(WORK IN PROGRESS: Liable to change)
A partial index of discussion notes is in
http://www.cs.bham.ac.uk/research/projects/cogaff/misc/AREADME.html
Most philosophers appear to have ignored the distinction between the broad concept
of Virtual Machine Functionalism (VMF) described in Sloman&Chrisley (2003) and the
better known, more restricted, version referred to there as Atomic State Functionalism (ASF),
which is often given as an explanation of what Functionalism is, e.g. in Block (1995).One of the main differences is that ASF encourages talk of supervenience of states
and properties, whereas VMF requires supervenience of machines that are arbitrarily
complex networks of causally interacting (virtual, but real) processes, possibly operating
on different time-scales, examples of which include many different processes usually
running concurrently on a modern computer performing various tasks concerned with
handling interfaces to physical devices, managing the file system, dealing with security,
providing tools, entertainments, and games, and possibly processing research data. Another
example of VMF would be the kind of functionalism involved in a large collection of possibly
changing socio-economic structures and processes interacting in a complex community, and
yet another is illustrated by the kind of virtual machinery involved in the many levels of
visual processing of information about spatial structures, processes, and relationships
(including percepts of moving shadows, reflections, highlights, optical-flow patterns and
changing affordances) as you walk through a crowded car-park on a sunny day: generating
a whole zoo of interacting qualia. (Forget solitary red patches, or experiences thereof.)Perhaps VMF should be re-labelled "Virtual MachinERY Functionalism" because the word
'machinery' more readily suggests something complex with interacting parts. VMF is
concerned with virtual machines that are made up of interacting concurrently active (but
not necessarily synchronised) chunks of virtual machinery which not only interact with
one another and with their physical substrates (which may be partly shared, and also
frequently modified by garbage collection, metabolism, or whatever) but can also
concurrently interact with and refer to various things in the immediate and remote
environment (via sensory/motor channels, and possible future technologies also). I.e.
virtual machinery can include mechanisms that create and manipulate semantic content[*],
not only syntactic structures or bit patterns as digital virtual machines do.[*] This semantic content is information in Jane Austen's sense, as explained here.
Note:
As P. F. Strawson pointed out (Individuals: An essay in descriptive metaphysics, 1959)
some of the semantic (intentional) relations may be partially implemented in causal
connections between things in the environment, including long dead things, and also, as
every mathematician knows (including Plato), structures processes and events on paper,
blackboards, in sand, and now computing devices outside the thinker too.Keywords:
asynchronous concurrent causation, atomic state functionalism, definability,
information processing, interfaces to environment, operating system, physics,
qualia, representation, self monitoring, virtual machine functionalism,
virtual machine supervenience,
NOTE ON TERMINOLOGY
This document does not use standard terminology for varieties of functionalism discussed,
since, as far as I know, the need to distinguish these varieties has not generally been
acknowledged, and the importance of the most complex type, Virtual Machine Functionalism
has not generally been recognised, though Rickles-IEP comes close.
So I have created my own labels for the cases I distinguish. I'll be happy to be informed
that the cases have already been described elsewhere and there are alternative labels in
use.
I am
not here talking about virtual machines that are surrogates for real machines in
that sense! Rather the sorts of virtual machine that I discuss include some that could not
exist in anything but a virtual machine form (i.e. implemented in some kind of lower level
machine) because the machine requires a kind of flexibility of structure that would either
be impossible, or too slow or too costly in a physical machine. For example a planning
program that creates and tests and tries to extend partial plans builds the plans in
software. Building them by creating new machinery for each plan fragment in order to check
its feasibility would be either too costly and slow, or may not even physically possible
given the constraints on the system that is to create and use the plans.
In the most interesting cases, discussed in more detail below, there are VMs whose
specification uses concepts that are not definable in the language of physics. So it is
not possible to go directly from the specification to a physical design: instead a more
complex process is required, of working out how to make something physical that performs
the required functions. That sort of engineering design process produces an implementation
of the design. In general there will be many possible physical implementations for a
machine required for a biological function and there are deep and interesting questions
about how we should describe the relationship between such a physical machine and the
virtual machine it implements. I'll return to that later.
For the rest of this paper, I shall consider only virtual machines for which there need
not be any non-virtual alternative that is practical. I believe that biological evolution
"discovered" the need for such things long before we did. From now on I'll use "virtual
machine" to refer to such indispensable information processing mechanisms. But I want to
further restrict the class of virtual machines under consideration to machines that could
not be fully specified in the language of physics even if all the working instances are
physical.
So, a virtual machine (VM) is a machine that does something by making use of physical
mechanisms, but whose states, processes, and functions are not defined using physical
concepts. Some aspects of that idea are very old but the idea has been considerably
enriched and has developed in new precise forms as a result of problems that had to be
solved in designing increasingly complex computing systems. Virtual machines on computers
include things like spelling checkers, web browsers, email systems, anti-virus systems,
and operating systems (e.g. Linux, MacOS, Windows, Solaris, and many others). The internet
is a very complex mixture of physical machines and virtual machines made of networks of
smaller physical and virtual machines.
I'll explain what I mean by "virtual machine functionalism" (VMF) as a theory about the
nature of minds, after introducing simpler, more familiar, variants of functionalism in
philosophy of mind, with which it can be contrasted. Almost everything I've read about
functionalism as a theory of mind has failed to allow for the possibility of VMF, even
though the key concepts and design techniques are already familiar to computer systems
engineers, having been developed over many decades to support increasingly complex
applications of computing technology.
In philosophy of mind "functionalism" is a label used in connection with different
theories that attempt to explain how minds, along with mental states, and mental
processes, can exist in a physical universe, there are several different forms, some of
them presented in http://en.wikipedia.org/wiki/Functionalism_(philosophy_of_mind)
A very simple example is to think of hunger as a state in which if the input is
information about the possibility of obtaining food by performing one of a set of possible
actions, or information from which the machine can infer such a possibility, then in the
hunger state that input will trigger an output that causes the machine to consume the
food, or, if that's not possible in the current situation, to take an action that brings
it closer to consuming the food. For this concept of hunger, it does not matter what the
actual mechanisms are inside the machine, as long as they operate in such a way that such
a dispositional state exists. More sophisticated variants can allow the current state to
include more than one desire or need, each of which can have a tendency to cause particular
actions to occur (actions that meet the need), and in that case what actually happens will
depend either on the relative strengths of the competing desires/needs, or some rule in
the system for taking control when a conflict exists.
