Current cog Lecture 5

Attention

• Selection of the sensory input

• Unattended → unaware, inattentional blindness

• Looking is not attending

Conversely: Attention improves perception

• Example from spatial cueing paradigm

• Fixate on middle dot. A cue then draws attention to a location.

• Task: discriminate the target (appearing at cued or non-cued location)

• Primary measures: RT (ms) and accuracy (% correct)

Reason for attention

• Inattentional blindness: shows we cannot process everything up to the same level of awareness

• Brain has insufficient capacity or mental resource to process all sensory information

• Must make a selection (hence often referred to as selective attention)

• Information processing has a bottleneck: From parallel to serial

Taxonomy I: Targets of attention

• Attention comes with a sense of direction. To what?

  • Internal attention: Own thoughts, memories, and action plans (intentions)

  • External attention: Sensory input

    • Different modalities

    • Space

    • Time

    • Features

    • Objects

Taxonomy II:Sources of attention

• What is then directing attention?

• Stimulus-driven attention: Attention drawn to salient stimulus

  • “Salience-driven”; “Bottom-up”; “Exogenous”; “Involuntary”; “Automatic”; “Feedforward-based”

• Goal-driven attention: Attention initiated by current behavioural requirements

  • “Top-down”; “Endogenous”; “Voluntary”; “Controlled”; “Feedback-based”

• Experience-driven attention: Attention driven by learned context or value

  • “Selection history”; “Reward”; “Priming”; “Automatic”

Example from the lab: Spatial cueing

• Task: Search for a target letter (e.g. “R” or “B”), press button accordingly

• Stimulus-driven cue (peripheral onset)

• Goal-driven cue (central arrow): may tell you where the target is

• Cue is either valid (target appears at indicated position) or invalid (target appears at other position)

Other taxonomies

• Selective attention versus divided attention

• Transient attention versus sustained attention

Attention in the real world

• What are the characteristics here that attract attention?

• Attention is “caught”: Attentional capture

• Attention driven by some feature contrast

• Attention driven to a location

In the lab: Color saliency

• Classic study from Theeuwes (1992)

• Task: Look for deviant shape; ignore deviant color

  report the orientation of the little line segment (horizontal/vertical), as fast as you can

• Primary measure: Reaction time (RT);

  Secondary measure: Response accuracy (% error)

• Two important conditions: No salient distractor vs. Salient distractor

• Conclusion: Salient distractor interferes, so must have captured attention

• Follow-up study looking at eye movements (Theeuwes et al. , 2003)

• Same design, stimuli & task

• Primary measures: First saccade direction, fixation duration

• Conclusion: saliency also briefly captures the eyes

In the lab: EEG

• Some event-related potentials (ERPs) in the EEG signal reflect attention

• Most prominent is the N2pc

• Posterior, contralateral, negative component, lateralized (left or right), around 250 ms after stimulus onset

• There also exists a positive potential, reflecting suppression, called the Pd

• Applying this EEG technique to saliency (Hickey et al., 2006; JOCN)

• Same design, stimuli & task as Theeuwes (1992)

• Primary measure: Attention-related contralateral negative component, called N2pc

• Results: N2pc contralateral to target, but also contralateral to salient distractor

• Conclusion: distractor attracts attention

In the lab: Abrupt onset saliency

• Grey items change to red, except one. Letters appear inside.

• Task: Make an eye movement to the grey object, and report its letter (C or reverse C) ignore the abrupt onset of a new object

• Primary measures: RT; accuracy of first saccade (% to target)

• Result: ~40% of saccades directed towards irrelevant onset

• Secondary measures:

  • Saccadic latency

  • Saccadic trajectory

• Conclusion: Onsets capture attention automatically; Occurs fast & early

In the lab: Orientation saliency

• Grid of orientations, of which two items deviate; Vary orientation of background items, so that one of the items is more salient than the other

• Task: make an eye movement to left-tilted item, and ignore the right-tilted item

• Primary measure: saccadic accuracy (% towards the target, versus % towards salient item)

• Two main conditions: target can be the more salient one, or the less salient one (and we also varied eccentricty)

In the lab: Visual search

• Task: Find the unique object (target)

• Condition 1: Here defined by unique feature: “feature search”

• Important: Vary the set size (= number of items in the display; also called display size)

• Primary measure: RT as a function of set size, referred to as search slope

• Results: Detection does not suffer from additional items, slope is flat

• Condition 2: Here defined by combination of features: “conjunction search”

• Results: Detection suffers from additional items, slope is steep

• Conclusions:

  • Salient feature contrasts are detected in parallel across the visual field: before the bottleneck

  • Nonsalient contrasts require serial scanning: bottleneck

Control over capture

• Can you prevent capture?

• Same design, stimuli & task as Theeuwes (1992)

• Two conditions: Distractor at any position (random, low probability) vs.

  distractor most frequently at one particular position (high probability)

• Results: Frequent distractor location suppressed or avoided

• Conclusion: Some control possible with extensive repetition

Feature maps

• Different areas have different visual functions

• Even for basic physical features like intensity, orientation, motion, color

• We can represent this schematically in maps like this, called feature maps

A Saliency Model

• Itti & Koch (2000) built a computer model that does the same

• How? → use image filters

• Feature values are then combined into an overall saliency map

• Activity in the saliency map represents the relative saliency of each location

• A next step is called winner- take-all, to make the model actually choose the most salient location

• It then makes an “eye movement” towards it

• Finally, the location is suppressed, and the cycle repeats

Where is this saliency map?

• Complex network of brain areas thought to be involved

• A more ventral system that detects the salient features` (what)

• A more dorsal system that flags important locations and triggers motor response (where)

• fMRI work from Corbetta & Shulman (early 2000s)

• A ventral network driving stimulus-driven selection

• A dorsal network driving goal-driven selection

• But consider overlap

Alternative models of saliency: The Information Theoretic approach

• Do not need to transmit what is known, only what is not known

• Let’s apply these principles to an image

• We can build these principles into a computational model, where salient = maximal information value

• And then compare it to human eye movements

Alternative models of saliency: The Bayesian Surprise approach

• Yes, we can predict part of an image from another image part (Bruce & Tsotsos, 2009)

• Better is to predict from the observer’s expectations

• Example: Watching TV

• From Shannon Information Perspective, white snow should continually attract attention, since maximally unpredictive

• Bayesian approach: Update our observer model from “I’m watching CNN” to “I’m watching snow”

• After that the snow loses its surprise value

• Under this model: salient = largest update of our beliefs

• Results: Predicts eye movements even better