Inference-driven attention in symbolic and perceptual tasks: Biases toward expected and unexpected inputs

Abstract

The aims of this paper are (a) to gather support for the hypothesis that some basic mechanisms of attentional deployment (i.e., its high efficiency in dealing with expected and unexpected inputs) meet the requirements of the inferential system and have possibly evolved to support its functioning, and (b) to show that these orienting mechanisms function in very similar ways in two perceptual tasks and in a symbolic task. The general hypothesis and its predictions are sketched in the Introduction, after a discussion of current findings concerning visual attention and the generalities of the inferential system. In the empirical section, three experiments are presented where participants tracked visual trajectories (Experiments 1 and 3) or arithmetic series (Experiments 2 and 3), responding to the onset of a target event (e.g., to a specific number) and to the repetition of an event (e.g., to a number appearing twice consecutively). Target events could be anticipated when they were embedded in regular series/trajectories; they could be anticipated, with the anticipation later disconfirmed, when a regular series/trajectory was abruptly interrupted before the target event occurred; and they could not be anticipated when the series/trajectory was random. Repeated events could not be anticipated. Results show a very similar pattern of allocation in tracking visual trajectories and arithmetic series: Attention is focused on anticipated events; it is defocused and redistributed when an anticipation is not confirmed by ensuing events; however, performance decreases when dealing with random series/trajectory—that is, in the absence of anticipations. In our view, this is due to the fact that confirmed and disconfirmed anticipations are crucial events for “knowledge revision”—that is, the fine tuning of the inferential system to the environment; attentional mechanisms have developed so as to enhance detection of these events, possibly at all levels of inferential processing.

Acknowledgments

We thank Carlo Umiltà, Kenny Coventry, Timothy Hubbard, Brian Scholl, David Fencsik, Jim McAuliffe, and an unknown reviewer for their helpful comments on previous drafts of this paper, and Sally Couchman for language revision.

Notes

A feature singleton is an element that has a feature value that is locally unique within a dimension (e.g., a red element in a background of blue elements; Pashler, 1988).

We do not commit ourselves to any specific algorithmic theory of reasoning or knowledge representation here: By using the term “rules”, we mean any possible way by which predictive environmental regularities are detected and stored, with no assumptions regarding their format, be it propositional, or based on models, or embedded in clusters of weighted associations within a network, or any other type of format.

Some scholars will not approve of considering “deductive” the generation of probabilistic anticipations; however, an inferential process is deductive as long as the falsity of its conclusions implies the falsity of at least one of the premises (i.e., pieces of previous knowledge) on which the conclusions were based. In this perspective, most knowledge-based anticipations are probabilistic only because they are grounded on contingent, uncertain knowledge; but, they are nevertheless deductive because discovery of their falsity implies the revision of the knowledge on which they were grounded.

Since a participant could not know whether a t-1 dot was part of a regular or a random series, the actual association of t-1 dots with imperative dots in the experiment was 33%.

The experiment used a central rectangular portion of the screen of size 14.36° × 10.52°. Therefore, asking the participants to maintain fixation on the centre would have caused many errors when the dot appeared twice consecutively in a location far from fixation. Experiment 3 uses a comparable task, modified so as to allow fixed gaze.

The effect has been observed for continuous movement and for inferred movement; in our experiment, the movement was inferred.

Yantis and Hillstrom (1994) suggested that it is not merely an abrupt onset that captures attention, but it is the presence of new objects in the visual fields. If our current inference-based distinction between expected, unexpected, and random events is used, and new objects are equated to unexpected objects, then the predictions by the abrupt-onsets account are similar to those by our current account, except for an important detail: Since in-target dots in regular trajectories are expected objects, they should not capture attention; therefore, responses to them should not be advantaged.

To minimize memory problems, at the beginning of each block participants were asked to read aloud and repeat to the experimenter the target number for the ensuing block.