A neuronally encoded structured version of the world which could potentially guide behaviour

'Representation' and 'internal representation' are used in multiple senses in philosophy, linguistics, and the cognitive sciences. The meaning of'internal representation' as used here deserves, therefore, careful clarification, especially as it is ardently "reductionistic. Generally speaking, 'representation' is the expression of things in one language transformed into another. 'Language' is any set of symbols with rules for putting them together (Marr 1982). We are not engaged here, however, in the formal treatment of representations at large, but rather in the application of the concept to memory research. In the context of brain sciences, 'representation' means encoding of things in the world, such as objects, events, and processes, in neuronal language(s). This encoding is done in a way that enables the nervous system to manipulate the representations, modify and transform them, while maintaining: (a) parsing, which is the distinctiveness of things represented, and (b) structural relationships between the things represented. Both are needed for useful interaction with the world.

'Representations' have a long and rich history in the philosophy of mind, referring to some kind or another of 'mental images', or elements of an inner 'private language' or a 'language of the mind'. Some aspects of this usage can be traced back to 'phantasia', which meant 'appearance' or 'perception' in Greek philosophies (Irwin 1991; Long 1991; Annas 1992). A limited yet highly varied selection of notable examples includes treatments by British "associationism (Hobbes 1651; Warren 1921), Kant (Kant 1781; Caygill 1995), Bergson (1908), Wittgenstein (McGinn 1997), and more recent philosophers and cognitive theorists (Stich and Warfield 1994; Markman and Dietrich 2000). In modern discourse, it is useful to distinguish between two "levels of treatment of'representation'. One is cognitive, mental, 'semantic', or 'symbolic'. 'Mental representations' in this sense are theoretical postulates invoked to account for 'propositional attitudes'. The latter are regarded as 'mental sentences' characterized by a specific content, being about something in the world ('intentional'), and conditions that can satisfy the proposition. For example, believing that x, being angry at y, or desiring z are all 'propositional attitudes'. The other level of analysis in which 'representations' are currently used is the computational or implementational level, which some philosophers of the mind refer to as 'subsymbolic'; here representations are activated vectors in neuronal coding space (e.g. Cooper 1973; Churchland and Sejnowski 1992; definition 1). The relevance of 'symbolic' representations to 'subsymbolic' ones is a matter of heated debates (e.g. Fodor and Pylyshyn 1988). Phylogenetic considerations as well as "Ockham's razor lead to the assumption that 'subsym-bolic' and 'symbolic' representations are the micro- and the macrolevels of the same brain and mental faculty. This is clearly a case in which interlevel translation via 'correspondence rules' is badly needed ("reduction). For our purpose, we should also note that often, even computational neuroscientists who manipulate 'sub-symbolic' representations do hope to explain complex mental states at the end of the day.

The stand taken here is strictly reductive. The tenet is that nervous systems, even the most primitive of them all, had evolved to encode knowledge about the world. 'Knowledge' is used in the most elementary sense and is devoid of "anthropomorphic connotations of awareness and "consciousness. It refers to structured bodies of information possessed by the organism about the world, and capable of setting the organism's reaction to the world. 'World' means both the external milieu and the external states of the organism, and the organism means, specifically, the nervous system. These neu-ronally encoded structured versions of the world that could potentially guide behaviour are the 'internal representations' (definition 2). Representation is hence an inherent and fundamental function of all nervous systems. Therefore, internal representations are expected to vary tremendously in their complexity. Some are very simple, for example, a neuronal circuit subserving withdrawal in response to pain (e.g. "Aplysia), encodes a representation of'no pain' or various intensities of pain, and the appropriate motor response programme.1 Other internal representations are far more complex, and many, for example representations of propositional attitudes, are highly complex. However, regardless of their complexity, all internal representations as considered here: (a) are encoded in some way or another in neuronal systems; (b) determine the behavioural output to an input; and (c) when altered, may modify the potential to react rather than immediate action or reaction to an input (Dudai 1989, 1992; "learning, "memory).

In computational neurosciences, a further distinction is sometimes being made, between 'representations' and 'internal representations'. This stems from "models of associative networks. The simplest associative network is composed of two layers, input and output. A set of input patterns arriving at the input layer is mapped directly into a set of output patters at an output layer. Under these conditions, the "system is said to lack 'internal representations', because the coding provided by the external world suffices to generate the output. In contrast, when the number of layers increase and intermediate, 'hidden' layers are added, the representation is said to be 'internal' (Rumelhart et al. 1986c). However, this distinction does not hold water when real nervous systems are considered. In even the simplest nervous systems, sensory information is recoded and manipulated in neuronal language, be it at the level of cells or circuits ("percept). Therefore, any representation of information by the nervous system should be considered 'internal'.

'Internal representation' is thus a generic term, referring to the most fundamental functional property of nervous systems. Indeed, with evolution, it became manifested in many forms and realized in a variety of codes. But as an umbrella term, it is highly valuable. At the conceptual level, it focuses our attention on the essence of brain function. At the practical level, it guides us to search for representational codes and look for the most important changes, i.e. the representational ones, in systems that learn. Admittedly, we are still short of even a mini-dictionary that translates neuronal activity from representational code into behavioural change.2 But that day will soon come. And from that day on, the philosopher shall dwell with the neurobiologist, and the mutual interest in 'representations' will no doubt lead them once-in-a-while to dare and attend each other's seminars.

Selected associations: Attention, Binding, Learning, Map, Memory

1In the intact organism, even in this simple case, the representation encoded in the reflex circuit is in fact but one element in a structured body of information encoded by the nervous system in toto.

2For a limited selection of expeditions for the neuronal Rosetta stones that might yield this interlevel translation: see Lee et al. (1988), Shadlen et al. (1996), Dan et al. (1998), Kitazawa et al. (1998), Stanley et al. (1999), Miyashita and Hayashi (2000), and Zhang and Barash (2000); see also *Aplysia, *hippocampus, and *honeybee.

0 0

Post a comment