=Paper=
{{Paper
|id=Vol-130/paper-13
|storemode=property
|title=The Philosophy of Information: A Methodological Point of View
|pdfUrl=https://ceur-ws.org/Vol-130/Greco.pdf
|volume=Vol-130
}}
==The Philosophy of Information: A Methodological Point of View==
https://ceur-ws.org/Vol-130/Greco.pdf
The Philosophy of Information
A Methodological Point of View
Gian Maria Greco, Gianluca Paronitti, Matteo Turilli, and Luciano Floridi
Information Ethics Group, Oxford University Computer Laboratory,
Oxford, United Kingdom
1 Introduction
The Philosophy of Information is a new area of research at the intersection of phi-
losophy and computer science [4]. It concerns (a) the critical investigation of the
conceptual nature and basic principles of information, including its dynamics (es-
pecially computation), utilization (especially computer ethics) and sciences; and
(b) the elaboration and application of computational and information-theoretic
methodologies to philosophical problems. Past work by members of our group
has concentrated on (a), and in this paper we explore (b). In a nutshell, we ask
what computer science can do for philosophy, rather than what the latter can
do for the former.
Applications of computational methods to philosophical issues may be ap-
proached in three main ways:
1. Conceptual experiments in silico, or the externalization of the mental the-
ater. As Patrick Grim has remarked “since the eighties, philosophers too have
begun to apply computational modeling to questions in logic, epistemology,
philosophy of science, philosophy of mind, philosophy of language, philoso-
phy of biology, ethics, and social and political philosophy. [. . . ] A number
of authors portray computer experimentation in general as a technological
extension of an ancient tradition of thought experiment” [10].
2. Pancomputationalism, or the fallacy of a powerful metaphor. According to
this view, computational and informational concepts are so powerful that,
given the right Level of Abstraction (see section 3), anything could be pre-
sented as a computational system, from a building to a volcano, from a forest
to a dinner, from a brain to a company, and any process could be simulated
computationally heating, flying and knitting. Even non-computable func-
tions would be representable, although by abstracting them to such a high
level that they would no longer count as a system (one would have to abstract
output and even termination and the existence of output, but a system has
to be allowed to terminate or not, even if one does not observe the output).
But then pancomputationalists (e.g. [2]) have the hard task of providing a
credible answers to the following two questions: how can one avoid blurring
all differences among systems, thus transforming pancomputationalism into
a night in which all cows are black, to paraphrase Hegel? And what would
it mean for the system under investigation not to be an informational sys-
tem (or a computational system, if computation = information processing)?
Pancomputationalism does not seem vulnerable to a refutation, in the form
of a possible counterexample in a world nomically identical to the one to
which pancomputationalism is applied.
3. Regulae ad directionem ingenii, or the Cartesian-Kantian approach. Are
there specific methods in computer science that can help us to approach
philosophical problems computationally?
In the following sections we answer this last question by introducing three
main methods: Minimalism, the Method of Abstraction and Constructionism.
Each one is discussed in a separate section.
2 Minimalism
Philosophical questions pose multi-faceted problems. According to Descartes,
a problem space can be decomposed by a divide-and-conquer approach. The
outcome is a set of more approachable sub-problems, interconnected in a sort
of Quinean web of dependencies. When dealing with a philosophical question,
the starting problem often presupposes other open problems and the strength
of the answer depends on the strength of the corresponding assumptions. A
minimalist starting problem relies as little as possible on other open problems,
thereby strengthening the final answer to the philosophical question.
Philosophers may improve the tractability of a problem space by choosing
discrete systems with which it can be studied. Minimalism outlines three criteria
to orientate this choice: controllability, implementability and predictability. A
system is controllable when its structure can be modified purposefully. Given this
flexibility, the system can be used as a case study to test different solutions for
the problem space. The second minimalist criterion recommends that systems be
implementable physically or by simulation. The system becomes a white box the
opposite of a black box (see section 4). Metaphorically, the maker of the system
is a Platonic “demiurge”, fully cognisant of the components of the system and its
state transition rules. The system can therefore be used as a laboratory to test
specific constraints on the problem space. The third criterion follows from the
previous two: the chosen system must be such that its behaviour is predictable.
