Previous: An Application and Some Empirical Results
Up: A Computational Model of Color Perception and Color Naming Next: An Architecture for Autonomous Agents
The goal for the research presented in this dissertation, as set forth in
Chapter , was twofold: firstly to contribute to theories of
autonomous agency, in particular to the study of symbol grounding or
embodiment, and secondly to do this by modeling a particular aspect
of perception and natural language semantics, viz. the domain of color
perception and color naming. From a methodological point of view, the
approach chosen was to study these phenomena from a ``vertically
integrated'' perspective, resulting in an experimental implementation of a
complete (albeit narrow-minded) color-naming and color-pointing agent,
ranging all the way from real visual stimuli captured by a camera to
symbolic descriptions using natural language terms, and back.
With respect to color perception, the goal was to model (an aspect of) human color perception by modeling known neurophysiological data on the responses of a certain class of color-sensitive neurons found in the human (and primate) visual system, relating these responses to existing color perception models of a psychological or psychophysical nature. The attempted explanation is of course only partial, as it does not deal with phenomena such as color constancy and the influence of context on color perception. In this respect the research presented was partly successful, in that it showed how one can derive a psychophysical color space (the NPP space) from neurophysiological data, and the resulting color space shows some remarkable similarities to existing psychological and psychophysical color spaces that have been derived independently, and using entirely different means. Various features of the derived color space, and the method used to derive it, are worthy of further investigation in and of themselves. The success in this area was only partial because the derived color space does not perform better for the purpose of color naming and color pointing than other psychophysical color spaces, in particular the CIE L*a*b* space. I believe this is mainly due to the fact that the NPP space is not fully perceptually equidistant, something which the L*a*b* space is explicitly constructed to be. In addition, the NPP space is based on measured responses from a single Macaque monkey only, and the CIE spaces are based on average experimental data from large populations of human subjects.
The category model that was developed, representing a perceptual color category as a normalized Gaussian function in three-dimensional color space (described by a focus location and a parameter determining the ``width'' of the function) performs well for our purpose. The model itself is not an arbitrary construct, but is based on independent psychological research in categorization, and presumably has a wider applicability to categorization in general. Since it assumes an underlying perceptually equidistant (Euclidean) space, it is not surprising that it works best in conjunction with the CIE L*a*b* space. The category model may also offer a tool to study bias in learning or development of visual (and perceptual) categories, but this connection is very tentative. The learning dimension of the color naming problem has only been touched on very briefly, and there certainly remains more work to be done in this area.
The color perception and naming model as implemented is of course limited. For instance, it does not deal with dynamically changing visual input, and for the most part assumes a constant adaptation state of the visual system. The computational approach adopted to deal to some extent with color constancy seems adequate for our purpose, but it is not clear whether it has any wider applicability. The more general question of how a color perception system comes about, with precisely this set of basic color categories, is not really addressed by the work presented. There are probably evolutionary reasons for why our color perception system works the way it does, but that is outside of the scope of the current research.
With respect to the larger goal of making a contribution to theories of cognitive agent architecture, I believe the research is successful. It shows how a particular set of terms (either terms from a natural language or terms from a knowledge representation and reasoning system) can be grounded or embodied in a perceptual categorization model, and presents a working implementation of such a model. Of course the scope of the model is limited, dealing only with Basic Color Terms, but it nevertheless represents a step in the direction of a general theory of (artificial) agency. The autonomous agent architecture my colleagues and I have developed as a tangent to the work presented in this dissertation hopefully contributes a little to this larger goal. One of the most important features of the color model in this respect is that it has a large bottom-up component (everything leading up to the color space representation of visual stimuli), causally connected to the outside world (the agent's environment), in addition to a top-down component (the categorization and naming model). These two components interact in a well-defined way, and while we can describe and study them separately to some extent, neither is eventually of any use without the other. I believe that the knowledge representation and reasoning community would do better to work on the problem of grounding some fundamental set of terms in this way than to worry about unicorns, round squares, possible worlds, and modeling logical inference (at least for the foreseeable future). The latter subjects may be interesting as an aid in modeling and understanding some of our own thought processes, but it is of no use to an artificial cognitive agent as long as it cannot relate to its environment in even the most basic of ways. In this respect I subscribe to the methodology of what has come to be known as ``nouvelle AI'' in some circles, emphasizing complete (albeit simple) artificial organisms that can function in a real world environment, before moving on to more ``high level'' problems. I believe AI can no longer afford to live in an ivory symbolic tower, and needs to deal with the ``low level'' grunge of the real world, and how an agent can relate to and interact with it.