analysis. Khalidi’s approach to cognitive ontology is “guided by Marr’s distinction between levels of analysis” (239), but with a particular focus on Marr’s top level: the level of computational theory, at which we can analyze the system in terms of overall function or goal of the cognitive task.
Marr’s tri-level approach is generally understood methodologically: his levels of computational analysis are taken to be epistemic tools for describing and analyzing computational systems. But Khalidi wants to “add an ontological dimension to Marr’s methodological and epistemic account of the computational level” (27). He argues that the taxonomy of Marr’s approach provides the metaphysical structure of cognition, such that “the computational level is the proper domain of the cognitive” (240). This is a bold move, for a number of reasons.
First, there has been a lot of push-back on Marr’s levels even as a methodological tool. Churchland (1986) argued that there is no useful purpose to be served in applying Marr’s levels to much work in cognitive science, and it has more recently been suggested that Marr’s levels might be “self-imposed barriers to investigation” (Love 2015). Second, those who accept the epistemic benefits of Marr’s tri-level characterization might still disagree with the top-down approach: Bickle (2015), for example, suggests that developments in neuroscience have challenged Marr’s assumptions about the necessity of starting with the computational level. Third, even those who accept Marr’s top-down approach as a way of taxonomizing cognition might deny that this taxonomy should be used as the basis for an ontology of the cognitive. As Danks points out, the computational specification of a cognitive theory “significantly underdetermines the commitments that are implied by it” (Danks 2014, 16). It is also worth mentioning that even those who do read Marr’s levels ontologically are not all convinced that we should be realists about theories at Marr’s computational level. Many proponents of Bayesian rational analysis, for example, point to the idealizations involved in characterizing Marr’s computational level as evidence that it should be given an instrumentalist interpretation (see Oaksford and Chater 2007; Icard 2018).
There is a further question about the relationship between Marr’s levels of computational analysis and their application to cognitive science. Khalidi assumes that Marr’s computational level provides an intentional characterization of the cognitive task. According to Egan (1995), the function described at Marr’s computational level is purely mathematical. Egan proposes that when we give an intentional characterization of the computational theory in order to explain how it counts as a cognitive process in a particular environment, this characterization is an extrinsic description: the intrinsic formal description is the one that matters for individuation and taxonomy purposes. It is unclear whether Khalidi’s own particular brand of realism can accommodate this sort of view.
For me, one of the most intriguing aspects of Khalidi’s book was his focus on ‘cognitive phenomena’ as “a subset of the mental or psychological realm” (1). He seems to want to distinguish cognition proper from “affective, perceptual, sensory, or experiential aspects of mentation” (3), while also claiming that Marr’s computational level of analysis provides the proper domain of the cognitive. It is difficult to see how to reconcile this characterization of Marr with the fact that he originally developed his levels of analysis specifically to argue for an account of visual perception. It is also noticeable that Khalidi’s examples of cognition in this book seem to be cases of ‘central processing’: capacities which rely on domain-general non-modular cognitive capacities. But the more central and less modular a cognitive process is, the more difficult it is to give it a precise formal characterization at the computational level of analysis (Fodor 1983; Egan 1995; Bermudez 2014), and the less helpful this formal characterization will be in helping us identify potential algorithms that compute the function.
References
Bermúdez, José Luis (2014). Cognitive Science: An Introduction to the Science of the Mind. (2nd edition.) Cambridge University Press.
Bickle, John (2015). Marr and Reductionism. Topics in Cognitive Science 7 (2):299-311.
Churchland, Patricia Smith (1986). Neurophilosophy: Toward A Unified Science of the Mind-Brain. MIT Press.
Danks, David (2014). Unifying the Mind: cognitive representations as graphical models. MIT Press.
Egan, Frances (1995). Computation and content. Philosophical Review 104 (2):181-203.
Fodor, Jerry A. (1983). The Modularity of Mind: An Essay on Faculty Psychology. Cambridge, MA: MIT Press.
