Lecture 2: Attention

[Disclaimer: These informal lecture notes are not intended to be comprehensive - there are some additional ideas in the lectures and lecture slides, textbook, tutorial materials etc. As always, the lectures themselves are the best guide for what is and is not examinable content. However, I hope they are useful in picking out the core content in each lecture.]

Part 1: Definitions

The quote by William James at the start of the lecture is used to motivate three key questions. Here it is again in full:

Attention is…the taking into possession of the mind, in clear and vivid form, of one out of what seem several simultaneously possible objects or trains of thought.

The three questions posed in the lecture are:

List of different distinctions

From these three questions, we obtain several quite distinct aspects to attention. By considering the "how many" question, we arrive at the distinction between:

The "what kind" question leads us to consider the difference between

External attention is usually subdivided by sensory modality: visual attention is when you're paying attention to something you can see, auditory attention is directed at things you can hear, and cross-modal attention occurs when attention is directed to multiple senses at once (e.g., to the smell and taste of a cup of coffee)

Finally, the "what controls" question leads us to consider


This list of attentional "types" provides a pretty effective classification system for attentional phenomena. Consider the following examples:

A car backfires in my street, causing me to startle and look out the window

The attention here is:

Part 2: Phenomena in auditory attention

A lot of the early work in attention focused on auditory attention, apparently because that was easier to study with the technology of the time. A lot of this work is inspired by the so-called cocktail party problem - trying to pay attention to a single "source" (e.g., one person speaking) when there are many different sources competing for your attention (e.g., it's a very noisy room and there are many people talking at once)

Much of this research relied on a paradigm called the dichotic listening task, in which a participant wears headphones and two different auditory signals are presented, one to the left ear and the other to the right ear. The task is to shadow the content in one ear (e.g., right ear), by repeating aloud what you hear in the target (right) ear, and ignore the content in the non-target ear (left).

[Empirical result] A typical pattern of results (e.g., Moray 1959) is that people are very good at the task, shadowing the target signal successfully. However, when tested, they remember very little of the content presented in the non-target ear. They can perhaps recall whether it was a male or female voice, and a few other "low level" perceptual features, but semantic information (i.e., the content!) seems not to be recalled.

[Theoretical idea] These kinds of findings led Broadbent (1954) to propose an early selection view of attention, sometimes referred to as "Broadbent's filter". The idea is that the perceptual system does some "simple" perceptual analysis of every signal, detecting things like pitch, loudness, frequency, etc. On the basis of this preattentive analysis the filter selects a single signal to pay attention to. Only this selected signal is processed further, and as a consequence we are only aware of the semantic information (i.e., meaningful content) within the attended stream. This idea explains the basic empirical results described earlier, but it has some problems.

Early selection theory predicts that we use low-level perceptual features to select the target for attention. Yet there is evidence that people use semantic information to do so. It also predicts that we don't process the semantic content of the unattended stream, yet that also seems not to be correct.

[Empirical result] Here's the relevant findings from the lecture:

[Theoretical interpretation] There are two alternative theories that were discussed that are able to account for these findings.

In the empirical literature there are follow up studies that seek to distinguish between these possibilities, but they're beyond the scope of this lecture.

Part 3: Phenomena in visual attention

In the third part of the lecture we asked whether the phenomena listed above are idiosyncratic to audition, or whether they are general phenomena that could be found in any sensory modality. We discussed two studies in particular.

The key take home message from this is that there are some remarkable similarities in how attention operates across different sensory modalities.

Part 4: Visual search

The final part of the lecture discussed "visual search" tasks, in which the goal is to find a "target object" hidden among a collection of distractors (e.g., find my child among the crowd of children in the playground)

[empirical results] When the target is defined by a specific feature (e.g., colour) it seems to "pop out". Attention is automatically (i.e., passively) drawn to the target item. The set size (i.e., number of distractor items) makes no difference to the search time. This phenomenon is quite general, and doesn't depend on what kind of feature differentiates the target item: colour, shape, size, orientation, motion, depth, all produce pop out effects.

In contrast, when the target does not possess any unique features, there is no pop-out effect. For instance, if we need to find a red horizontal rectangle in a field of red vertical rectangles and blue horizontal rectangles, you need to make use of both features (i.e., orientation and colour) to solve the search problem. Search is slower, and now the set size matters: the more distractors there are, the slower you are to find the target.

[theoretical interpretation] The explanation for this proposed by Treisman (1986) is referred to as "feature integration theory". The idea is that the perceptual system has many different "feature analyzers" that detect perceptual features (e.g., red, blue, horizontalness etc). These operate quickly and in parallel - so if you can solve a visual search problem using only a single feature, then set size is irrelevant because the parallel nature of the feature analyzers means that you're processing every part of the visual input at the same time.

However, the feature analyzers are distinct from one another. Just because one analyzer has detected "redness" at a particular location and another has detected "horizontalness" at the same location doesn't mean that we automatically bind those two pieces of information together into a unified representation of the object. In order to do this feature binding, we need to direct attention to the location in question. Because attention is a slow, serial process (i.e., does one thing at a time), any visual search problem that requires feature binding (i.e., we need to use multiple features to solve it) will not produce a pop out effect, and visual search time will be slower when the set size is larger

[empirical data] One piece of evidence for feature integration theory (FIT) is the illusory conjunction phenomenon. FIT predicts that Ffature extraction occurs automatically and in parallel; but object recognition requires feature binding, a process that requires slow serial attentional processing of stimuli to be done accurately. If this is not allowed (e.g., stimuli are presented too quickly for attention to come into play), then errors in binding will occur and will be based on features extracted automatically during early perceptual processing. See lecture slides for illustration.