This is a book about minds. It is also about computers. Centrally, we will be interested in examining the relation between minds and computers.

The idea that we might one day be able to construct some artefact which has a mind in the same sense that we have minds is not a new one. It has featured in entertaining and frightening fictions since Mary Shelley first conceived of Frankenstein's monster.

In the classic science fiction of the early to mid-twentieth century, this idea was generally cashed out in terms of 'mechanical men' or robots - from the Czech word robata, which translates roughly as the feudal term corvée, a term which refers to the unpaid labour provided to one's liege lord.

In more modern fiction, the idea of a mechanical mind has given way to the now commonplace notion of a computational artificial intelligence. The possibility of actually developing artificial intelligence, however, is not just a question of sufficiently advanced technology. It is fundamentally a philosophical question.

It is this question that we will be centrally concerned with throughout this volume. In order that we might be in an informed position to consider the possibility of artificial intelligence, we will need to answer a number of related questions.

Firstly, we will be asking just what the human mind is. The twentieth century saw a succession of philosophical theories of mind, culminating in the currently dominant theory which accommodates the possibility of artificial intelligence. Our first goal, which we will spend Chapters 1-10 pursuing, is to clearly articulate this theory.

Philosophically responsible engagement with this theory requires a sound understanding of precisely what a computer is. Consequently, we're going to spend three chapters developing a rigorous technical account of computation. Although this material is technical, the introduction is slow and gentle and will be readily accessible to a reader with no background in mathematics or computer science.

Along the way to our target theory, we're going to survey the space of available philosophical theories of mind, weighing the merits and flaws of each. This will provide a comprehensive introduction to the philosophy of mind.

We are also going to take a couple of empirical diversions along the way. We're going to tell the story of the rise of empirical psychology and we will spend a chapter developing a rudimentary understanding of functional neuroanatomy.

Once we are armed with a sound philosophical understanding of our target theory, the remainder of the book will be given to evaluating it. We will see that a wide range of material from the empirical disciplines bears importantly on the tenability of the theory. As such, this book is overarchingly an exercise in cross-disciplinary analysis.

We're going to focus on two mental capacities that are distinctively human - our capacity for reasoning and our facility for language. Our aim first of all in Chapters 11-20 will be to compare what we know of the human rational and linguistic capacities with methods for implementing these computationally.

We will see how computers can be programmed for strategic game play, for reasoning about novel situations based on known information and for certain functions implicated in language production and comprehension. This will expose us to some introductory material in linguistics and a tiny bit of formal logic, and we will touch on some material from cognitive psychology.

In the final chapters of the book we will examine some more advanced philosophical material concerning the notions of meaning and representation. Lastly, we will introduce artificial neural networks and see how they can be employed in the pursuit of artificial intelligence - again with particular respect to rational and linguistic functions.

All told,we will be examining material from philosophy, psychology, linguistics, neuroscience and computer science - the disciplines which constitute cognitive science. It is to be expected that most readers will find some of this material more approachable and some less so; however, I have aimed for maximal accessibility to the introductory reader throughout.

Each chapter from Chapter 11 onwards engages with an issue which by all rights deserves a dedicated volume. As such, the coverage is less than comprehensive and I have frequently simplified explanations in the name of accessibility. There are suggestions for further reading at the end of the book for readers who want to further their understanding of the issues we cover.

Comprehensive coverage of the relevant issues, however, is not our primary concern here. Our main aim is to develop and evaluate the philosophical theory of mind which allows for the possibility of artificial intelligence. By the end of the book, the reader should find themselves in a sound position from which to make informed decisions concerning the possibility of developing artificial intelligence. This book should also provide a solid foundation for philosophically responsible engagement with cognitive science broadly.

We're now going to begin our tour of the space of available philosophical theories of mind with a theory which most people implicitly, and pretheoretically, subscribe to: dualism.

We are going to begin our examination of the available theories of mind with Cartesian dualism. There are at least two good reasons for doing so. One is that the presentation of theories of mind in the following chapters will be broadly chronological and - at least as far as modern philosophy is concerned - to begin with Cartesian dualism is to begin at the beginning.

