Neurocognitive Mechanisms

Gualtiero Piccinini

Reviewed by Matteo Colombo

Neurocognitive Mechanisms: Explaining Biological Cognition
Gualtiero Piccinini
Oxford: Oxford University Press, 2020, £80.00
ISBN 9780198866282

Gualtiero Piccinini’s Neurocognitive Mechanisms clarifies and addresses several foundational questions in the cognitive sciences, including: How does a multi-level mechanism like the brain produce cognitive capacities? Is there a fundamental level of explanation for cognition? How does explanation in psychology relate to explanation in neuroscience? What’s a promising route to explain consciousness? (Its answers in reverse order: ‘A noncomputational version of functionalism’, ‘They are mutually constraining and should be integrated in multilevel mechanistic explanations’, ‘No, all levels are ontologically equal’, ‘By performing neural computations over neural representations’.)

What ties all these different threads together is the claim that ‘biological cognitive capacities are constitutively explained by multilevel neurocognitive mechanisms, which perform neural computations over neural representations’ (p. 1). The fourteen chapters of Neurocognitive Mechanisms unpack, situate historically, and build a sustained argument in support of this claim with a characteristically neat writing style, offering a synthesis of Piccinini’s twenty-year research programme at the intersection of new mechanism and the philosophy of computation. (All but the first and third chapter draw on previously published work by Piccinini and his collaborators; but they also include novel ideas and substantive revisions relative to their published ancestors.)

The book starts, in Chapter 1, from the observation that cognitive neuroscientists often explain the cognitive capacities of biological systems in terms of their constitutive causal structures (or, mechanisms), which span several levels of organization, from single cells up to the entire nervous system. Cognitive capacities consist of high-level causal properties of brains defined only by their regular effects in normal conditions. Entities at higher levels in a mechanism are composed of entities at the next lower level, and higher-level properties are realized by lower-level properties. But—argues Piccinini—appealing to the part–whole relationships of composition and realization to make sense of how multi-level mechanisms constitutively explain cognitive capacities does not commit one to the conclusion that there is some ontologically fundamental level. Piccinini suggests that higher levels are invariant aspects of lower levels, because at least some higher-level properties of any given mechanism are invariant under some range of transformations to some of its components.

Piccinini acknowledges that the notion of X being an invariant aspect of Y deserves more probing and articulation than what he offers in this chapter—and one route to sharpen that notion might be to deploy tools and results available in the philosophy of science on explanatory depth (for example, Hitchcock and Woodward [2003]; Weslake [2010]; Strevens [2011]; Woodward [2021]). Yet, the egalitarian ontology of levels introduced in this chapter suffices to help us rethink traditional views in the metaphysics of mind, such as (anti-)reductionism and functionalism.

In fact, Chapter 2 uses this ontology to argue against both reductionism and anti-reductionism. Because higher-level properties are aspects of their lower-level realizers, cognitive capacities are not strictly identical with neural properties. So, reductionism should go. On the other hand, the same type of higher-level causal property can be multiply realized, when it is an aspect of different types of lower-level properties possessed by different kinds of causal structures. But the multiple realizability of cognitive capacities does not entail that they are ontologically autonomous of neural properties, because brains do not possess global causal powers that their organized lower-level components, considered collectively, lack.

Chapter 3 explicates the idea that cognitive capacities are causal roles, by introducing a goal-contribution account of teleological functions. According to this account, the cognitive capacities of any biological system consist of causal roles that make regular contributions to either the biological or non-biological goals of the system. This teleology grounds the normativity of functions, where some functional ascriptions are objectively correct and some incorrect, and also justifies treating certain features of a system as functionally irrelevant or as malfunctioning.

One set of issues Piccinini’s goal-contribution account does not address concerns the complex interplay between biological and non-biological goals of organisms. A related set of issues is epistemic. Given that organisms achieve the biological goals identified by Piccinini—namely, survival, development, reproduction, and helping others—over time and at relatively longer time scales, and given that these goals often trade off or conflict with other shorter-term, biologically insignificant goals of organisms, how should investigators discover a system’s teleological function(s)? What evidence, methodologies, and experimental designs will help us to find out? How should researchers resolve disagreements about, say, the teleological function(s) of the primary visual cortex? (Is the function performing Gabor transforms, sparsifying the neural code, detecting edges and other features of visual scenes, making probabilistic causal inferences, making action-oriented predictions, all of the above, or…?)