I call that "Atomic state functionalism" because each state is a unit (though it may have
components, like a vector of numbers representing coordinates in a space).
Figure ASVM, below, crudely illustrates Atomic State Functionalism, where each state is
labelled by a letter ("a". "b", "c", ... etc.). The diagram represents the functionality
of a machine that cycles between reading its input and then switching state and possibly
producing some output, in accordance with a fixed set of rules: essentially a Turing
Machine, with some input output devices that can copy symbols from the tape to a motor
controller or copy symbols from a digitised sensory device to the tape. The diagram
deliberately leaves open the mode of implementation of the virtual machine in a physical
machine. However, during the last eight or so decades, suitable mechanisms have been
designed and built with steadily increasing sophistication (including reducing physical
size, cost, power consumption, and increasing speed and number of possible states).
Figure ASVM
[An Atomic-State Virtual Machine running on a Physical Machine.]
For example, if an organism is very hungry and slightly thirsty and believes there is food
in front of it, that may trigger a transition to a state in which the hunger is decreased,
the thirst is increased (e.g. because the food was salty) and there is no longer a belief
that food is available. A different input might have led to a transition in which thirst
was decreased and hunger left unchanged.
What atomic state functionalism and molecular state functionalism have in common is the
notion that the transitions that occur form a single (linear) sequence of simple or
compound states and the performance of an action is synchronised with a state transition.
So Figure ASVM could be modified to represent a molecular state machine simply by
replacing some of the letters labelling states, with groups of letters labelling
sub-states, and allowing some letters to occur in several different groups.
There are several further subdivisions between varieties of functionalism that have been
made in the philosophical and cognitive science literature, some of them presented in this
Wikipedia page:
http://en.wikipedia.org/wiki/Functionalism_(philosophy_of_mind)
More detailed accounts are given in the Stanford Encyclopaedia of Philosophy:
http://plato.stanford.edu/entries/functionalism/
Additional references are included below.
The focus shifts from changing states of the individual, or properties of the individual,
to multiple internal machines (or "sub-machines"), where a machine is a collection of
enduring entities and relationships including causal relationships in which processes of
various kinds can occur, and which can have causal interactions with other machines with which
they are in some way connected, or with which they may overlap (like two people sharing an
appointments book, or a note-pad for example).
There can be many, even changing numbers of, sub-machines constituting such a virtual
machine, and many of the sub-machines may themselves be made up of (sub-)sub-machines,
some of them switching between discrete states, others changing continuously. Some of the
sub-machines may be partly within the individual, partly external, providing sensory or
motor interfaces to the environment, or sensory-motor interfaces, such as hands that have
to acquire some information by manipulating objects, and eyes that constantly change
direction of gaze. One rather abstract depiction of such a multi-component virtual machine
is in Fig MultiVM, below.
--
Figure MultiVM
(A Multi-component Virtual Machine running on a Physical Machine.)
Sensory information streams in asynchronously (i.e. not waiting for the next 'read input'
instruction) through many channels, triggering responses in internal concurrently active
virtual machines to which the sensors concerned are linked. Additional streams of
information can fan out to various other machines (which may or may not ignore some of
what they receive).
The contents of information streams need not be restricted to scalar values (like
currents, voltages, or utility measures) or to bit patterns, but could include, for
example, logical expressions, instructions, descriptions, image fragments, diagrams, and
in some cases complex structures such as molecules or information encoded in molecules (as
in genes).
In addition, some of the sub-machines may create local information stores whose contents
are not immediately transmitted anywhere else, but can be accessed by other sub-systems as
required. The information stores may be of many different kinds varying according to
whether they include factual information or control information, general specifications or
instance information, detected or inferred regularities or records of contents of
particular space-time regions (sometimes called "episodic memory"), explanatory theories,
predictions awaiting testing, "compiled" or learned procedures for rapid performances,
grammars for internal languages, and many more.
A system composed of multiple concurrently active, interacting virtual machines with
multiple external interfaces, also concurrently active, is depicted crudely in Fig MultiVM.
A different view is provided in Fig RobotVM, below, emphasising that perceptual and action
subsystems can include information-processing sub-systems and not just physical devices.
In the system depicted by the above figure (MultiVM), the set of virtual (sub-)machines
need not be fixed. The red polygons are intended to represent states of sub-machines in
which they can spawn new virtual machines (also setting up communication between them
and between the new machines and old machines, or sensory motor mechanisms). A system
in which existing machines can spawn new machines as required can vary its complexity as
required to deal with new problems, or new sub-problems of current problems. (In computers this
is commonplace, insofar as running procedures can invoke other procedures by creating new
activation records for them, and that process can recurse. So the number of currently active
procedures is unbounded in principle, though physical resource limits (or the size of the
universe) can impose a limit in practice. (Note that computers, including domestic desk-top
PCs increasingly make use of multi-core CPUs in addition to parallel external interfaces
and device controllers, extending possibilities for changing the number of running processes
as needed).
It is also possible in principle for some of the enduring or temporary VMs to have continuous
state changes e.g. changing measures of compression, temperature, velocity, etc., instead of
only discrete steps as in a Turing machine or the CPU of a typical mono-processor computer.
As James Gibson and others have pointed out, many of the sensory devices are not passive
receptors but active explorers, e.g. hands exploring surfaces of objects, eyes using
saccades and other movements, to select from the optic array. This is crudely depicted in
in "Fig RobotVM", where all the components indicated will have their own internal
state-transitions, and possibly internal architectures composed of yet more VMs.
Figure RobotVM
(A Multi-component Virtual Machine
Linked to specific sensorimotor morphology.)
