Oron Shagrir

A central claim in contemporary cognitive science is that the neural mechanisms that bring about cognitive capacities and behavior are computations. It is also widely assumed that computations are not sensitive to the content, or the semantic properties of representations. From this insensitivity premise, some infer that mental content is causally irrelevant—a position we term the sensitivity argument against the causal relevance of content. This conclusion, however, sits uneasily with everyday situations in which the contents of our representations appear indispensable to our actions. In this paper, we reject the view that such experiences are illusory and instead identify a flaw in the inference from insensitivity to content to its causal irrelevance. Drawing on the phenomenon of computational indeterminacy—the fact that a single physical state can realize multiple formalisms—we argue that content can be causally relevant even when computation itself is content‐insensitive. While remaining neutral on the precise link between higher level and physical causation, we show that computational indeterminacy offers a principled route to restoring the causal relevance of mental content.

A key component of scientific inquiry, especially inquiry devoted to developing mechanistic explanations, is delineating the phenomenon to be explained. The task of delineating phenomena, however, has not been sufficiently analyzed. We contend that Marr’s characterization of the computational level (CL) provides a valuable resource for understanding what is involved in delineating phenomena. Unfortunately, the distinctive feature of Marr’s computational level, his dual emphasis on what is computed and why it is computed, has not been appreciated in philosophical discussions of Marr. Accordingly we offer a distinctive account of CL. This then allows us to develop two important points about delineating phenomena. First, the accounts of phenomena that figure in explanatory practice are typically not qualitative but precise, formal, or mathematical representations. Second, delineating phenomena requires consideration of the demands the environment places on the mechanism—identifying, as Marr put it, the basis of the computed function in the world.

Accelerating Turing machines have attracted much attention in the last decade or so. They have been described as “the work-horse of hypercomputation” (Potgieter and Rosinger 2010: 853). But do they really compute beyond the “Turing limit”—e.g., compute the halting function? We argue that the answer depends on what you mean by an accelerating Turing machine, on what you mean by computation, and even on what you mean by a Turing machine. We show first that in the current literature the term “accelerating Turing machine” is used to refer to two very different species of accelerating machine, which we call end-stage-in and end-stage-out machines, respectively. We argue that end-stage-in accelerating machines are not Turing machines at all. We then present two differing conceptions of computation, the internal and the external, and introduce the notion of an epistemic embedding of a computation. We argue that no accelerating Turing machine computes the halting function in the internal sense. Finally, we distinguish between two very different conceptions of the Turing machine, the purist conception and the realist conception; and we argue that Turing himself was no subscriber to the purist conception. We conclude that under the realist conception, but not under the purist conception, an accelerating Turing machine is able to compute the halting function in the external sense. We adopt a relatively informal approach throughout, since we take the key issues to be philosophical rather than mathematical.

What are the limits of physical computation? In his ‘Church’s Thesis and Principles for Mechanisms’, Turing’s student Robin Gandy proved that any machine satisfying four idealised physical ‘principles’ is equivalent to some Turing machine. Gandy’s four principles in effect define a class of computing machines (‘Gandy machines’). Our question is: What is the relationship of this class to the class of all (ideal) physical computing machines? Gandy himself suggests that the relationship is identity. We do not share this view. We will point to interesting examples of (ideal) physical machines that fall outside the class of Gandy machines and compute functions that are not Turing-machine computable.

Computational physical systems may exhibit indeterminacy of computation (IC). Their identified physical dynamics may not suffice to select a unique computational profile. We consider this phenomenon from the point of view of cognitive science and examine how computational profiles of cognitive systems are identified and justified in practice, in the light of IC. To that end, we look at the literature on the underdetermination of theory by evidence and argue that the same devices that can be successfully employed to confirm physical hypotheses can also be used to rationally single out computational profiles, notwithstanding IC.

What does it mean to say that an object or system computes? What is it about laptops, smartphones, and nervous systems that they are considered to compute, and why does it seldom occur to us to describe stomachs, hurricanes, rocks, or chairs that way? Though computing systems are everywhere today, it is very difficult to answer these questions. The book aims to shed light on the subject by arguing for the semantic view of computation, which states that computingsystems are always accompanied by representations. This view is presented as an alternative to non-semantic views such as the mechanistic account of computation.

It is often indeterminate what function a given computational system computes. This phenomenon has been referred to as “computational indeterminacy” or “multiplicity of computations.” In this paper, we argue that what has typically been considered and referred to as the challenge of computational indeterminacy in fact subsumes two distinct phenomena, which are typically bundled together and should be teased apart. One kind of indeterminacy concerns a functional characterization of the system’s relevant behavior. Another kind concerns the manner in which the abstract states are interpreted. We discuss the similarities and differences between the two kinds of computational indeterminacy, their implications for certain accounts of “computational individuation” in the literature, and their relevance to different levels of description within the computational system. We also examine the inter-relationships between our proposed accounts of the two kinds of indeterminacy and the main accounts of “computational implementation.”

