This week the Brains Blog is hosting a symposium on Neal Anderson and Gualtiero Piccinini’s new book The Physical Signature of Computation: A Robust Mapping Account (Oxford University Press; eBook available here). Today’s post from Anderson and Piccinini provides introductory remarks and an overview of the content of the book. Through this week, we will have four commentary posts from André Curtis-Trudel (University of Cincinnati) on Tuesday, Marcin Milkowski (Polish Academy of Sciences) on Wednesday, Danielle Williams (Washington University in St. Louis) on Thursday, and Mazviita Chirimuuta (University of Edinburgh) on Friday.
-Trey Boone, Associate Editor
It is notoriously difficult to articulate what it is for a physical system to implement a computation. What exactly is the “secret sauce” that distinguishes physical systems that can compute from those that cannot? To some, the line is drawn by purely physical considerations — by whether and in what sense the transformations of a system’s physical states “mirror” those of abstract computational states in a putatively implemented computation. To others, the line is drawn at least in part by agents who (might) use a physical system for computation — by the usability or active use of a physical system by an agent as means to their computational ends. For yet others, there is no line at all: simply being a physical system—a brain, a laptop computer, a bucket of water, or anything else — makes it computational, either because the physical is taken to be already inherently computational or because one or more computational interpretations of any physical system’s microscopic dynamics can be patched together.
The range of possible answers to the question of what computational implementation amounts to is thus extraordinarily wide, and the dust has yet to settle. This may seem like an implausible state of affairs in the present era, given that so many of our technologies are computational and so many scientific approaches — perhaps most notably in the mind sciences — have adopted computational paradigms. Yet, appeals to computing technology or cognitive science cannot fully answer the implementation question. No amount of knowledge or experience with extant examples of computational implementation, such as those provided by everyday computing devices, can alone answer the question of what defines the general class of physical systems to which these examples belong, just as no amount of knowledge about zebras can define the genus Equus to which zebras belong. The mind sciences are even less well positioned to provide answers. Not all mind scientists regard cognition as computational, and even if cognition is computational then it is itself a specific form of physical computation as is any particular computing technology. Only appeals to more general considerations are likely to yield adequate answers to the implementation question.
In our book The Physical Signature of Computation: A Robust Mapping Account, we reconsider implementation through appeals to such considerations. We first examine physical and computational descriptions separately, and then consider their hybridization in computational descriptions of physical systems — i.e., descriptions that ascribe putative computations to physical systems. We critically assess the rationale for and consequences of various options for connecting the physical to the computational in such hybrid descriptions. While doing so, we identify several precise criteria that a mapping from physical states and transitions to computational states and transitions must satisfy if the physical states of a system are to implement the computational states. The result is our robust mapping account (RMA) of implementation.
The criterion that is most central to our account is what we call physical-computational equivalence (PCE): evolving physical states bear the same (neither more nor less) information about the evolving computation as do the computational states they map onto. This criterion selects those microphysical states that can be legitimately grouped together and granted “computational statehood” in attributions of computations to physical systems, filtering out the arbitrary and unprincipled state groupings upon which arguments for pancomputationalism and the indeterminacy of computation depend.
Such robust mappings, we argue, must be satisfied by legitimate physical implementations of computations — systems that bear genuine physical signatures of the computations they implement — distinguishing them from physical systems that can be regarded as computational only through interpretive excess. We support this conclusion by motivating and developing the RMA from the ground up, and by examining various implementation claims and forms of conventional and unconventional computation — including reversible and reservoir computation — through the lens of this account. We then bring the account to bear on foundational issues in physical computation including pancomputationalism, the alleged indeterminacy of computation, and computational theories of cognition.
The exposition unfolds as follows: After motivating our work through a philosophical introduction to physical computing (Chapter 1), we examine the elements of physical and computational descriptions and emphasize similarities and differences relevant to their hybridization (Chapter 2). We then discuss the hybridization of these two kinds of description into computational descriptions of physical systems (CDPSs) and the physical-to-computational mappings on which they are based (Chapter 3). Next, we consider various possibilities for what one might require — beyond the existence of a minimal physical-computational mapping — for acceptance of an implementation claim as adequate, and we introduce physical-computational equivalence as one such potential requirement. We express these possibilities in the form of a scheme for classifying CDPSs according to their descriptive strength (weak, robust, or strong), and show how this classification scheme can be used to elucidate and evaluate implementation claims (Chapter 4).
Through this exploration, we conclude that physical-computational equivalence—the defining criterion for robust computational descriptions—is an essential ingredient of implementation: physical systems that satisfy robust computational descriptions bear genuine physical signatures of the computations that they implement. This leads us to the robust mapping account of implementation, which we articulate in detail, use to argue for the objectivity and determinacy of physical computation, place in the context of other accounts of implementation, and connect to results from physical information theory (Chapter 5).
With the robust mapping account in hand, we evaluate pancomputationalism in its three dominant forms: unlimited (every physical system implements lots of computations), limited (every physical system implements at least one computation), and ontic (every physical system is essentially computational). We first consider the first two forms (Chapters 6 and 7), to which our account can be directly applied. We conclude that our robust mapping account rules out both unlimited and limited pancomputationalism. We then consider the third (Chapter 8), which can be illuminated by our account only indirectly since it regards all physical reality as already essentially computational. We conclude that ontic pancomputationalism is unwarranted and, as a corollary, that we don’t live in a computer simulation.
Our argument against ontic pancomputationalism relies in part on the thesis that consciousness outstrips computation. To defend that thesis, we consider the relation between computation and the mind, appealing to our RMA to illuminate the foundations of the mind sciences where we can (Chapter 9). We argue that cognition involves computation in at least our robust sense, with empirical investigation of neural systems required to find its physical signature and study its properties. We also distinguish between causal powers and physical qualities and argue that consciousness most likely involves not only computational properties, which can be accounted for in terms of causal powers, but also physical qualities. This last conclusion further supports our argument that the nature of the physical universe cannot be wholly computational (Chapter 8). We finish the book by summarizing our conclusions and the intended takeaways from our work, and by pointing out that our robust mapping account can be added as an ingredient to improve other, more specialized accounts of physical computation, such as semantic and mechanistic accounts (Chapter 10).
Our primary aims in The Physical Signature of Computation were to clarify the nature of implementation, place it on a more solid footing, identify the specific feature(s) of physical computing systems that distinguish them from other physical systems, and bring the resulting account (and, where warranted, additional considerations) to bear on unresolved problems in the foundations of physical computation. We hope to have contributed on all of these fronts, and that our book advances the discussion — starting with this Brains blog book symposium.
— Neal Anderson and Gualtiero Piccinini