Neurocognitive Mechanisms says relatively little about epistemic questions concerning evidence, confirmation, discovery, and scientific technologies and methodologies for model testing and model selection, concentrating on questions in metaphysics.

This is evident in Chapter 4, where Piccinini distinguishes and situates historically different versions of functionalism, and argues that in the context of cognitive neuroscience, one should understand functionalism in terms of the teleological functions of neural mechanisms. Chapter 5 distinguishes functionalism from computationalism—namely, the idea that cognitive capacities are explained by computations performed by the brain—offering a close reading of (McCulloch and Pitts [1943]). This and the previous chapter were perhaps my favourites, as Neurocognitive Mechanisms is at its best when theoretical insight and tight argument are combined with illuminating historical detail.

To help us appreciate what McCulloch and Pitts got wrong, but also right, Chapter 6 distinguishes information processing from different species of computation (including digital, analogue, quantum, and generic computation), and argues that ‘a physical computing system [in the generic sense] is a functional mechanism whose [teleological] function is to manipulate medium-independent variables in accordance with a rule defined over the variables’ (p. 144). While this formulation is uncommitted to the idea that computations must be defined over semantically evaluable items, it preserves McCulloch and Pitts’s original insight that neural computing explains cognition.

Chapter 7 articulates this insight, arguing that all adequate computational explanations of cognitive capacities are constitutive causal explanations, which always involve all and only the relevant details—‘the ones that make the most difference to the behaviour of the whole’ (p. 160)—of the multi-level, spatially and temporally organized causal structures ‘internal to’ the capacities of interest. Since constitutive explanations are meant to offer an analysis of causal capacities, and functional explanations in psychology offer such an analysis, functional explanations in psychology are species of constitutive explanation. If all adequate constitutive explanations must include mechanistic information, then ‘functional analysis in psychology is neither distinct nor autonomous from mechanistic explanation’ (p. 163).

The general point that explanation in psychology and neuroscience can fruitfully constrain each other is persuasive; and it is also uncontroversial that evidence from neuroscience matters when evaluating the degree of confirmation of competitive psychological explanations of a given capacity. Less clear is the intended scope of mechanistic explanation.

In line with Salmon’s ([1984], p. 275) seminal work on causal explanation, Piccinini acknowledges that there are non-constitutive forms of causal explanation such as aetiological explanations, which offer causal histories leading up to the occurrence of events or behaviours of interest (pp. 156–57). And consistently with, for example, Chirimuuta ([2018]), Piccinini recognizes that not all analyses of neural computation or information processing are mechanistic (p. 190). These remarks make conceptual space for species of computational explanations in the cognitive sciences that are neither causal nor mechanistic. There might be constitutive, non-causal, computational explanations of certain capacities, which would provide information about non-causal properties ‘intrinsic’ to the target capacity; and there might be aetiological, computational explanations of cognitive events and behaviours, which would offer information about the ‘extrinsic’ causes or pathways leading up to those events. These possibilities acquire plausibility if we appreciate the diversity of epistemic goals and explanatory practices in different corners of the cognitive sciences (for example, Kording et al. [2020]; Silberstein [2021]), as well as that different species of computational explanation may be especially fitted for furthering only some of those goals and practices.

Chapters 8 to 14 combine Piccinini’s integrationist, mechanism-centred account of explanation in psychology and neuroscience with his mechanistic account of concrete computation to formulate and defend a neurocomputational theory of cognition. In particular, after Chapter 8 illustrates how cognitive neuroscience integrates psychological and neuroscientific explanations in the search for multi-level mechanisms, then Chapters 9, 12, and 13 argue, respectively, that neurocognitive processes are computational, that they operate over neural representations, and that they are ‘sui generis’ (neither digital nor analogue) computations. Along the way, Chapter 10 forestalls fallacious arguments, which are sometimes cited in support of the idea that brains are computers; and Chapter 11 identifies, evaluates, and rejects the objection that computation is either insufficient or unnecessary to explain cognition. Chapter 14 concludes the book by considering whether and how the neurocomputational theory developed in the previous chapters may explain not only cognition but also consciousness. The chapter draws on distinctions between functionalism, mechanism, and computationalism to canvass possible approaches to explaining conscious experience, highlighting a non-computational version of mechanistic functionalism as a promising path that one might take.