It is generally accepted that, in the cognitive and neural sciences, there are both computational and mechanistic explanations. We ask how computational explanations can integrate into the mechanistic hierarchy. The problem stems from the fact that implementation and mechanistic relations have different forms. The implementation relation, from the states of an abstract computational system to the physical, implementing states is a homomorphism mapping relation. The mechanistic relation, however, is that of part/whole; the explaining features in a mechanistic explanation are the components of the explanandum phenomenon and their causal organization. Moreover, each component in one level of mechanism is constituted and explained by components of an underlying level of mechanism. Hence, it seems, computational variables and functions cannot be mechanistically explained by the medium-dependent states and properties that implement them. How then, do the computational and the implementational integrate to create the mechanistic hierarchy? After explicating the general problem, we further demonstrate it through a concrete example, of reinforcement learning, in the cognitive and neural sciences. We then examine two possible solutions. On one solution, the mechanistic hierarchy embeds at the same levels computational and implementational properties. This picture fits with the view that computational explanations are mechanistic sketches. On the other solution, there are two separate hierarchies, one computational and another implementational, which are related by the implementation relation. This picture fits with the view that computational explanations are functional and autonomous explanations. It is less clear how these solutions fit with the view that computational explanations are full-fledged mechanistic explanations. Finally, we argue that both pictures are consistent with the reinforcement learning example, but that scientific practice does not align with the view that computational models are merely mechanistic sketches.

The mechanistic view of computation contends that computational explanations are mechanistic explanations. Mechanists, however, disagree about the precise role that the environment – or the so-called “contextual level” – plays for computational explanations. We advance here two claims: Contextual factors essentially determine the computational identity of a computing system ; this means that specifying the “intrinsic” mechanism is not sufficient to fix the computational identity of the system. It is not necessary to specify the causal-mechanistic interaction between the system and its context in order to offer a complete and adequate computational explanation. While the first claim has been discussed before, the second has been practically ignored. After supporting these claims, we discuss the implications of our contextualist view for the mechanistic view of computational explanation. Our aim is to show that some versions of the mechanistic view are consistent with the contextualist view, whilst others are not.

The semantic view of computation is the claim that semantic properties play an essential role in the individuation of physical computing systems such as laptops and brains. The main argument for the semantic view rests on the fact that some physical systems simultaneously implement different automata at the same time, in the same space, and even in the very same physical properties. Recently, several authors have challenged this argument. They accept the premise of simultaneous implementation but reject the semantic conclusion. In this paper, I aim to explicate the semantic view and to address these objections. I first characterize the semantic view and distinguish it from other, closely related views. Then, I contend that the master argument for the semantic view survives the counter-arguments against it. One counter-argument is that computational individuation is not forced to choose between the implemented automata but rather always picks out a more basic computational structure. My response is that this move might undermine the notion of computational equivalence. Another counter-argument is that while computational individuation is forced to rely on extrinsic features, these features need not be semantic. My reply is that the semantic view better accounts for these extrinsic features than the proposed non-semantic alternatives.

An underlying assumption in computational approaches in cognitive and brain sciences is that the nervous system is an input–output model of the world: Its input–output functions mirror certain relations in the target domains. I argue that the input–output modelling assumption plays distinct methodological and explanatory roles. Methodologically, input–output modelling serves to discover the computed function from environmental cues. Explanatorily, input–output modelling serves to account for the appropriateness of the computed function to the explanandum information-processing task. I compare very briefly the modelling explanation to mechanistic and optimality explanations, noting that in both cases the explanations can be seen as complementary rather than contrastive or competing.

Theodore Sider distinguishes two notions of global supervenience: strong global supervenience and weak global supervenience. He then discusses some applications to general metaphysical questions. Most interestingly, Sider employs the weak notion in order to undermine a familiar argument against coincident distinct entities. In what follows, I reexamine the two notions and distinguish them from a third, intermediate, notion (intermediate global supervenience). I argue that (a) weak global supervenience is not an adequate notion of dependence; (b) weak global supervenience does not capture certain assumptions about coincidence relations; (c) these assumptions are better accommodated by the stronger notion of intermediate global supervenience; (d) intermediate global supervenience, however, is also not an adequate notion of dependence; and (e) strong global supervenience is an adequate notion of dependence. It also fits in with anti-individualism about the mental. It does not, however, serve to rebut arguments against coincident entities.

Jaegwon Kim contends that global supervenience is consistent with non-materialistic cases. Paull and Sider, Horgan, as well as Kim, attempt to defend it from these charges. It is shown here that their defense is only partially successful. Their defense meets one challenge to global supervenience---the hydrogen-atom case---but fails to meet other, ‘local’, cases. It is suggested that the other challenges can be met if global supervenience is combined with weak supervenience. The combination of global and weak supervenience constitutes a viable picture of psychophysical relations, and is especially attractive to nonreductive materialists who are also anti-individualists.

