An optimal-fitness framework for modeling perceptual compression

Abstract

AbstractPerceptual systems are constrained by their information transmission capacity. Accordingly, organismal strategies for compressing environmental information have been the subject of considerable study. The efficient coding model posits maximized mutual information between stimuli and their neural representation. The reward maximization model posits minimized signal distortion, operationalized as reward foregone due to stimulus confusion. The matched filters model posits the preferential transmission of information that informs evolutionarily important decisions. Unfortunately, the efficient coding model is sometimes at odds with empirical findings, and all three models struggle with recapitulating each other’s predictions. Here I aim to reconcile the models by developing a framework for modeling compression in which: compression strategies dictate stimulus representations, compressed stimulus representations inform decisions, decisions deliver rewards, environments differ in decision-reward associations and fitness function, and therefore, different environments select for different compression strategies. Using this framework, I construct environments in which the fittest compression strategy: optimizes signal distortion, optimizes both signal distortion and mutual information, and optimizes neither but nevertheless is fit because it facilitates the avoidance of catastrophically risky decisions. Thus, by modeling compression as optimal with respect to fitness, I enable the matched filters model to recapitulate the predictions of the others. Moreover, these results clarify that mutual information maximization and signal distortion minimization are favored by selection only under certain conditions. Hence, the efficient coding model is reconciled with the findings that it fails to predict, because those findings can now be understood to derive from outside the model’s proper scope of application. Going forward, the optimal-fitness framework is poised to be a useful tool for further developing our understanding of nature’s perceptual compressions; a salient reason why is that it enables empirical findings to be bridged not only with concepts from information theory, but also economics.Author SummaryPerceptual systems are constrained by their information transmission capacity. Thus, stimuli are not transmitted in full detail, but are instead compressed. Presently, there are several extant models of compression that are supported by empirical results. However, they do not recapitulate each other’s predictions, and are not bound by any common conceptual framework. In the present study, I create a common conceptual framework: the optimal-fitness framework, which allows for the evaluation of the evolutionary fitness of a particular compression in a particular environmental context. This framework, in turn, allows me to define the features of the environments that favor the compressions predicted by the extant models. These findings serve to refine the extant models by defining their domain of applicability, and to unify the models by demonstrating the existence of environments in which their predictions overlap. Furthermore, the optimal fitness framework accommodates the expression of, and the demonstration of the evolutionary value of, various naturalistically plausible compressions that are not predicted by the existing models.