Different machines can use the information acquired in different ways, e.g. triggering a
motor response (in the case of reflexes) but also triggering goals to be achieved,
questions to be answered, processes of analysis and interpretation to be initiated,
modifications of ongoing processes, 'waking up' dormant machines, decomposing and
analysing new information (e.g. parsing), checking new information against previously
constructed hypotheses, or questions, or goals, and many more.
[Some of these ideas were presented in Chapter 6 of Sloman 1978, online hereIn the case of visual information, various visual sub-machines will concurrently, and
http://www.cs.bham.ac.uk/research/projects/cogaff/chap6.html
Many of the theoretical ideas were later used to design and implement a software
toolkit, developed by the author, students, and colleagues to support research in
multi-component virtual machine architectures for individual agents, as described here:
http://www.cs.bham.ac.uk/research/projects/poplog/packages/simagent.html
Some videos of 'toy' systems developed are in
http://www.cs.bham.ac.uk/research/projects/poplog/figs/simagent ]
The scientific/philosophical task of discovering the variety of types of concurrent interacting
virtual machinery can be informed by introspection, reading good novels, doing experiments,
studying psychology, studying brain mechanisms, trying to build working models to explain
observed animal behaviours (e.g. "Betty" the hook-making New Caledonian crow who made
headlines in 2002) and other things. [REF Elske van der Vaart, 2013]
But trying to unravel such a complex multiply linked, highly, but not entirely,
integrated, mixture of different sorts of machinery is usually intractable, though we can
make progress by combining approaches of different disciplines, and extending
evidence-based information with powerful, testable conjectures about mechanisms. Moreover,
we can try to understand what processes of evolution (supplemented by learning,
development, social processes) could have led to it, by going back to earlier stages and
finding out what transitions occurred: what new mechanisms were added and what problems
they solved, benefits they produced, side effects (possibly dysfunctional) they had etc.
I.e. the design history of current machines can help us understand what's in current
machines and why. (Later insert examples of how the failure to think that way leads people
to ask the wrong questions: e.g. they ask why changes are not noticed in the 'change
blindness' experiments instead of asking about how changes can be detected, or asking
questions about tool-use or 'theory of mind' in young children or other animals, instead
of asking about matter-manipulation and meta-semantic competences.)
I have some online discussions of difficulties in using current computing ideas to
implement some types of biological virtual machinery, e.g. mechanisms used in geometrical
reasoning and some motivational mechanisms. [REFS]
I don't claim the difficulties are insurmountable, though it is worth noting that
biological information-processing makes heavy use of chemical computations that are very
different from manipulations of bit patterns.
However, we have learnt a huge amount in the last half century by encountering problems
that required invention/creation of new forms of virtual machinery, and that's still going
on. That includes creation of VMs whose functions and internal operations are not
definable in the language of physics, even though the machines are implementable in
physical machines.
An example is a chess VM in which it is possible for a threat to be detected and an
attempt made to find a defence against that threat by searching the space of possible
subsequent moves (the game tree). The notions of "detecting", "threat", "defence", "move",
and "searching" as understood by the VM designer, are not definable in terms of notions of
physical particles, mass, velocity, electric charges, currents, voltages, etc., or even
definable in terms of operations of bit patterns in a computer, since designers can come
up with new working implementations of the old strategies, implying that they are not
using an implementation specification, to define the concepts used. If they tried to
produce such a definitional specification using known physical mechanisms (e.g. by
constructing a disjunction of all possible descriptions of currently known types of
physical implementation, using only the language of physics) that could not include
future Chess VMs that use new kinds of computer whose physical states are very
different from those of previously used computers. Moreover, because of the
fine-grained multiple realisability of VMs, any adequate disjunction would probably
be too large to be expressible in this universe, in any physically implemented language.
So, since we do not know how physics will advance, or how new technology based on current
physics will produce new future implementations, the concepts we use now cannot be
equivalent to any disjunction of physical descriptions of implementations.
More discussion of the indefinability claim is needed. Compare Block's "Functional Reduction" paper.
Note:Figure Levels
This amounts to the claim that there are patterns that can exist in physical
structures and processes (including patterns of causal interaction) that may either
exist naturally or be created by us, where the patterns can be described in a
language developed for talking about those structures and processes, whose concepts
require substantive extension beyond the subject matter of the physical sciences:
the new concepts are not definable in terms of old (physical) ones, but have to be
introduced in the context of theories about certain sorts of entities.A familiar example is the concept of an English sentence, such as "The cat sat on
the ancient mat" written in ink or paint or a collection of thumb-tacks on a white
painted wall. It is not possible to specify in general using only the language
of physics what constitutes an instance of that sentence. That would require use of
concepts like "word", "noun", "verb", "subject", "indirect object", "tense", which
(I claim, though I'll not argue here) are not definable in the language of physics.Nevertheless every written or printed instance of that sentence will have a physical
description which could be used by a machine to produce a copy, even if the machine
has no understanding of what words, sentences, cats or mats are. That sentence is a
static structure. In a virtual machine there are not only static structures but also
operations on those structures, including constructing them, modifying them,
interpreting them, correcting mistakes in them, and those VM processes are not
describable in the language of physics, even if their physical instantiations, e.g.
changing patterns on a screen, are.Nevertheless the concepts can be related to concepts of physics, or concepts of
digital electronics, or concepts of computation in a particular programming
language, since system designers can implement special cases of English sentences
and processes that produce, modify or interpret those sentences, The process of
implementation is often very difficult to automate and sometimes new things that are
learnt in the process of generating and fixing bugs can transform the concepts, e.g.
by subdividing cases, or revealing new abstractions. People with no experience of
software design, testing, debugging, and development may find this hard to
understand.)(Note: Need to compare this with Block's argument that Functionalism and Physicalism
are incompatible, in his "Functional Reduction".)
One of the problems about this debate is that it is not clear what counts as
physics, since physics has been extended in very surprising ways (e.g. providing
mechanisms for trans-continental conversations that would probably have been
unimaginable to most of Newton's contemporaries.) That's why I've given physics
different possible lower levels in the following figure illustrating some of the
variety of virtual machinery, natural and human-made.