Substance dualism is a metaphysical view. It is the view that the universe consists of two different kinds of stuff - two metaphysically distinct substances. As such, substance dualism is a commitment to a particular ontology - an ontology which sees the universe as comprised of both material and immaterial substances. As well as all the material stuff which makes up the physical world, the dualist holds that there is also non-physical, immaterial stuff to be taken account of.

Cartesian dualism is a view about the mind, the body and the relation between them. It is a particular kind of dualism which takes its name from its original proponent, René Descartes. It is essentially the view that while the body is a material object, the mind is not. According to the Cartesian dualist, the mind is composed entirely of immaterial stuff. As such, Cartesian dualism is clearly a kind of substance dualism as the Cartesian dualist is committed to an ontology which admits of both material and immaterial substances.

There are a number of reasons why one might endorse Cartesian dualism. I will consider the three strongest arguments in its favour: the argument from religion, the argument from introspective appearance and the argument from essential properties.

We have now seen three arguments in favour of dualism. The argument from introspective appearance and the argument from essential properties both seek to establish a broad mind-body dualism. Coupled with certain common-sense intuitions concerning the efficacy of the mind in bringing about changes in the body and vice versa, they become arguments supporting Cartesian dualism. The argument from religion seeks to establish Cartesian dualism in particular, as the interaction between mind and body is essential for the argument.

Recall from section 2.2 the four propositions [D1]–[D4] which characterise Cartesian dualism. One way to recover the core ontological intuitions of Cartesian dualism from the damning criticism of the problem of interaction is to give up the commitment to propositions [D3] and [D4], leaving us in want of an account of the relation between the physical body and the non-physical mind. This strategy leads to the theistic dualist theories known as parallelism and occasionalism.

A final theory we should at least mention before concluding this chapter is anomalous monism, otherwise known as double aspect theory or, simply, property dualism.

The next theory of mind we’re going to examine is philosophical behaviourism. Before we do so, however, it will serve us to take a short detour into the history of psychology.

The treatment of the history of psychology here will be rather cursory. For our purposes, it serves to identify a few key figures and seminal contributions which led to the birth of psychology as a distinct academic speciality. Roughly and broadly speaking, we can divide the infancy and early childhood of psychology into three stages, distinguished by the methodologies employed by the discipline’s progenitors.

Although we will reserve the title of ‘founder of psychology’ for another, Gustav Fechner (1801–87) must be credited with the incep- tion of the empirical tradition in psychology and the delivery of the first quantitative psychological law.

It is Wilhelm Wundt (1832–1920) who clearly deserves the appellation ‘founder of psychology’. Wundt established the first psychological laboratory, inaugurated the first journal of psychology – Philosophische Studien – in 1881, and founded an Institute for Experimental Psychology at Leipzig in 1894. He also wrote the first textbook in psychology and supervised legions of graduate students from around the world who would become the first generation of psychological practitioners.

Aside from the foundational work of the physiological and introspectionist psychologists, there are two further important historical preconditions which led to the emergence of psychological behaviourism.

Psychological behaviourism, as we have seen, is a methodological view – a doctrine concerning the way in which one should go about doing psychology. Philosophical behaviourism, in contrast, is an analytic view – a substantive theory of what mental states are. Henceforth, when I make reference simply to ‘behaviourism’, I will be referring to the philosophical variety.

We’ll begin with the three objections which merely problematise behaviourism and move on to the further three objections which are insurmountable for the behaviourist.

We’re now going to take a brief diversion from our examination of philosophical theories of mind and develop a rudimentary understanding of functional neuroanatomy.

The human central nervous system can be broadly divided into three areas. The spinal cord, the brain stem, and the rest of the brain, including the cerebral hemispheres which constitute the cerebrum.

The final aim of this chapter is to briefly describe the operations of neurons. Neurons are individual nerve cells which conduct electrical impulses and the brain consists of a very large number of them.

Now that we have at least a rudimentary understanding of just what an amazing thing the human brain is, it is time to examine a philosophical theory which posits a very strong connection between the neural and the mental.