The demiurge can predict the behaviour of the system in that she can infer the
correct consequences from her explanations of the system. The system outcomes
become then the benchmarks of the tested solutions.
The following three elements characterise Minimalism. First, Minimalism is
relational. Problems and systems are never absolutely minimalist, but always
connected with the problem space posed by the philosophical question. Second,
Minimalism provides a way to choose critically the starting problem for the anal-
ysis of a problem space, thus guaranteeing the strength of the next step in the
forward process of answering the philosophical question. According to a mini-
malist approach, the tractability of a philosophical problem is a function of the
three criteria outlined above. They allow the use of dynamic systems to test
possible solutions and to derive properties of the problem space. Finally, Mini-
malism is a matter of inferential relations between a problem and its space, but
it is not a way to privilege simple or elementary problems. Minimalist problems
may be difficult or complex.
Minimalism is an economic method that may be confused with Ockham’s
razor. The two methods are compatible, but while Ockham’s Razor avoids in-
consistencies and ambiguities by eliminating redundant explicative or ontological
elements in a theory, Minimalism provides a set of criteria for choosing problems
and systems relative to a given specific question. Moreover, Ockham’s principle
of parsimony is absolute and is applied to any theoretical element, while Mini-
malism’s main maxims of strength and tractability are always relative to a given
problem space.
A practical example of Minimalism applied to the philosophy of perception
may be helpful:
1. The identification of the question. We begin by asking e.g. “what is visual
perception?”. This question poses a wide problem space, hitherto approached
with different methods.
2. The Cartesian decomposition of the problem followed by a Quinean construc-
tion of the problem space. Some well-known sub-problems of this problem
space are the nature of internal representations, the role of mind in percep-
tion, vision as computation.
3. The identification of the starting problem. The standard representational
interpretation of perception is rich in assumptions about open problems.
Perception is based, for example, on the presumed existence of internal rep-
resentations. The sensorimotor approaches to visual perception are less de-
manding. Perception is chained to action while information is externalised.
James Gibson [9], one of the main advocates of the sensorimotor hypothe-
sis, cannot explain the nature of perceptual errors. This problem does not
rely on other open problems and therefore can be assumed as a minimalist
starting problem. We shall label it the “Gibson problem”.
4. The selection of the system to be used to study the starting problem. This
system has to be consistent with the requirements of Gibson’s sensorimotor
theory and with the criteria for Minimalism. The subsumption architecture,
proposed by Rodney Brooks [1], fits these requirements. The architecture of
Brooks’ robots is reactive, parallel and decentralised. Perception and action
are directly connected without any explicit internal representation or cen-
tralised inferential engines. Moreover, subsumption architectural behaviour
is fully specified by the topological structure of its layers composed of single
behavioural units. Its demiurge has full control and predictability power over
the system she has built. The Gibson problem can therefore be studied by
means of Brooks’ mobots.
5. The solution of the problem. In the sensorimotor approaches to vision, seeing
is something done by agents in their environments. The definition of percep-
tual errors must be shifted from a representational interpretation errors are
wrong computations made over internal representations to an action-based
interpretation errors are unsuccessful actions made by agents in their envi-
ronment. If the mobot’s sensorimotor features enable it to move randomly
in its environment then perception is successful, otherwise its perception is
erroneous. The mobot that bumps against a window lacks either the right
features or the sensorimotor capabilities relative to a given specific environ-
ment and its task of moving around randomly.
3 The Method of Abstraction
The process of making explicit the Level of Abstraction at which a system is
considered is called Method of Abstraction [8]. This epistemological method
applies both to conceptual and phenomenological systems and it is based on the
key concept of Level of Abstraction (LoA).