This argument that we can build physically implemented VMs whose description requires use
of concepts that are not definable in the language of physics, really requires a much more
elaborate, extended discussion. Descriptions of chess playing are a rather simple example.
There are many more examples related to the design not only of games but also many kinds
of
software functionality that people now take for granted, e.g. word processors, spelling
checkers, theorem provers, email systems, tutorial systems, internet mechanisms,
protection against malware, other security and privacy functions, provision of banking
services, remote buying and selling, social networks, and many more.
[Papers on this are on my web site. Use google. But there's still work to be done.]
The task of clarification also includes showing how causation can go upwards, downwards
and sideways in virtual machinery, and how some VMs can have self-monitoring (introspective)
capabilities -- a type of competence that engineers have put into some of their products,
but evolution seems to have provided for organisms very much earlier, and in more complex
forms that are still not understood.
Unfortunately, all that is not yet a standard part of philosophy curricula, so I keep
meeting philosophers who haven't a clue what I am talking about, or assume it must be the
kind of thing they have already studied (e.g. layering of turing machines) and so jump to
inappropriate conclusions. Unfortunately when typing on their word-processors or using
email, or web browsers, or chess programs, they don't ask 'How is this possible?' Or if
they do, they assume wrong, drastically over-simplified answers.
One of the features of virtual-machine functionalism is that it allows some VMs processing
'low level' sensory input to incrementally and collaboratively construct enduring internal
changing interpretations of the information, including references to states and processes
outside the whole system. Other VMs can interrogate, challenge, use, or modify some of
those interpretations. Many of the phenomena that lead to theories about qualia,
sense-data, phenomenal consciousness, can be seen as pointers to such mechanisms.
The parallel flow of information structures, some of which is control information,
specifying what something should do next, can produce some routes and stores that are
transient and constantly being over-written, along with others that are preserved for
various lengths of time, some available for use in social interaction or interaction with
the individual's 'future self' i.e. being available later as 'memories' or 'unfulfilled
goals', or unanswered questions or untested conjectures, etc.
Note: the Sloman/Chrisley 2003 paper enlarges on some of these points, including explaining
how an intelligent agent can develop an internal language for describing its own low level
sensory states in a manner that uses 'causal indexicality' and implies that its internal
descriptions are inherently private -- a possibility not considered by Wittgenstein in his
discussion of the possibility of a logically private language. So it is possible for an
engineer to design a system, in such a way that that system develops concepts the engineer
cannot possibly share, even though she knows quite a lot about them and why they are useful
for her machine. She may be able to make guesses as to some of the characteristics of the
concepts, e.g. how many different colour experiences the robot will learn to distinguish,
and the conditions under which the number can be changed.
Insofar as the entities described by such a private language are contents of the
internally detected and recorded virtual machine states, e.g. results of grouping of
features produced by self-organising networks (e.g. Kohonen nets) they would appear to be
concepts of types of qualia, or sense-data, or phenomenal experience. If that's right, we
have found a way of justifying much philosophical talk that is often taken to be woolly,
anti-scientific metaphysics. Instead it turns out to be part of biology, suitably extended
to deal with recent products of evolution.
The case of internal classifications of intermediate states of perceptual virtual
machinery developed by some kind of self-organising classifier mechanism is different from
another case, namely where the internal perceptual states have a very complex structure
that varies systematically with changing relationships between the perceiver and a complex
environment. For example if you walk through a full car park on a sunny day, whether you
notice or not, your visual experiences will change in very complex ways that are
systematically related to your changing relationships to a multitude of surfaces of parts
of cars, parts of the ground on which they are parked, the direction of ambient sunlight,
the shadows cast by static and moving occupants of the car park, etc. In particular, there
will
be surfaces that become more or less occluded as you walk, there will be multiple
reflections and highlights visible in curved surfaces on car-bodies, all changing in
systematic but complex ways.
The details of the changing optic array are much more complex than I have described, and
most of the details are not noticed by most people. But it seems that human visual
mechanisms make good use of them to acquire rich and fairly precise information about the
occupants of the car park (i.e. the many parts and visible surface fragments of cars, lamp
posts, fences, side-walks, etc.) Some of the information will be used consciously, e.g. in
taking care that your trolley does not scratch a car, while much of it will be used
unconsciously to compute various properties, including location, orientation, and
curvature of surfaces, the materials of which they are made, and your relationships to
them. Some of the patterns of optic flow may be used unconsciously in posture control, as
shown by David Lee's moving wall experiment several decades ago. [REF] (It may be possible
for special training (e.g. for athletes) to make the information consciously accessible. I
am not sure about this.) In cases like this the supervenience of perceptual states on the
mixture of physical relationships between things in the environment and parts of your
body, including your optical mechanisms, is a very finely honed product of evolution
followed by development and learning.
More about qualia
The above discussion does not bring out clearly enough the fact that there
really are intermediate (virtual, non-physical) information structures created
by the visual information processing architecture and some of those structures
don't merely contribute causally to the construction of more 'high level'
information structures referring to cars, their wheels, their curved bonnets,
etc., but can, under certain circumstances be interrogated by more central
processes. The result, in many cases, is experience of qualia that are closely
related to, but distinct from, the objects seen in the environment. In the case
of dreams, hallucinations or visual illusions similar intermediate level
information structures may exist and may be attended to, without corresponding
to objects in the environment.An artist trying to depict a scene faithfully, by replicating much of its
appearance (not its actual structure, which cannot be replicated on 2-D paper),
has to learn to attend to features of his/her qualia that most people ignore,
which correspond to intermediate information structures organised in a 2-D
fashion in registration with the retinal information, and whose retinal location
can change as the direction of gaze changes. Those internal information
structures in the visual virtual machinery really exist and have different kinds
of causal roles depending on what else is going on. Nadia, the autistic child
described by Lorna Selfe in her 1977 book, was exceptionally good at attending to
and depicting on paper some of the contents of her visual experience.