The canonical exposition of the causal theory is given by David Armstrong in his 1968 monograph, A Materialist Theory of Mind. Armstrong, together with J.J.C. Smart and U.T. Place, is one of the three major figures associated with Australian materialism.

Australian materialism rose to prominence in the late 1950s with the publication of two very influential papers: Place’s Is Consciousness a Brain Process? (1956) and Smart’s Sensations and Brain Processes (1959).

Let’s begin with some fairly weak objections to Australian materialism. We might argue that we have the capacity to introspect our mental states and that when we do so, we learn about our mental states. We don’t, however, learn anything about our neurophysiology through introspection, so mental states can’t be identical to neural states.

Thought experiment plays an important role in the philosophy of mind. Since this is the first time we are seeing one in this volume, it is worth very briefly discussing their role.

There is very little explanatory work to be done in this chapter and, consequently, it will be comparatively short. The reasons for this are twofold.

We begin – as we did with the causal theory – by reflecting on the defining characteristics of certain terms.

In order that we fully appreciate the structure of the theoretical framework, it is useful to represent the three levels of description – and the identities that obtain among them – diagrammatically (see Figure 6.1).

The two objections we will raise here target not any particular kind of functionalism but, rather, the claim at the heart of the functionalist framework. These are objections to the contention that there is nothing more of importance to know about mental states beyond their function and that carrying out such a function is sufficient for something being a mental state.

In the previous chapters, we have considered the question of what minds might be and sketched out the space of possible responses to this question. In doing so we have seen a progression of philosophical theories of mind and considered arguments and objections pertaining to each.

It is highly likely that every reader of this book has at some stage in their life played a game of at least one of the following: chess, draughts, backgammon, go, Chinese checkers or – at the very least – tic-tac-toe (aka noughts and crosses). If you understand how at least one of these games is played (most of us can grasp tic-tac-toe), then regardless of how good or bad you are in playing them, you already understand the principles underlying formal systems. We’ll begin our examination of formal systems by simply making explicit what you already grasp implicitly.

Formal systems are composed of two collections: a collection of states and a collection of rules. The specification of any given formal system consists in the specifications of its states and of its rules.

We want to discuss the properties and operations of formal systems, so let’s define a symbol system to play with. Let’s call it [STR] – it will be a string system.

The operations of formal systems consist in successive applications of rules to states. Given an initial state, we can help ourselves to a distinction between states which will arise during the operations of the system, and those which, while they meet the criteria for possible states, never actually arise during the operations of the system.

The answer to Exercise 7.4(b) brings out an important feature of the system [STR] – that it has no terminal states.

There is one last point to make concerning formal systems before we move on and do something more interesting with them. It is important to appreciate that the only important or relevant properties of formal systems are formal properties – properties of form.

The formal systems we have looked at so far have been very rudimentary string systems. Consequently, their useful application is rather limited. We can, however, employ formal systems to rather more interesting and useful ends. In particular, we can use formal systems to do computation. In this chapter,we are going to use a particular kind of deterministic formal system – a register machine – to rigorously define computability.

Register machines are theoretical entities. They can, however, be physically implemented (as can any formal system). Modern digital computers as we know them are implementations of a special kind of register machine, as we will see in the following chapter.

Register machines are formal systems. We now know what register machine states are – for all intents and purposes they are simply finite sequences of natural numbers. We next need to know what the rules are.

Now that we know what register machine states are and what it is to run a register machine program, let’s examine the operations of a simple register machine. Consider the following program:

Having completed Exercise 8.1 you would have noticed that the program [ADD] always terminates with a number in R₁ which is the sum of the numbers which were in R₁ and R₂ in the initial state. In other words, the program [ADD] adds the numbers in R₁ and R₂ and terminates with their sum in R₁.

We have so far introduced computation informally as the operations of a register machine program. We have also seen an example of something which is not computable. There is no effective procedure (a fortiori no register machine program) which will determine whether a given program will halt. Consequently something which cannot be computed is the question of whether a given computation will halt.

In the remainder of this chapter we will be developing methods for constructing register machine programs to implement algorithms. We are going to write programs to compute various functions. In each case, I will set an exercise and then work through a possible solution. Attempting the exercises before reading the solution will aid significantly in consolidating your understanding of this material.