Many of the arguments contained in these chapters hinge on the key notion of medium independence, which ‘occurs when […] the inputs and outputs of a higher-level property are multiply realizable, so that the higher-level property can be realized by any lower-level structure with the right degrees of freedom organized in the right way’ (p. 66). If the low-level functional properties that make the most difference to the higher-level properties of whole brains are properties of sequences of action potentials in neurons—for example, their rate or timing—and these functional properties are individuated only in terms of certain degrees of freedom and their organization, without reference to any other properties of their realizing, physical medium, then they are medium independent. And if neurocognitive processes are functional manipulations of these medium-independent vehicles in accordance with a rule, then neurocognitive processes are computations.

Piccinini offers helpful remarks on medium independence, especially in Chapters 2, 6, and 9. But there are several outstanding questions. One thing to notice is that individuating medium-independent properties by removing, or abstracting away from, any property of its realizing medium except certain degrees of freedom of the medium does not make medium-independent properties causally inefficacious, non-concrete properties, as some critics might mistakenly conclude. (My financial advisor assures me that interest rates are making things happen in my bank account and elsewhere.) Another thing to notice is that medium independence seems to be a graded, contrastive notion. Depending on the relevant number of degrees of freedom of a physical medium, and the different ways they could be organized, a given functional property may be more or less dependent on the specific physical properties of that medium compared to other media and some other functional property. At a more basic level, evaluating whether and in what sense properties like the rate of spiking in a population of neurons are medium-independent requires we get clearer on what degrees of freedom are for a biological system, what degrees of freedom electrochemical signals in the brain possess, how these degrees of freedom are taken into account in empirically successful computational models of the brain, and, more generally, how an embodied, environmentally embedded, complex system like the brain may gain or lose certain degrees of freedom.

All in all, Neurocognitive Mechanisms is a welcome addition to the philosophy of the cognitive sciences. It offers a pellucid example of how ideas about concrete computation can fruitfully be applied to foundational questions about biological cognition. As this book develops a comprehensive, metaphysically sophisticated, and empirically informed neurocomputational theory of cognition, it should be of particular interest to philosophers of the cognitive sciences with a taste for metaphysical questions, and, more generally, to anybody interested in a clear-headed, up-to-date account of computationalism.

Matteo Colombo
Tilburg Center for Logic, Ethics, and Philosophy of Science
m.colombo@uvt.nl

References

Chirimuuta, M. [2018]: ‘Explanation in Computational Neuroscience: Causal and Non-Causal’, British Journal for the Philosophy of Science, 69, pp. 849–80.

Hitchcock, C. and Woodward, J. [2003]: ‘Explanatory Generalizations, Part II: Plumbing Explanatory Depth’, Noûs, 37, pp. 181–99.

Kording, K. P., Blohm, G., Schrater, P. and Kay, K. [2020]: ‘Appreciating the Variety of Goals in Computational Neuroscience’, Neurons, Behavior, Data Analysis, and Theory, 3, available at .

McCulloch, W. and Pitts, W. [1943]: ‘A Logical Calculus of the Ideas Immanent in Nervous Activity’, The Bulletin of Mathematical Biophysics, 5, pp. 115–33.

Salmon, W. [1984]: Scientific Explanation and the Causal Structure of the World, Princeton, NJ: Princeton University Press.

Silberstein, M. [2021]: ‘Constraints on Localization and Decomposition as Explanatory Strategies in the Biological Sciences 2.0’, in F. Calzavarini and M. Viola (eds), Neural Mechanisms: New Challenges in the Philosophy of Neuroscience, Cham: Springer, pp. 363–93.

Strevens, M. [2011]: Depth: An Account of Scientific Explanation, Cambridge, MA: Harvard University Press.

Weslake, B. [2010]: ‘Explanatory Depth’, Philosophy of Science, 77, pp. 273–94.

Woodward, J. [2021]: ‘Explanatory Autonomy: The Role of Proportionality, Stability, and Conditional Irrelevance’, Synthese, 198, pp. 237–65.