Computational neuroscientists not only employ computer models and simulations in studying brain functions. They also view the modeled nervous system itself as computing. What does it mean to say that the brain computes? And what is the utility of the ‘brain-as-computer’ assumption in studying brain functions? In previous work, I have argued that a structural conception of computation is not adequate to address these questions. Here I outline an alternative conception of computation, which I call the analog-model. The term ‘analog-model’ does not mean continuous, non-discrete or non-digital. It means that the functional performance of the system simulates mathematical relations in some other system, between what is being represented. The brain-as-computer view is invoked to demonstrate that the internal cellular activity is appropriate for the pertinent information-processing task.Keywords: Computation; Computational neuroscience; Analog computers; Representation; Simulation.

It is generally assumed that everything that can be said about dependence with the notion of strong global supervenience can also be said with the notion of strong supervenience. It is argued here, however, that strong global supervenience has a metaphysically distinctive role to play. It is shown that when the relevant sets include relations , strong global supervenience and strong supervenience are distinct. It is then concluded that there are claims about dependence of relations that can be made with the global notion of strong supervenience but not with the “local” (individual) one.

A key component of scientific inquiry, especially inquiry devoted to developing mechanistic explanations, is delineating the phenomenon to be explained. The task of delineating phenomena, however, has not been sufficiently analyzed, even by the new mechanistic philosophers of science. We contend that Marr’s characterization of what he called the computational level provides a valuable resource for understanding what is involved in delineating phenomena. Unfortunately, the distinctive feature of Marr’s computational level, his dual emphasis on both what is computed and why it is computed, has not been appreciated in philosophical discussions of Marr. Accordingly we offer a distinctive account of CL. This then allows us to develop two important points about delineating phenomena. First, the accounts of phenomena that figure in explanatory practice are typically not qualitative but precise, formal or mathematical, representations. Second, delineating phenomena requires consideration of the demands the environment places on the mechanism—identifying, as Marr put it, the basis of the computed function in the world. As valuable as Marr’s account of CL is in characterizing phenomena, we contend that ultimately he did not go far enough. Determining the relevant demands of the environment on the mechanism often requires detailed empirical investigation. Moreover, often phenomena are reconstituted in the course of inquiry on the mechanism itself.

Are all three of Marr's levels needed? Should they be kept distinct? We argue for the distinct contributions and methodologies of each level of analysis. It is important to maintain them because they provide three different perspectives required to understand mechanisms, especially information-processing mechanisms. The computational perspective provides an understanding of how a mechanism functions in broader environments that determines the computations it needs to perform. The representation and algorithmic perspective offers an understanding of how information about the environment is encoded within the mechanism and what are the patterns of organization that enable the parts of the mechanism to produce the phenomenon. The implementation perspective yields an understanding of the neural details of the mechanism and how they constrain function and algorithms. Once we adequately characterize the distinct role of each level of analysis, it is fairly straightforward to see how they relate.

Is the mind/brain a kind of a computer? In cognitive science, it is widely believed that cognition is a form of computation--that some physical systems, such as minds/brains, compute appropriate functions, whereas other systems, such as video cameras, stomachs or the weather, do not compute. What makes a physical system a computing system? In my dissertation I first reject the orthodox, Turing-machine style answer to this question. I argue that the orthodox notion is rooted in a misunderstanding of our pre-theoretic notion of computation and of Turing's characterization of it. I then offer an alternative--semantic --theory of computation for physical systems. I suggest that to view a system as a computing system is to identify its processes and states, as computational, with respect to their semantic relations to external objects. Lastly, I examine the ramifications of my theses about computation for cognitive science. I argue that the level at which we specify psychological processes/mechanisms is defined over semantic, rather than syntactic or algorithmic, types. As a result of this, I go on to claim that cognitive scientists take semantic properties as those which explain behavior, not those which are in need of explanation

There is an intensive discussion nowadays about the meaning of effective computability, with implications to the status and provability of the Church–Turing Thesis (CTT). I begin by reviewing what has become the dominant account of the way Turing and Church viewed, in 1936, effective computability. According to this account, to which I refer as the Gandy–Sieg account, Turing and Church aimed to characterize the functions that can be computed by a human computer. In addition, Turing provided a highly convincing argument for CTT by analyzing the processes carried out by a human computer. I then contend that if the Gandy–Sieg account is correct, then the notion of effective computability has changed after 1936. Today computer scientists view effective computability in terms of finite machine computation. My contention is supported by the current formulations of CTT, which always refer to machine computation, and by the current argumentation for CTT, which is different from the main arguments advanced by Turing and Church. I finally turn to discuss Robin Gandy's characterization of machine computation. I suggest that there is an ambiguity regarding the types of machines Gandy was postulating. I offer three interpretations, which differ in their scope and limitations, and conclude that none provides the basis for claiming that Gandy characterized finite machine computation.