Icard, Thomas F. (2018). Bayes, Bounds, and Rational Analysis. Philosophy of Science 85 (1):79-101.
Jones, Matt & Love, Bradley C. (2011). Bayesian Fundamentalism or Enlightenment? On the explanatory status and theoretical contributions of Bayesian models of cognition. Behavioral and Brain Sciences 34 (4):169-188.
Love, Bradley C. (2015). The Algorithmic Level Is the Bridge Between Computation and Brain. Topics in Cognitive Science 7 (2):230-242.
Marr, David (1982). Vision. W. H. Freeman.
Oaksford, Mike & Chater, Nick (2007). Bayesian Rationality: The Probabilistic Approach to Human Reasoning. Oxford University Press.
I’ll try to reply briefly to two points raised by Zoe Drayson’s very valuable comments, which I interpret to, first, raise doubts about taking Marr’s levels of description or explanation ontologically seriously, and second, pose questions about the proper domain of cognition and the point of characterizing cognitive capacities in computational terms.
When it comes to the first worry, Drayson cites Egan (1995), who argues that the function described at Marr’s computational level is purely mathematical and that the computational characterization is an extrinsic description and that the intrinsic formal description is the one that matters for individuation and taxonomy purposes. But although I think that Egan has done a lot to clarify the relationship between the computational and algorithmic level, I don’t buy Egan’s claim that the computational characterization is an extrinsic description and doesn’t matter for taxonomy. This seems to rest in part on her claim that computationally characterized states don’t have determinate content, but that they “have determinate content only by appeal to various pragmatic considerations such as ease of explanation and connections to our commonsense practices” (Egan 2014, 130). In my view, their determinate content comes from their individuation by reference to etiological and environmental factors (as I try to say in the original post). So I think that a computational characterization is determinate and enters into explanation and prediction, which makes it indispensable for taxonomy. And since I’m a (pluralist) realist, I take taxonomy to be our best guide to ontology.
A second point made by Drayson is that I argue in the book that the computational level is the proper domain of cognition, though Marr developed his levels of analysis to argue for an account of visual perception. Moreover, my examples of cognition seem to be cases of domain-general, non-modular, central processing, but the more central and less modular, the more difficult it is to give it a precise formal characterization at the computational level (and the less helpful it is in identifying potential algorithms that compute the function). I would respond by saying that I should have made it clearer in the book that the cognitive domain is not the only domain situated at the computational level of description and explanation. I also would add that many of the examples of cognition in the book are not ones of domain-general, non-modular, central processing. Consider domain specificity itself, which is one of my central examples, as well as (arguably) language-thought (LT) processes and cognitive biases and heuristics. Also, it doesn’t seem too difficult to give computational accounts of such cognitive constructs as concepts, episodic memory, etc., and that’s partly what I try to do in the book. Having said that, it was a shortcoming of the book that I didn’t attempt to provide a characterization of cognition in general, or indeed try to determine whether cognition was a real kind in its own right. But luckily, there’s a body of work that addresses this question, which I’m happy to tap into. For example, in the book, I quote Camp (2009, 303-304), who characterizes “the most basic aim of thought” (which I take to be roughly the same as cognition) as “using information about the world to solve problems and facilitate one’s survival and flourishing.” Similarly, Buckner (2015) identifies a cluster of properties characteristic of cognition, all of which are related to behavioral flexibility. Briefly, I would say that the cognitive domain is characterized by individuals flexibly processing information, solving problems, performing tasks, and manipulating their inanimate and animate environments. When characterized thus, I think it becomes clearer why it is best investigated from Marr’s computational level, which asks: “What is the goal of the computation, why is it appropriate, and what is the logic of the strategy by which it can be carried out?” I would maintain that these inquiries are scientifically fruitful and they characterize a very large proportion of inquiries in the cognitive sciences), and they do not hinder, and often actively help, inquiries that identify potential algorithms.