REPRESENTATION IN NATURAL AND ARTIFICIAL AGENTS - AN EMBODIED COGNITIVE SCIENCE PERSPECTIVE

The goal of the present paper is to provide an embodied cognitive scie nce view on representation. Using the fundamental task of category lea rning, we will demonstrate that this perspective enables us to shed ne w light on many pertinent issues and opens up new prospects for invest igation. The main focus of this paper is on the prerequisites to acqui re representations of objects in the real world. We suggest that the m ain prerequisite is embodiment which allows an agent - human, animal o r robot - to manipulate its sensory input such that invariances are ge nerated. These invariances, in turn, are the basis of representation f ormation. In other words, the paper does not focus on representations per se, but rather discusses the various processes involved in order t o make learning and representation acquisition possible. The argument structure is as follows. First we introduce two new perspectives on re presentation, namely frame-of-reference, and complete agent. Then we e laborate the complete agent perspective and focus in particular on emb odiment and situatedness. We argue that embodiment has two main aspect s, a dynamic and an information theoretic one. Focusing on the latter, there are a number of implications: Representation can only be unders tood if the embedding of the neural substrate in the physical agent is known, which includes morphology (shape), positioning and nature of s ensors. Because an autonomous mobile agent in the real world is expose d to a continuously changing high-dimensional stream of sensory stimul ation, if it is to learn category distinctions, it first needs a focus of attention mechanism, and then it must have a way to reduce dimensi onality of this high-dimensional sensory stream. Learning is very hard because the invariances are typically not found in the sensory data d irectly - the classical problem of object constancy: it is a so-called type 2 problem. Rather than trying to improve the learning algorithms - which is the standard approach - the embodied cognitive science vie w suggests a different approach which focuses on the nature of the dat a: the agent is not passively exposed to a given data distribution, bu t, by exploiting its body and through the interaction with the environ ment, it can actually generate the data. More specifically, it can gen erate correlated data that has the property that it can be easily lear ned. This learnability is due to redundancies resulting from the appro priate interactions with the environment. Through such interactions, t he former type 2 problem is transformed into a type 1 problem, thus re ducing the complexity of the learning task by orders of magnitude. By observing the frame-of-reference problem we will discuss to what exten t these invariances are reflected - represented - in the ''neural subs trate'', i.e. the internal mechanisms of the agent. It is concluded, t hat representation is not a concept that can be studied in the abstrac t, but should be elaborated in the context of concrete agent-environme nt interactions. These ideas are all illustrated with examples of natu ral agents and artificial agents. In particular, we will present a sui te of experiments on simulated and real-world artificial agents instan tiating the main arguments.