(Discussed further here.)Not all visual qualia need correspond to external reflective surfaces. If you stare
at a coloured patch for a while and the patch is removed you may experience an
after-image of a different colour. Does the after-image exist? Yes: intermediate
visual information structures can be produced in many different ways, and the
normal production by an external stimulus is not the only one (as is shown
clearly by dreams, hallucinations and after-images). Moreover, you can make your
after-image move by tapping your eyeball. Is there something that moves? Yes.
Does it move in physical space? No: it moves in a 2-D information space in which it
changes its relationships. Some of the contents of that space are produced by
external surfaces, e.g. qualia corresponding to a perceived car bonnet. Those
qualia will also move when you tap your eyeball, even though the actual car bonnet
does not. In some cases left-eye qualia and right-eye qualia that are normally
stereoscopically fused come apart during such tapping and there is only motion of the
qualia derived from the tapped eye.NOTE ON Dennett: (Added 23 Apr 2013)
These remarks seem to flatly contradict, with supporting examples, the claims
made by Dan Dennett in this presentation:http://www.youtube.com/watch?v=AaCedh4Dfs4though I agree with him on many other topics.
"A Phenomenal Confusion About Access and Consciousness",Note on Block (revised 29 Apr 2013)
I don't yet know how closely what I've written corresponds to the views of Ned
Block on "Phenomenal consciousness", which Dennett finds so puzzling. What I also
find puzzling is Block's claim (e.g. in his 1995 paper) that contents of phenomenal
consciousness (as opposed to the contents of access consciousness) are allegedly
not suited to having cognitive functions. The only sense I can make of this is that for
many cognitive functions, other than merely being attended to, they normally need
further processing, e.g. to produce perceptual information about visible surfaces in the
environment. If that's what Block means, I have no quarrel with him, since it would
be an example of the general phenomenon in software engineering that information in
its original form cannot be processed in a certain way, though information derived by
it via intermediate mechanisms can be. For example, it is impossible to find English
syntactic structure in an auditory information stream, though after several layers of
processing to "extract" a sequence of words the parsing process can be applied.
Sometimes there has to be two-way cooperation between high level and low level
processing. (A similar point is made regarding multi-level processing of visual
information in the Popeye program mentioned above.) I used to think Block was
saying something much more obscure, and probably wrong.In his "Functional Reduction" paper Block discusses varieties of functionalism and
their compatibility with varieties of physicalism, in a very interesting way. However
I think a consideration of examples of virtual machine functionalism where the
virtual machinery includes several concurrently (but not necessarily synchronously)
active sub-systems with their own causal relationships, both within the VM and also
across its boundaries (e.g. to internal physical memory, to physical interfaces and
even to things referred to in the environment) would show that there is no version of
physicalism as normally defined that survives, even though all the virtual machinery
is ultimately implemented in physical processes. Part of the reason for this is that
patterns that can exist in physical structures and processes need not all be definable
in the language of physics. For example, what makes something an expression of the
English sentence "Today is Fred's birthday" depends not only on how current users of
English read text but also who Fred is, which calendar is in use and other complex
social facts. What makes the state of part of a virtual machine a case of someone
wondering what Fred thinks of him is even more remote from being translatable into a
physical description, but I shall not argue that here.Summary:
a well designed robot, with a visual system whose functionality is similar to
ours, in a cognitive architecture similar to ours, will, at any time, include
collections of intermediate structures that have many of the properties often
derided in the notion of a 'Cartesian Theatre'. Maybe it's time to stop deriding
it and finding out how to make one that works. (Murray Shanahan has made an
interesting attempt in his 2010 book 'Embodiment and the inner life', but I
don't think he has considered enough of the required functionality.) I have
tried to present some aspects of human mathematical consciousness that most
researchers ignore, here.)It should be clear that my position is very different from Dan Dennett's, even
though we share a great deal.
The experiment on unconscious seeing, available here, was designed in part to probe some
of those mechanisms.
The information processed by sophisticated machinery is not just patterns in physical or
virtual machines (like the bit-patterns used in computers), which may be conveyors or
encoders of information, but are not the the same things as the information they express
or encode (although in special cases one information conveyor can refer to another, or
even to itself).
I think there is a deep and complex notion of being able to acquire, manipulate, derive,
evaluate, analyse, communicate, combine, and use information, but much of its complexity
cannot simply be tamed by introspecting what we think we mean by "information" as many
people attempt to do.
The project of tracing the developments of information processing in biological evolution,
including identifying intermediate stages that nobody has thought of looking for, will
give us far deeper understanding of the problems in the long run.
It seems clear that the very earliest use of information was for control, e.g. a
microbe with mechanisms (using chemotaxis) for deciding things roughly like "should this
stuff be let in or not?", a decision that may initially be binary but could later be
refined e.g. "how much of this stuff should be let in" or "under what conditions is it
useful to let this stuff in?", or later "would this stuff be better to ingest than that
stuff in my current state?" etc. This doesn't imply that the machines construct English
sentences or sentences in any language that could be translated into English. Rather they
have mechanisms that perform functions that we can approximately describe in sentences,
though in some cases we may need new terminology -- e.g. the servo-control functions of
some brain mechanisms.
As evolution progressed and organisms, and their environments, became more and more
complex and varied, some of them acquired information processing capabilities that became
more complex and varied, including abilities to learn about, detect, and make use of what
J.J.Gibson called "affordances" (positive and negative, e.g. opportunities and obstacles)
in the environment. We are still in the process of developing good theories about what
organisms and machines can do with information, what forms the information can take, how
it can be represented or encoded, where and how it can be obtained, how it can be
manipulated, what other information can be derived from it, how it can be used, etc. etc.
Human intentionality sits near the peak of a mountain of biological resources for dealing
with a mountain of biological needs in many mountains of different situations.
