38 Informational environments
Visual processing is one contributing factor to an organism’s successful behaviour, but, taken by itself, it only allows for inferences as to how the world stands,
which may or may not be borne out by how the world actually stands. These
inferences are consolidated for the perceiving organism by the repeated success of
his behaviours towards what is perceived over the course of repeated perceptual
instances. They are consolidated for the external observer by matching different
perceptual instances against each other and against the physical knowledge at
hand. To some extent, this picture resembles W.V.O. Quine’s image of “how the
human subject of our study posits bodies and projects his physics from his data”
on the one hand (1969a, 83), and of “total science” as “a field of force whose
boundary conditions are experience” on the other (1961, 42).
In order to get this “cognitivist” view of visual perception as information processing to work as proposed by Marr, and in order to counter the underdetermination problem, information has to be reified, so as to keep a mark of identity for
what is being processed throughout all stages involved while, at the same time,
keeping it attached to the original stimulus and the outgoing response. Marr does
not present an explicit concept of information, but only on a reading of information as entities that can be subject to a well-defined sequence of formal operations,
traceable from input to output, the complex array of symbols and their transformations involved in perception will be properly grounded. And only to the extent that
there is a parallel between representational and computational processes, in terms
both of the nature of information and of the methods of processing involved, and
in the same sense as computation is cognition to Pylyshyn (1980), the perceptual
processing in Marr’s theory has a claim for an empirical grounding (see the schematic representation of the parallel in Marr 2010, 332).
Where information is reified in Marr’s account, its receivers, in certain
respects, are not. Information processing, Marr maintains, must be understood
on three clearly distinct levels: first, the goal of the computational process has
to be established; second, the algorithms and the representations for input and
output need to be determined; and, third, the physical realisation of representation and algorithms has to be identified. Although they will provide an understanding of some information-processing task only in conjunction, these levels
are relatively independent of each other (Marr 2010, 24–27). Accordingly, and
in alignment with the computational paradigm, the formal, algorithmic structure
of the subject matter can be analysed in abstraction from its concrete physical
realisation, and may be realised in a variety of physical arrangements. Hence,
one could not only provide machine models and computer simulations of perceptual processes but also construct a “general purpose vision machine” (Marr
2010, 331). The concrete physical realisation of a perceptual system, being part
of an organism interacting with his environment, is relevant to be sure, but
remains underdetermined by the computational theory proper, and is described
in abstract formal terms. Together with the presumed parallelism between perceptual and computational processes, it is this assumption that bears witness to
the alignment of Marr’s theory with AI and placed it at the origins of computational neuroscience.
Visual processing is one contributing factor to an organism’s successful behaviour, but, taken by itself, it only allows for inferences as to how the world stands,
which may or may not be borne out by how the world actually stands. These
inferences are consolidated for the perceiving organism by the repeated success of
his behaviours towards what is perceived over the course of repeated perceptual
instances. They are consolidated for the external observer by matching different
perceptual instances against each other and against the physical knowledge at
hand. To some extent, this picture resembles W.V.O. Quine’s image of “how the
human subject of our study posits bodies and projects his physics from his data”
on the one hand (1969a, 83), and of “total science” as “a field of force whose
boundary conditions are experience” on the other (1961, 42).
In order to get this “cognitivist” view of visual perception as information processing to work as proposed by Marr, and in order to counter the underdetermination problem, information has to be reified, so as to keep a mark of identity for
what is being processed throughout all stages involved while, at the same time,
keeping it attached to the original stimulus and the outgoing response. Marr does
not present an explicit concept of information, but only on a reading of information as entities that can be subject to a well-defined sequence of formal operations,
traceable from input to output, the complex array of symbols and their transformations involved in perception will be properly grounded. And only to the extent that
there is a parallel between representational and computational processes, in terms
both of the nature of information and of the methods of processing involved, and
in the same sense as computation is cognition to Pylyshyn (1980), the perceptual
processing in Marr’s theory has a claim for an empirical grounding (see the schematic representation of the parallel in Marr 2010, 332).
Where information is reified in Marr’s account, its receivers, in certain
respects, are not. Information processing, Marr maintains, must be understood
on three clearly distinct levels: first, the goal of the computational process has
to be established; second, the algorithms and the representations for input and
output need to be determined; and, third, the physical realisation of representation and algorithms has to be identified. Although they will provide an understanding of some information-processing task only in conjunction, these levels
are relatively independent of each other (Marr 2010, 24–27). Accordingly, and
in alignment with the computational paradigm, the formal, algorithmic structure
of the subject matter can be analysed in abstraction from its concrete physical
realisation, and may be realised in a variety of physical arrangements. Hence,
one could not only provide machine models and computer simulations of perceptual processes but also construct a “general purpose vision machine” (Marr
2010, 331). The concrete physical realisation of a perceptual system, being part
of an organism interacting with his environment, is relevant to be sure, but
remains underdetermined by the computational theory proper, and is described
in abstract formal terms. Together with the presumed parallelism between perceptual and computational processes, it is this assumption that bears witness to
the alignment of Marr’s theory with AI and placed it at the origins of computational neuroscience.