Vision

8. Computational models of vision: problems and prospects

The most common criticism of computational models of human cognitive capacities, including accounts of our perceptual abilities, is that they are unable to approximate actual human performance. It is true that many impressive computer models fail miserably in the real world. Sometimes they fail because the information required is not available to the mechanism. As Marr emphasized, the computational theorist can try to avoid this problem by first attempting to characterize the computational problems that perceptual mechanisms, in their natural context, are required to solve, a process that involves discovering general environmental constraints that perceptual mechanisms of adapted organisms can be expected to exploit.

But the study of biological visual systems faces additional hurdles. Even if the information on which a posited process runs is in some abstract sense ‘in the data’, the input may be too ‘noisy’ for the mechanism to make use of it. Computational theorists are of course aware of this problem. Some of the processing posited by computational accounts, especially in early vision, involves the elimination of extraneous or irrelevant information in the image. (For example, the primal sketch in Marr’s account, which represents intensity changes in the image, does not preserve the absolute values of intensity gradients at every point in the grey-level array.) Bayesian models, in particular, attempt to isolate and discount confounding variables. Additionally, the theorist must eventually find neural hardware capable of doing the computationally characterized job, before being confident that the model is biologically feasible. Given the difficulty of the task it is unlikely that a complete computational account of vision is just around the corner. Nonetheless, computational theorists make an important contribution to our understanding of vision by their careful study of the nature of the problems to be solved by visual mechanisms, although the solutions they offer are properly evaluated by their performance in the real world.

An alternative style of computational model may ultimately prove better suited to explicating human vision than models, such as Marr’s, that treat perceptual processing as rule-governed operations defined over representations. In ‘connectionist’ computational architectures information is typically represented by patterns of activation over a connected network of units or nodes. Connectionist processes are explicated at a level distinct from the neurological or implementational. Connectionist cognitive models typically appeal to representations, memory and learning, hence they qualify as indirect; although connectionist accounts of representation, memory and learning differ in significant respects from more traditional computational accounts (see Connectionism). Connectionist theorists have claimed that their models are better able to handle noisy input and ‘multiple simultaneous constraints’ characteristic of real-world processing situations, though traditional computationalists have disputed this claim. Many Bayesian models lend themselves to implementation in parallel networks. Indeed, despite the significant idealization imposed by the Bayesian framework itself, Bayesian models may prove more amenable to integration with neurological accounts than traditional ‘representationalist’ models such as Marr’s. Some Bayesian models are designed specifically to be consistent with known neural mechanisms, with the prior and likelihood functions implemented in the model by synaptic weights. Whether the ‘transparency’ of these models from the neurological perspective proves ultimately to be a virtue will depend on whether the empirical predictions the models make possible are borne out.