This chapter completes our exposition of computational theory. So far, we have discussed formal systems broadly and we have used a special kind of formal system – a register machine – to define computation and computability.

The halting problem is a particular instance of the Entscheidungsproblem – or decision problem – which was of interest to mathematicians and logicians well before there was a formal theory of computation. The decision problem for a particular formal system refers to the question of whether or not there is an effective procedure for determining, of any given state of the system, whether or not it is generated in the system. If there is such a procedure, the system is said to be decidable.

We have seen that for any algorithm one can specify, we can design a register machine program to compute it. The final aim for this chapter is to specify a single register machine program which can itself compute any algorithm. In other words, we want to specify a register machine program which can emulate or instance any other register machine program.

Armed with the method of Gödel coding, we are now ready to describe the register machine program which can compute any algorithmic function.

So far, we have considered the question of what minds might be and in doing so, have examined a number of philosophical theories which aim to provide an answer to this question. We suspended that discussion after introducing functionalism and turned our efforts to developing a rigorous account of what computation is. It is now time to pick up where we left off.

Computationalism is the view that to have a mind is to instantiate a particular formal system or collection of systems. In other words, since mental operations are held to be the operations of formal systems, mental operations are held to be computations. So to have a mind, claims the computationalist, just is to be engaged in certain computational processes.

Computationalism is often described as a ‘software’ view of the mind. The human brain is seen as providing the biological computational hardware – or wetware – which confers on humans the capacity to have a mind. Having a mind, on this view, is a matter of having the right program running in one’s wetware.

The objection from variation runs as follows. Computationalism says that humans have minds by virtue of implementing [MIND], but human minds vary greatly. How can this be, given all minds are held to be isomorphisms of the same formal system?

Minds can learn. Consequently, different minds can do different things. In other words, some minds can perform functions which other minds cannot (or can perform certain functions better than most other minds). How can computationalism account for this?

Another standard prima facie objection appeals to the human creative capacity, as follows. The operations of formal systems are entirely mechanical but minds are creative. Minds create great works of art, music, architecture and literature, and have an enormous capacity to innovate. This characteristic creativity of human minds seems to be compelling evidence against computationalism which seeks to account for mentality in terms of purely mechanical operations.

How do we determine whether or not something has a mental life? When we are not in the grip of philosophical scepticism, most of us are quite convinced that those around us have mental lives – at the very least we habitually act as if they do. Given that we have privileged access to only our own mental states, what is it about other people that leads us to believe that they have minds?

Given the computationalist hypothesis as we have described it, investigating mentality involves investigating the operations of the formal system [MIND] and its constituents.

One method of determining whether a state of a formal system is generated, or finding a derivation for a particular state of the system, is to construct the entire generation tree for the system. If the state we are interested appears on the tree it is a generated state and we can read off its derivation(s) by following the branches back up to the root node.

Having decided between a top-down search and a bottom-up search, we then have to further decide on a procedure for searching the tree. Keep in mind that any formal system of sufficient interest for artificial intelligence researchers to be investigating will be considerably more complex than the toy examples we are looking at here. Consequently, choosing an appropriate search procedure can have a significant impact on the computational resources required to carry out the search. In fact, as we shall see, it can mean the difference between a successful search and a search which never halts.

Breadth first and depth first are both what we call blind searches – they are conducted without any consideration of the closeness to solution of the nodes being searched.

Now that we have a basic understanding of search procedures and heuristic functions, it is time to apply this understanding in consideration of automated methods of game play.

Ifa game is sufficiently simple, we can very easily automate a procedure for playing it. It is fairly trivial,for instance,to determine an algorithm for playing tic-tac-toe such that following the algorithm will always result in either a win or a draw.

We saw in section 11.3 that we can apply a heuristic function to a node which evaluates, according to criteria relevant to the system, the closeness to solution of that node. In the context of a two-player game, the states which will count as solution states are those representing wins for (an arbitrary) one ofthe players. In this case,the furthest a game could be from a solution state would be a state representing a win for the other player.