We'll understand those resources best when we know how to replicate their functionality,
but there's still a very long way to go. However it's already clear that we know how to
make relatively simple machines that have types of intentionality that many non-human
animals do not (yet) have. For example, given the current state of software technology, we
need have no hesitation in describing a chess virtual machine running on my computer by
saying things like:
It has detected the new threat created by my last move and is now looking for aThis is not a way of speaking in metaphors. This sort of thing may be an accurate
suitable defence against the threat. In doing that it thinks about various possible
moves open to it and their consequences. It has noticed that there are three different
moves open to it that postpone the threat, insofar as I'll have to make one or more
moves to reinstate the threat. But it notices that one of its moves would produce a
situation in which it looks at first as if my threat has been completely blocked and
the only way to see that it isn't blocked is to notice the opportunity I have of
creating a diversion on the left flank that will require the computer to reallocate
resources that will allow me then to return to my original threat and force a mate.
So it chooses that last way of blocking the threat.
I am not an expert on computer chess, but I know enough about it to believe that there are
chess playing programs for which that kind of description could be true at a certain point
in a game, and in principle the sorts of software tools I've developed with colleagues and
students would make it possible to design machines with such capabilities, as would many
other tools.
However, that would not justify me in replacing the last sentence with:
So it chooses that last way of blocking the threat in the hope that its opponentThat change would require the intentional competences of the chess virtual machine to be
will not notice the possibility of reinstating the threat.
Giving the machine that meta-intentional kind of intentionality will require its
information processing architecture to be extended to include something like a model or
representation of the opponent as a thinking machine.
There are simple versions of such things in existing AI systems, and no doubt there will
be more sophisticated versions later on.
One of the assumptions of the Meta-Morphogenesis project is that long before
philosophers got interested in these matters, and long before AI researchers started
trying to give machines these capabilities, biological evolution 'discovered' the
advantages of organisms being able not only to perceive and think about certain subsets of
what actually exists, but also to perceive and think about some of the possible ways in
which things (e.g. spatial configurations) can change -- and also acquired the ability to
evaluate the possibilities in order to select a subset as worth attempting to realise.
(This is a long-winded summary of Gibson's claim that animals can perceive and make use of
affordances.)
There are lots of examples of animal behaviour that are impossible to make sense of
without assuming that they have such information-processing capabilities, even if they lack
the meta-competences required to detect that they have the capabilities and are using
them. Something similar can be said about very young children, before they are capable of
going through anything remotely like philosophical reflection on their thought processes,
but can think about which box to open to retrieve a toy that is out of sight.
I don't expect many philosophers to agree with me. But if and when we have made more
progress in designing such machines most of the philosophers who interact with them will
feel compelled to think of them as perceiving, thinking, intending, preferring, etc. I.e.
they will treat them as having intentionality. Moreover, they will be justified in doing
so because of how the machines work.
That will not be as weak as what Dennett suggests we do to other human beings, namely treat
them 'as if' they have intentionality -- as a useful way of making predictions (i.e.
adopting the intentional stance). It's no more a stance than believing unsupported apples
near the surface of the earth fall is adopting a gravitational stance.
The strategy I am proposing, namely doing philosophical analysis by designing, building
testing and explaining working systems, will show that intentionality is not just one
thing: there are different levels of complexity/sophistication that require different
sorts of underpinning information-processing machinery. This will be one of the important
results of the Meta-Morphogenesis project.
The same can be said about being conscious of something. Both consciousness and
intentionality, like efficiency, reliability, accuracy, ease of use, are polymorphic
concepts: they don't refer directly to properties of things, but to higher-order
relationships between properties and relations. The higher-order relationships may be the
same even if what they relate varies: For example, the efficiency of a lawnmower in
serving its purpose is very different from the efficiency of a proof in serving its
purpose. But the relation of economy of means is similar.
States and processes involving intentionality, like the states and processes in the chess
machine, often require the existence of virtual machinery, because physical machinery is
incapable of performing the right functions, even though the virtual machinery is
implemented in physical machinery. Typically the implementations require constantly
changing mappings between virtual and physical processes.
"Having tested all versions of functionalism, we recommend that you get yourselfTheir ideas were apparently developed quite independently of the work reported here. One
basic functionalism plus mechanisms plus neural representations and computations
plus naturalistic semantics based on information and control plus properties that
are powerful qualities. You'll have a complete account of the mind. Mental states
are representational functional states within the appropriate kind of computational
mechanistic system. Some mental states even have a qualitative feel."
(Quoted with permission.)
Other points stressed here, apparently not covered in that paper are
Maley and Piccinini appear to share the explanation of the existence of qualia presented
here, namely that some virtual machines have components that are capable of
inspecting/monitoring some of of the intermediate information structures in layered
perceptual systems. Since the existence of the intermediate structures does not depend on
their being monitored, the implication is that there can be unnoticed qualia, some
unnoticed because some subsystem did not switch attention, some unnoticed because there is
no subsystem capable of monitoring them, even if there could be. In humans, some of those
self-monitoring subsystems seem to develop a long time after birth. (Of course,
philosophers who define 'qualia' in terms of being objects of attention, or something like
that, will regard the notion of 'unnoticed qualia' as self-contradictory. I suggest that should
be compared with claiming that whales cannot be mammals because allowing mammals to
have fins is self-contradictory.)
I'll be interested to learn whether the idea of Virtual Machine Functionalism is included by
any other philosophers who discuss functionalism, or virtual machinery. They may use the
idea
but refer to it by some other label than "virtual machine functionalism".
John Pollock came close, in What Am I? Virtual machines and the mind/body problem,
Philosophy and Phenomenological Research., 76, 2, pp. 237--309, 2008,
But I am not sure he noticed the need for multi-process virtual machinery with multiple
concurrently active causal mechanisms as illustrated in the architecture diagrams and
the more complex supervenience diagram below.
I was introduced to the idea by R.M. Hare when I was one of his 'moral tutees' in Balliol
College Oxford, around 1959. By then he had started talking about whether ethical truths
supervened on non-ethical truths -- though he was a prescriptivist, not an objectivist
about ethics.
Later, the idea of supervenience was extended (by Donald Davidson) to the relationship
between mental and physical phenomena, and much discussion since then has been
concerned with whether such supervenience exists and if so what sort of supervenience.