The metaphor of interface in a computer system is helpful to illustrate what
a LoA is. We all know that users seldom think about the fact that they use a
variety of interfaces between themselves and the real electro-Boolean processes
that carry out the required operations. An interface may be described as an
intra-system, which transforms the outputs of system A into the inputs of system
B and vice versa, producing a change in data types. LoAs are comparable to
interfaces because:
1. they are a network of observables;
2. the observables are related by behaviours that moderate the LoA and can
be expressed in terms of transition rules;
3. they are conceptually positioned between data and the agents’ information
spaces;
4. they are the place where (diverse) independent systems meet, act on or
communicate with each other.
LoAs can be connected to form broader structures of abstraction, from hi-
erarchy of abstractions to nets of abstraction. One of the possible relations be-
tween LoAs is simulation. A simulation relation [13] is the relation between the
observables of a simulator system and a simulated one. This relation must oc-
cur between pairs of observables in order to guarantee a satisfactory degree of
congruence not only for the current state of the two systems but also for their
evolution. In the simulation relation, the epistemic agent is coupling the state
evolution of two systems by observing these two systems at different LoAs. This
means that an epistemic agent tries to construct an equivalence relation be-
tween the two systems, seeking to understand at what LoA those systems could
be considered congruent.
As example, let us now apply the Method of Abstraction and the simulation
relation in order to re-define functionalism. Functionalism argues that a phys-
ical or abstract entity is identified by its causal or operational role. From this
viewpoint, a system is not evaluated by its structures and their interactions, but
rather by the functions it shows. If the “matter” constituting a system is irrele-
vant for its identification, then the same functional organization can be realized
by different systems and substrates, which are usually called realizations [12].
This is the multi-realizability thesis.
Some philosophers try to rule out multi-realizability from the functionalist
approach. They argue that multi-realizability could lead to a weakening of a neu-
roscientific approach in the explanation of human behaviour. Why bother with
actual neural structures if one can execute an algorithm to instantiate the same
behaviours shown by these neural structures? It is argued that a computational
approach is therefore more suitable for processing those algorithms.
Unfortunately, multi-realizability cannot be disconnected from functionalism
since, without it, functionalism becomes inexplicable. This is clear if we consider
the mathematical concept of function. A function is usually expressed by an
operation on one or more variables. The well-known scheme is f (x) = y, but
this simply means that the variables in the equation could be realized by an
infinite class of numbers or by points over the Cartesian plane or by means of a
Turing machine or by set theory. Without all these instantiations, it would be
impossible to explain the function f (x) = y. We shall therefore conclude that
functionalism entails multi-realizability.
Now, in the classic account of functionalism we deal with relata (the func-
tional organization and the realizations) and relations (the realization relation
between the functional organization and the realizations, and the simulation
relation between the various realizations).
Our goal is to show that realization and simulation are equivalent. One can
say that an epistemic agent can observe any functional organization at a specific
LoA and the realization of that functional organization at another LoA. Then the
realization relation between the two LoAs is characterized by: (a) the codification
of the inputs of the functional organization LoA into the inputs of the various
realizations LoAs, and (b) the de-codification of the outputs of the latter into
the outputs of the former. Basically, simulation relation and realization relation
are equivalent because they are relations which describe the same processes.
Given that multi-realizability and functionalism are coupled concepts and
that a simulation relation is equivalent to a realization relation, it follows that a
common functional organization does not exist at a higher LoA than that of its
realizations. The functional organization is the net of abstraction constructed
by the epistemic agents with the simulation relation between the various realiza-
tions conceived at different LoAs. This means that a functional organization is
the relational structure produced by various realizations and by the simulation
relation that connects them. For example, a carpenter who is making a piece
of furniture by following a blueprint is not handling a functional organization
(the blueprint) and a realization (the piece of furniture), but two realizations at
different LoAs, which are related in a simulation relation specified by his work.
This new interpretation of functionalism leads us to reconsider functional-
istic explanations within the philosophy of AI and the philosophy of mind by
introducing simulation relation as a new player. The functionalistic explanation
is configured as a specification of simulations between the LoAs at which the
realizations are disposed by the epistemic agent.