We now have the makings of a procedure for searching for strategies for playing a game as complex as chess. We search as far ahead as is computationally tenable, apply a heuristic function to the nodes at the search horizon, then use a minimax procedure to work back to the node we are searching from.

We have now given some consideration to an activity which is generally taken as constitutive of intelligence – the ability to (at least learn to) play well a complex game such as chess. We’ve seen – at least in broad strokes – how to employ symbol systems and search methods to automate the strategic play of such games.

In this chapter we are going to begin to investigate the rational capacity – the ability to reason. A thorough survey of automated reasoning methods would require dedicated volumes. We’re going to concentrate on one kind of automated reasoning project which suits our purposes well – the design of expert systems.

The first thing to appreciate is that there is a distinction between logic and logics. Logic is a research tradition whose objects of investigation are logics. These logics are formal systems and there are very many of them.

Natural language conditionals are statements of the form ‘if . . . then. . .’. The study of conditionals, and the determination of an adequate formal account thereof, is of central importance to logic. Many logics are distinguished solely by virtue of their treatment of the conditional.

If you were able to follow the example deductions in the previous section, then you already grasp the important aspects of the operations of expert systems. In fact, the example cases used to introduce predicate notation were actually themselves miniscule expert systems.

The example expert system of the previous section is greatly simplified but, nonetheless, it is clear that it is able to capture the deductive process that we engage in when reasoning about kinship relations.

Devising computational procedures for handling natural language is arguably the most significant problem facing artificial intelligence researchers. In this chapter we’re going to begin by considering the various computational problems which need to be solved in order to facilitate this.

Let’s reflect on the various procedures involved in the comprehension of a spoken utterance.

Noam Chomsky revolutionised the discipline of linguistics in the 1950s by taking a new approach to the study of grammar.

We’re going to specify a phrase structure grammar for a fragment of English. The states of our phrase structure grammar will be finite strings of those symbols which feature in the rules. The initial state will be the symbol ‘S’. The rules of the system are as follows.

Determining the grammaticality of sentences of language according to a generative grammar is clearly a computational procedure. It should also be clear that given a phrase structure grammar and an arbitrary string of its symbols, we can conduct a bottom-up search to determine whether or not the string is generated by a phrase structure tree.

In the previous two chapters, we approached the rational and linguistic faculties with a view to analysing their constituent mechanisms and accounting for these mechanisms in computational terms.

It is a recognised empirical fact that people generally perform quite poorly on this set of reasoning problems. The reader of this volume is clearly more intelligent than average – simply by virtue of having purchased this book if nothing else – but I would still be surprised if you made no mistakes on the problem set (unless you’ve previously been exposed to these problems).

People generally spot the validity of simple inferences, such as in problem 1. This is just modus ponens so it does follow that Mike is happy. Problem 11 also instances a very simple inference form – modus tollens. Given a true conditional with a false consequent, we can always validly infer the falsity of the antecedent. However, it is common to see mistakes on this problem.

It seems we have plenty of evidence that when people reason, they do not ordinarily explicitly follow formal rules. Rather, they construct mental models of the problem situation and interrogate these mental models to determine a solution.

If it were the case that people always reasoned formally, according to the dictates of some logic, then providing a computational account of rational mechanisms would be very straightforward, since logics just are formal systems.

The human linguistic capacity is really quite amazing. The mechanisms which facilitate linguistic production and comprehension are surprisingly complex given that our capacity for language is so natural to us as to appear incredibly simple. No doubt you’ve begun to appreciate just how much cognitive processing is involved in linguistic behaviour after reading Chapter 14.

The study of phonology is the study of the speech sounds and sound patterns of spoken language. Central to the study of phonology is the identification and classification of the phonemes of a given language.

All sonorant phonemes are ipso facto voiced. They also all have the same manner of articulation – they are open sounds. In other words, the passage of air through the articulatory apparatus is not impeded but resonates freely in the oral cavity.

Phonemes are idealisations. Actual speech sounds – phones – approximate to phonemes and may vary significantly between speakers. As language users, we are very good at detecting the distinctive contrasts in speech sounds and assimilating phones to phonemes.

We can also find plenty of evidence of rule-governed linguistic activity in the first-language acquisition literature.

