Attention and Action

Attention and Working Memory

Most contemporary theories link attention with a similarly multifaceted concept of working memory. In these theories, working memory (WM) denotes the set of processes that, together, maintain a model [Page 63]of the organism's current environment, as related to its ongoing behavioral goals. Obviously, the content of WM needs to be continuously updated. However, this updating is highly selective. On this view, the content of WM is itself provided by (i.e., is the outcome of) attentional processes. At the same time, WM is seen also as the primary source of top-down attentional control, such that goal-relevant processing is prioritized (selected).

In most laboratory studies of attention, relevant and irrelevant stimuli (and actions) are specified by the experimenter. Outside the lab, however, behavioral relevance is determined by an organism's entire goal hierarchy, with continuously varying priority accorded to unexpected threats and opportunities, as well as to the more orderly sequences of attention shifts needed for any skilled action. Ericsson and Kintsch argue that skilled human performance, from playing chess to making tea, requires not only information held in short-term WM but also rapid access to a vast set of retrieval structures in long-term memory. They call this long-term working memory.

Interest has centered on the role of various frontal (and parietal) brain structures, in particular the prefrontal cortex, in working memory functions. The emphasis here has been on the role of these brain structures in the active maintenance of patterns of activity representing current behavioral goals, and the means to achieve them. Influential models put forward by Desimone and Duncan, Miller and Cohen, and others propose that these activity patterns in the prefrontal cortex (and other structures, including the posterior parietal cortex) are the source of top-down bias signals, which serve to guide (i.e., selectively prioritize) the flow of neural activity throughout much of the brain. A brief summary of one of these models is provided in the next section.

The Integrated Competition Hypothesis

Desimone and Duncan's integrated competition hypothesis rests on three general principles.

• 1.
  Neural activity within a cell population is (locally) mutually competitive. Thus, increased activation in cells responding to one stimulus is accompanied by reduced activation in neighboring cells responding to others.
• 2.
  Top-down priming of cell responsiveness biases this competition in favor of activations relevant to current behavioral goals. Thus, sensory neurons that encode a relevant attribute or location show both enhanced spontaneous firing rates and increased gain in stimulus-evoked responses. Top-down bias signals are carried by “backward” cortico-cortical connections. (Backward connections are those running from higher to lower levels, e.g., in perceptual systems, running in the direction of the sensory inputs.) Backward connections are in fact more numerous anatomically than the classical “forward” connections. Their role in contextual biasing of conflict-prone neural processing is fundamental.
• 3.
  Finally, local competition is integrated across widely distributed components of the processing system. Thus, as one focus of processing gains dominance within a given (e.g., sensory) cell population, this advantage is propagated to other parts of the network, where cells coding for other (e.g., motor) properties consistent with or associated with the same entity (including its location, reward-value, etc.) gain further support, so that the network as a whole tends to cascade into a state in which the neural representation of one consistent object-goal-and-action is dominant throughout. This property has been explored in many different connectionist network models. It suggests a basis for the broadly “one thing at a time” character of much of perception and perceptual-motor integration.