4 Constructionism
A black box is a system whose internal structure, rules and composition remain
undisclosed. A white box is a system about which one knows everything, be-
cause one has constructed it. This perspective lays in the wake of the so-called
maker’s knowledge tradition, according to which: (a) one can only know what
one makes and therefore (b) one cannot know the genuine nature of reality in
itself. Philosophers who stress (b) argue that, since any attempt to know the
intrinsic nature of the world will inevitably fail, it is better to concentrate on
those sciences whose subject is created by us, such as politics and social sciences.
Philosophers who stress (a) argue that it is possible to improve our knowledge
of reality through the improvement of our knowledge of the techniques by which
reality is investigated. This tradition finds its champion in Francis Bacon’s phi-
losophy of technology. Following Bacon, technology becomes the main subject
of philosophical enquiry, because it is both a human product and the means
through which the world is investigated. Constructionism explicitly refers to the
maker’s knowledge tradition. Its method consists of the following five principles:
1. The Principle of Knowledge: only what is constructible can be known. Any-
thing that can not be constructed could be subject, at most, to a working
hypothesis.
2. The Principle of Constructability: working hypotheses are investigated by
(theoretical or practical) simulations based on them.
3. The Principle of Controllability: simulations must be controllable.
4. The Principle of Confirmation: any confirmation or refutation of the hy-
pothesis concerns the simulation, not the simulated.
5. The Principle of Economy: the fewer conceptual resources used, the better.
In any case, the resources used must be fewer than the results accomplished.
Constructionism suggests that, given a theory, one implements and tests it
in a system. Because one constructs the system, one can also control it. As
Newell and Simon remarked “neither machines nor programs are black boxes;
they are artefacts that have been designed, both hardware and software, and
we can open them up and look inside” [11] (for constructionist approaches in
Cybernetics and proto-Cybernetics see [3]). Consider for example behaviour-
based robotics. One may observe an ant and offer a hypothesis about its internal
structures in order to explain its behaviours. Then one may build a system to
test that hypothesis. The resulting system is controllable in that it is modifiable,
compositional and predictable. This means that, as far as the constructed system
is concerned, one can change its internal structures and rules; the system can be
implemented by adding or removing new parts; and since one knows the rules
of the system, one can know its behaviour. Suppose that the mobot we have
constructed behaves like an ant. The Principle of Confirmation prevents us from
generalizing the working hypotheses, as if the simulation were the real cause (or
internal structure) of the simulated. From this, the sub-Principle of Context-
dependency is derived: isomorphism between the simulated and simulation is
only local, not global. The mobot accounts for the behaviour of the ant only
under the constraints specified by the simulation. If the constraints change, so
does the evaluation of the hypotheses.
Constructionism is in plain contrast to any mimetic approach in epistemol-
ogy. The latter assumes that reality is approached through some reproductive
or representational mechanism: ideas, mental images, corresponding pictures,
concepts and so forth are copies or portraits of some otherwise mysterious real-
ity in itself. Constructionism, on the contrary, considers knowledge a modelling
process, which shapes and edit reality to make it intelligible. It therefore re-
jects mimetic theories such as Plato’s, Descartes’ or Locke’s. The Principle of
Economy refers to the “careful management of resources”. On the one hand, in
defining knowledge processes, mimetic theories use a large amount of resources.
Assuming that there is a reality and that it works in some particular way means
making a heavy ontological commitment. On the other hand, Constructionism
does not state anything about reality in itself. A more modest commitment
makes errors less likely.
The Turing Test (TT) is an enlightening example of how the methodology
outlined in this paper and, more specifically, the constructionist method work,
for it respects the minimalist criterion, uses the LoAs and is constructionist.