See
http://plato.stanford.edu/entries/supervenience/
Questions discussed are whether properties of one kind (e.g. mental properties, such as
being hungry) can supervene on properties of another kind (e.g. being in some complex
physiological state), or whether one kind of state or process can supervene on another
kind of state or process (though processes are rarely discussed in detail). Similar
questions had arisen, using different terminology, in the Social Sciences, e.g. questions
about whether social or economic facts were reducible to or in some non-reducible way
supervened on non-social facts, including psychological facts and individual human
behaviours. "Wholism" and "Individualism" were among labels used for alternative answers
to such questions.
Our discussion of Virtual Machine Functionalism makes the claim that not just states,
properties, and processes, but complex working virtual machines, with their internal
causal interactions, can supervene on physical machinery -- e.g. operations in the working
chess machine supervene on aspects of the physical machinery in the computer on which it
runs. A key feature of the claim being made here is that the running virtual machinery,
and some of its parts, can have causal powers that affect not only parts of the virtual
machine, but also parts of the underlying machine and sometimes also its physical
environment, e.g. when the virtual machine is a chemical plant control system.
This kind of "downward" causation (causation of physical processes by VM processes)
may seem to be very mysterious: but learning how it is done in computers can help to
remove the mystery, and help us consider what to look for in how minds work. One of
the implications is that the same event can be caused in different ways. (This is
closely related to Elizabeth Anscombe's example of the same process being
describable in multiple ways in her 1957 book Intention.)
Here's a better depiction of the supervenience of a working virtual machine on a physical
system, showing more of the required structure in the supervenience relation, with lots of
bi-directional causation:
Figure Complex:
(Compare the complexity of Figure MultiVM above.)
It's not only causation that reaches out from the contents of a running virtual machine to
external entities: semantics can also. For example, software on my computer can use an
email address referring to someone in another country. The remote provision of many
products and services on the internet would be impossible but for the use of complex forms
of reference to things outside the referring machine, including clients, things bought and
sold, obligations, legal constraints, and many more. Work on the so-called "Semantic Web"
attempts to formalise such capabilities, but it seems to me so far to be trapped in
syntactic manipulations because the practitioners have not yet understood the requirements
for semantic competences. (A task in which good philosophers could help. But I don't work
on the semantic web, and my information, based on attending workshops and informal
discussions as
an 'outsider', may be out of date, if there has been progress recently.)
Notes:
It is possible to distinguish more kinds of supervenience than are normally considered in
philosophical discussions, including the following (with somewhat arbitrary labels):
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
Various horizontal, vertical and diagonal lines, and various rectangular, triangular and
There are very many kinds of mechanism supervenience studied in mechanics, including cases
where forces are amplified, speeds or directions of motion are controlled, energy is
transferred, etc. These are normally all thought of as cases of physical causation, but
the important point is that different sorts of physical causation occur on different
scales. The situation is even more complex in chemical causation.
Much of the power and beauty in Newtonian mechanics derived from the use of mathematics
(integral and differential calculus, for example) to demonstrate that various global
physical processes necessarily supervene on micro-processes with certain structures.
We don't yet have that for biology.
NOTE: Interacting/conflicting/causes in VMs:
There are problems about how some of the kinds of causation in virtual machines work that
still require investigation. I have begun to discuss some of them in this online slide
presentation:
http://www.cs.bham.ac.uk/research/projects/cogaff/talks/#talk86
Talk 86: Supervenience and Causation in Virtual Machinery
The Abstract includes:.... Some of the problems are concerned with concurrent interacting subsystems within a virtual machine, including co-operation, conflict, self-monitoring, and self-modulation. The patterns of causation involving interacting information are not well understood. Existing computer models seem to be far too simple to model things like conflicting tastes, principles, hopes, fears, ... In particular physical opposing forces and other well understood interacting physical mechanisms are very different from these interactions in mental machinery, even though they are fully implemented in physical machinery. This is likely to be "work in progress for some time to come." .... Two notions of real existence are proposed (a) being able to cause or be caused by other things (existence in our world) and (b) being an abstraction that is part of a system of constraints and implications (mathematical existence). Some truths about causal connections between things with the first kind of existence can be closely related to mathematical connections between things of the second kind. (I think that's roughly Immanuel Kant's view of causation, in opposition to Hume.)Thanks to Hugh Noble for reminding me of the importance of this.
The Birmingham CogAff (cognition and affect) project and its successors (now the
Meta-Morphogenesis project)
developed some theories about the layering of functionality
that overlap partly, but not completely, with architectural theories developed by others,
e.g. Minsky, Newell and his students, Langley, and many others. (An attempt by the BICA
society to collect and collate information on theories about architectures may still
be on-going: http://bicasociety.org/cogarch/
)
There's an introduction to the CogAff ideas here:
http://www.cs.bham.ac.uk/research/projects/cogaff/#overview
The layers of virtual machinery described there are sub-divided both in terms of
differences of function (including differences in semantic content) and differences of
evolutionary age. But it is very likely that the Meta-Morphogenesis project will show that
the three layers proposed ('Reactive' (oldest, most common), 'Deliberative' (newer and
rarer) and 'Meta-Management' (newest and rarest) perhaps better called Meta-semantic?)
layers all have internal subdivisions and there may be more more intermediate layers to be
added. (Compare Minsky's proposed six layers in The Emotion Machine). The differences
between perception, central processing and action also need to be more blurred than some
of our diagrams suggest. Here's a recent attempt to combine the horizontal layers and the
vertical divisions of function in a diagram Figure CogArch, acknowledging that perception
and
action sometimes overlap (as pointed out by Gibson and others).
(With thanks to Dean Petters.)
Arrows representing perceptual inputs at different levels reflect the fact that there is
information of different sorts in the environment requiring different kinds of perceptual
processing, illustrated by the different information levels involved in seeing marks on
paper, seeing familiar letters, seeing familiar words, reading phrases, reading and
understanding sentences, understanding stories or arguments, etc. New perceptual
mechanisms are required, building on old ones, in order to access the more abstract kinds
of information. Similar remarks can be made about actions with different levels of
sophistication e.g. twitching, chasing an insect away, signalling irritation to another
person, etc. (Compare Elizabeth Anscombe's 1957 book Intention.)