While we’ve mostly concentrated in this chapter on phonological processes, other areas of linguistics are also rich with examples of rule-governed behaviour.

This chapter marks a return to philosophical material after six chapters of technical material.

It is a crucial and defining feature of our mental states that they have semantic content – that they are meaningful states. Any adequate theory of mind must be able to account for the semantic contents of mental states.

On the strength of the Chinese Room thought experiment, we might be tempted to mount this argument against computationalism:

Our mental states are meaningful by virtue of being about things. In other words, meaningful mental states are representational states – they represent or stand for things. In previous chapters I’ve made reference to mental representations, such as the phonemic, syntactic and – crucially – semantic representations which facilitate linguistic production and comprehension. In this chapter, I want to briefly discuss the structure and nature of mental representation.

Intentionality is the technical philosophical term for the representational nature of mental states. Intentional states are those which are about something, which represent something.

To give an account of the relation between mental representations and their intentional objects is to give part of a semantics for mental representation. There are numerous theories of the semantics of mental representation but I’m not going to give a balanced exposition of the available theories here. Instead, I want to give just the barest sketch of two kinds of theories.

Theorists who endorse an account of mental representations as distributed patterns understand representations to be essentially interrelated. The semantics of mental representation that such a theorist will advance are such that the mechanism by which content is conferred on a representation is essentially mediated by relations to other representations.

So far in this chapter I’ve discussed two distinct views of mental representation and used this distinction as an entryway into understanding the competing symbolic and connectionist paradigms in artificial intelligence research.

The connectionist paradigm in artificial intelligence research rose to prominence in the last two decades of the twentieth century. Artificial neural networks were shown to be quite efficacious in modelling certain cognitive phenomena that had been problematic to implement with symbolic computational architecture.

Classical symbolic computational architecture – which we described at length in Chapters 7 to 9 and have seen many examples of since – admits of the following essential features.

Let’s take a look at some basic examples to exemplify these operations. To keep things simple, I’m going to use integers for connection weights and threshold values. Figure 19.1 depicts the simplest artificial neural network that does something interesting.

In this section we’re going to design an artificial neural network to function as an English speech synthesiser. This is a nice example of the kind of contextually sensitive processing tasks at which connectionist networks excel.

You should now have a sense of precisely how complex a processing task it is to convert English orthography to phonemics. We’ve considered only one phoneme and only a tiny fraction of relevant cases and even this quickly became quite a complex task.

In our described example case of training an artificial neural network to translate orthography to phonemics, the network learns how to map orthographic contexts to phonemes by learning to recognise certain patterns.

Although I’ve described the symbolic and connectionist approaches to artificial intelligence as fundamentally distinct – and, by implication, incommensurate – paradigms, it may well be the case that these views concerning information processing merely engage at different levels of description.

The introduction to artificial neural networks in this chapter has been very basic indeed. We’ve considered only the simplest kinds of networks and functions in order to avoid unnecessary mathematical complexity. I’d strongly recommend that the interested reader continue their investigations with the suggestions for further reading as a guide. A proper introduction to artificial neural networks requires a dedicated textbook.

We have now learned a lot about minds, having surveyed the space of available philosophical theories of mind and considered the advantages and disadvantages of each theory.

Although I’ve helped myself in places to an intuitive distinction between the mental processes we are consciously aware of and those which occur below the level of consciousness, I’ve not said much at all about consciousness per se.

On any given day, I clearly differ in a number of ways from the way I was the day before since I will have a number of different properties. I will have a distinct spatiotemporal location, I may have altered or augmented beliefs, I will have extra memories, small bits of my body – skin, hair, fingernails and the like – will have been lost and new bits will have grown, and so on.

It is generally considered that to lack the capacity for emotional states and responses is to lack one of the requirements for having a mind in the sense that we have minds. A deficit in emotional behaviour is one of the characteristic symptoms of certain psychopathologies and, while we hold such people to have minds, we believe their minds to be importantly qualitatively distinct from our own.

Now that we’ve reached the end of the book, it is time to reflect on what determinations we are able to make concerning the possibility of artificial intelligence.