Turing refuses even to try to provide an answer to the question “can a machine
think?”. He considers it a problem “too meaningless to deserve discussion” [14],
because it involves vague concepts such as machine’ and thinking’. Turing sug-
gests replacing it with the imitation game, which is exactly more manageable
and less demanding from the minimalist point of view. By so doing, he specifies
a LoA and asks a new question, which may be summed up thus: “can we con-
sider that a computer is thinking, at this Level of Abstraction?”. The rules of
the game define the conditions of observability [8]. If we observe the behaviour
under those conditions, we can accept an operational definition of a thinking ma-
chine for that LoA. By changing the rules of the game, one changes the LoA and
consequently the answer. Note how TT respects the constructionist principles:
1. By satisfying Minimalism, Turing also respects the Principle of Knowledge.
2. Turing makes a hypothesis based on the common assumption that conver-
sation skills require intelligence, and then he devises a system to evaluate
whether a machine is intelligent comparatively.
3. The system is controllable. We know how it works and how it can be modi-
fied.
4. Whether a machine passes the test implies only that the machine can, or
cannot, be considered intelligent at that LoA.
5. Finally, in tackling the problem of artificial intelligence, Turing refuses to
consider those ways requiring a large amount of conceptual resources. This
is, for instance, why he refuses to deal with any psychological assumption
about intelligence.
5 Conclusion
In this paper, we have introduced three methods and shown how they can be
imported from computer science into philosophy, in order to model and analyse
conceptual problems. We have outlined their main features and advantages. The
methods clarify implicit assumptions, facilitate comparisons, enhance rigour and
promote the resolution of possible conceptual confusions. Some applications of
the methods discussed in this paper have already been successfully provided in
computer ethics [7], in epistemology [5], and in the philosophy of information [6].
Of course, the adoption of the methods raises important further questions. We
mention only three of them that seem to us particularly pressing: (a) What is the
logic of problem spaces? (b) What are the logical relations between LoAs? (c)
How can Constructionism avoid solipsism? We have not attempted to answer
these questions, which we hope to address in future work? .
References
1. Brooks, R.: Intelligence without Representation. Artificial Intelligence 47 (1991)
139-159.
2. Chalmers, D. J.: The Conscious Mind: In Search of a Fundamental Theory. Oxford
University Press, Oxford (1996)
3. Cordeschi, R.: The Discovery of the Artificial. Behavior, Mind and Machines
Before and Beyond Cybernetics. Kluwer, Dordrecht (2002)
4. Floridi, L.: What Is the Philosophy of Information? Metaphilosophy 1-2 (2002)
123-45
5. Floridi, L.: On the Logical Unsolvability of the Gettier Problem. Synthese 142.1
(2004a) 61-80
6. Floridi, L.: Open Problems in the Philosophy of Information. Metaphilosophy 4
(2004b) 554-82
7. Floridi, L., Sanders, J. W.: On the Morality of Artificial Agents. Minds and
Machines 3 (2004a) 349-79
8. Floridi, L., Sanders, J. W.: The Method of Abstraction. In Lang, P. (ed.): Year-
book of the Artificial. Issue II: Models in contemporary sciences (2004b) 177-220
9. Gibson, J. J.: The ecological approach to visual perception. Houghton Mifflin,
Boston (1986).
10. Grim, P.: Computational Modeling as a Philosophical Methodology. In Floridi,
L. (ed.): The Blackwell Guide to the Philosophy of Computing and Information.
Blackwell, Oxford (2004)
11. Newell, A., Simon, H. A.: Computer Science as Empirical Enquiry: Symbols and
Search. Communications of the ACM 3 (1976) 113-126
12. Putnam, H.: Psychological Predicates. In Captain, W. H., Merrill, D. D.: Art,
Mind and Religion. Pittsburgh University Press, Pittsburgh (1967)
13. Roever, W.-P. de and Engelhardt, K.: Data Refinement: Model-Oriented Proof
Methods and their Comparison. Cambridge University Press, New York (1998)
14. Turing, A. M.: Computing Machinery and Intelligence. Mind 49 (1950) 433-460
?
For Italian legal requirements, Gianluca Paronitti must be considered the author
of section 3, Matteo Turilli of section 2, Luciano Floridi of sections 1 and 5, Gian
Maria Greco of section 4 and the first author of the whole paper. We wish to thank
Jeff Sanders for all his fundamental input when discussing previous versions of this
paper.