The examples of perceptual contents that have provoked most discussion of qualia,
sense-data, or phenomenal consciousness are probably elaborated versions of intermediate
perceptual information buffers that originally began to evolve in organisms that had only
reactive architectures. As more sophisticated forms or processing evolved the modes of
access to the contents of those buffers diversified and became more complex, both as a
result of evolutionary changes and also because of the ways in which various kinds of
learning can modify or extend the mechanisms provided by evolution. (In humans the
development of some of the mechanisms is 'deliberately' delayed to allow information
acquired in some parts of the system to be used to influence the growth of other parts, in
ways that lead to some of the phenomena Annette Karmiloff-Smith refers to as
"Representational Redescription" in her book "Beyond Modularity" discussed in more detail
here.)
Good analytic philosophers with a deep scientific background and some experience of
non-trivial software design and testing could make a major contribution.
In other species, sometimes called "altricial", the opposite is the case: they hatch or
are born helpless and lacking most of the competences they will later need (though the
sucking competence in new-born mammals is more sophisticated than it may seem). Jackie
Chappell and I have a paper presenting some sketchy ideas about a spectrum of patterns of
development including patterns where some genetically influenced competences are delayed
until other competences have been acquired so that they can provide some of the material
needed for the later competences. Likewise genetically partly pre-specified meta-competences,
and meta-meta-competences may begin to develop after the individual's architecture has
developed the necessary supporting mechanisms and acquired some of the information
required to match the new high level competence to features of the environment that can
vary. Human language development is the best known example, but there are others presented
in the work of Karmiloff-Smith mentioned above.
We argued that most species are a mixture of precocial and altricial features, and the
labels "altricial" and "precocial" should be applied to competences or features, not to
species or whole individuals. To avoid confusion we use different words for this, and
refer to competences as more or less pre-configured or meta-configured.
Our sketchy theory is (crudely) summarised in the diagram below showing alternative routes
from genome to behaviours, with increasing amounts, from left to right, of involvement of
learning and development based on results of previous interaction with the environment
using previously developed competences. The pre-configured competences are on the left,
increasingly meta-configured competences to the right.
Figure EvoDevo
[Chris Miall helped with the diagram.]
One implication of all this is first that in such species the relationship between the
genome and the virtual machines that develop is very complex, very indirect and liable to
influence by the environment, and consequently also very varied.
Another implication is that the later meta-meta...configured virtual machinery has functions
that can depend in complex ways on the functions of virtual machinery developed earlier in
that individual, which in turn can depend ultimately not only on the functions of the most
directly specified virtual machinery but also on various increasingly abstract features
of the environment -- for instance that some kinds of material can be used for building
shelters, and that some of those materials are more durable than others.
And finally, in the case of humans and possibly other social animals the functions of
virtual machines developed in the young can depend not only on the functions of virtual
machinery in other individuals in the community, but also on social and cultural products
of developments in many earlier generations. Any hope that there is a way of expressing
these functions in terms of the language of physics is a pipe dream.
In part, this is an indication that what we mean by "cause" has not been properly analysed.
Jackie Chappell and I have argued that animals and intelligent machines need two concepts
of causation, one Humean, based on evidence of correlations and one Kantian based on
understanding of structural relations. An example of the former (Humean causation) might
be a child's notion that pushing a light switch up or down makes a light go on or off. An
example of the latter (Kantian causation) is the fact that moving a vertex of a planar
triangle further from the opposite side causes the area to increase. Causation in and
between physical and virtual machinery can be of both kinds. However, that is a topic for
discussion elsewhere. This does not affect the notions of freedom/free-will presented in 1992.
The ideas presented here imply that the same event can have different causal explanations
in different explanatory contexts. In the discussion of how a chess VM is implemented, it
may be appropriate to answer the question "What caused the machine to detect the threat?"
by referring to some changes at an implementation level, e.g. changes in bit patterns in
the computer memory. At a higher level of description the answer to the question might
refer to a strategy deployed by the chess virtual machine for thinking about unexpected
moves made by its opponent.
The possibility of different correct answers to "What caused that?" is closely related to
the possibility of different correct descriptions answering "What's happening there?", as
discussed by Anscombe.
Although the statement that X caused Y may not be definitionally equivalent to any
statement about what would have happened if X had not happened, or if X had happened but
in different circumstances, there are connections between causation and counterfactual
truths. This is ignored by attempts to argue that any physical object can be interpreted as
performing any computation, by arbitrarily mapping portions of the object to portions of a
VM performing the computation. Such mappings prove nothing if they don't correspond to
causal relationships, with implications about what would have happened if.... This needs
to be taken into account in any arguments for or against functionalism or computational
theories of mind. (There is a lot more to be said about mental causation.)
Other gaps include requirements for replicating aesthetic enjoyment, including both
experiencing performances by other musicians, painters, dancers, etc., and enjoying being
a producer (as opposed to labelling things "good" or "bad" as a result of some training
process). What it is to find something funny is also a challenge (though I have not yet
looked closely enough at the Hurley, Dennett and Adams theory).
Why did the robot cross the road?There's lots more work to be done, and it requires close collaboration between high calibre
It was trying to understand the joke.
(To be continued....)
Mind in a Physical World,
MIT Press, 1998,
Note added 30 Apr 2013: He has just published a very relevant book which I have
not yet read, but intend to:
http://mitpress.mit.edu/books/explaining-computational-mind
Explaining The Computational Mind
By Marcin Milkowski,MIT Press, 2013.
Creative Commons License
This work, and everything else on my website, is licensed under a
Creative Commons Attribution 3.0 License.
If you use or comment on my ideas please include a URL if possible,
so that readers can see the original (or the latest version thereof).
Maintained by
Aaron Sloman
School of Computer Science
The University of Birmingham
.
