Trichotomy revisited: A monolithic theory of attentional control

Abstract

The control of attention was long held to reflect the influence of two competing mechanisms of assigning priority, one goal-directed and the other stimulus-driven. Learning-dependent influences on the control of attention that could not be attributed to either of those two established mechanisms of control gave rise to the concept of selection history and a corresponding third mechanism of attentional control. The trichotomy framework that ensued has come to dominate theories of attentional control over the past decade, replacing the historical dichotomy. In this theoretical review, I readily affirm that distinctions between the influence of goals, salience, and selection history are substantive and meaningful, and that abandoning the dichotomy between goal-directed and stimulus-driven mechanisms of control was appropriate. I do, however, question whether a theoretical trichotomy is the right answer to the problem posed by selection history. If we reframe the influence of goals and selection history as different flavors of memory-dependent modulations of attentional priority and if we characterize the influence of salience as a consequence of insufficient competition from such memory-dependent sources of priority, it is possible to account for a wide range of attention-related phenomena with only one mechanism of control. The monolithic framework for the control of attention that I propose offers several concrete advantages over a trichotomy framework, which I explore here.

Keywords: Attentional control; Learning; Memory; Selection history; Visual attention.

Fig. 1.

Trichotomy models of attentional control. (A) Basic trichotomy model (adapted from Awh et al., 2012). The focus is on three sources of input into an integrated priority map. (B) Updated trichotomy model proposed by Anderson et al. (2021). The adapted version depicted here is a simplified version of the Anderson et al. (2021) model that focuses on the different sources of input impinging upon the computation of attentional priority, which are assumed to compete at multiple stages of priority computation (not necessarily all integrated at the same processing stage).

Fig. 2.

Mechanisms of assigning attentional priority without any organizational framework, eschewing a distinction between the influence of current goals and selection history. Rather than fractionate selection history as in Fig. 1B, each factor influencing attentional priority is represented as functionally distinct and therefore gets its own box. Models like this have not been given realistic consideration in the field but may not be without some virtue that will be better captured by the proposed model.

Fig. 3.

The proposed model of attentional control. There is only one mechanism for assigning attentional priority, subsumed within the attentional control state. Different memory systems collectively comprise the attentional control state, several of which contain representational overlap. Note that the overlap in the circles depicts representational overlap with respect to the architecture underlying the attentional control state, not overlap in the features or locations to which adjustments in priority are applied (which can covary; e.g., the same feature can be prioritized according to both its positive valence and relevance to current goals). Physical salience is depicted in the weight (thickness) of the arrows reflecting different sources of sensory input, which is modified by the control state. Priority weights for features are depicted by colored circles with a plus or minus sign, and priority weights applied to representations of spatial information are depicted with a heat map; the choice of which to link to a given memory system is simply intended to depict one possible manifestation of the attentional control state and is not intended to reflect the full scope of any given source of priority (e.g., there is strong evidence that attentional priority based on statistical learning tied to historical relevance can be spatial in nature in addition to feature-based; see, e.g., Wang & Theeuwes, 2018a, 2018b, 2018c). An “endpoint representation” is depicted as the target of perceptual input that has been modulated by the attentional control state rather than a priory map specifically, as a priority map reflects only one of many perceptual representations that could be affected by the control state.

Fig. 4.

Example of attentional priority computations unfolding over time as hypothesized by the proposed model. Nodes in a hypothetical neural network are indicated by donut-shaped circles, and the flow of neural signals by arrows, with dotted lines reflecting neural signals that are tonically active in a particular context and solid lines reflecting sensory-evoked activity emanating from the eyeball. The thickness of the lines and donut-shaped circles denotes the strength of corresponding neural activity. (A) Tonic attentional control signals in which memory-related activity influences how cells within nodes of the network for visual information processing will respond to input consistent with the color indicated, reflecting red as a goal-relevant color and blue as a color prioritized via selection history. (B) As visually evoked signals move through the network, signals prioritized by the attentional control state are amplified at the points at which they come into contact with tonic attentional control signals. This amplification reflects phasic attentional control signals, which then feed forward through the network. Although the figure only depicts feedforward signals for the sake of simplicity, phasic attentional control signals also likely involve some degree of back-propagation as well. Note that the nodes depicted in the network are not intended to be inclusive; rather, a select few notes were chosen to illustrate the concept.