Visual adaptation and face perception

Abstract

The appearance of faces can be strongly affected by the characteristics of faces viewed previously. These perceptual after-effects reflect processes of sensory adaptation that are found throughout the visual system, but which have been considered only relatively recently in the context of higher level perceptual judgements. In this review, we explore the consequences of adaptation for human face perception, and the implications of adaptation for understanding the neural-coding schemes underlying the visual representation of faces. The properties of face after-effects suggest that they, in part, reflect response changes at high and possibly face-specific levels of visual processing. Yet, the form of the after-effects and the norm-based codes that they point to show many parallels with the adaptations and functional organization that are thought to underlie the encoding of perceptual attributes like colour. The nature and basis for human colour vision have been studied extensively, and we draw on ideas and principles that have been developed to account for norms and normalization in colour vision to consider potential similarities and differences in the representation and adaptation of faces.

Figure 1.

An illustration of a negative after-effect from light adaptation.

Figure 2.

Examples of stimuli used to measure face after-effects. (a) A single face distorted by locally stretching or shrinking the face along the vertical or horizontal axis. (b) Faces and their ‘antifaces’ formed by projecting the individual face through an average face.

Figure 3.

After-effects induced at different points in face space predicted by renormalization (left plots) or repulsion (right plots). Uniform renormalization predicts a constant shift in the appearance of all faces after adapting to a non-average face (top left), while no after-effects after adapting to the norm (bottom left). Repulsion predicts no shift in the appearance of the adapting image, and shifts away from the adapt for surrounding faces, with similar after-effects predicted for both non-average (top right) and average (bottom right) faces.

Figure 4.

After-effects along trajectories passing either through (A1-T) or not through (A2-T) the norm, predicted by normalization or repulsion. (a) Normalization predicts each adapting face shifts appearance relative to the average. This produces the largest shifts along trajectories through the norm (A1-T) and weaker shifts along axes that do not intersect the norm. Along each axis the predicted shift (solid arrows) is given by the projection of the adapt vector (dashed arrows) onto the test vector, and thus by the cosine of the angle between the adapt and the test vectors. (b) Repulsion instead predicts that each adapt image will bias the test away from the adapt and thus predicts equivalent after-effects along either adapt–test axis for images at equivalent distances.

Figure 5.

Norms and channel coding along a single perceptual dimension. The dimension may be represented by a small number of broadly tuned channels (left) or a large number of narrowly tuned channels (right). (a) The norm is represented implicitly by equal responses in the two channels. Adaptation to a biased stimulus reduces the response in one channel more than the other and shifts the norm from the one prevailing under neutral adaptation (N) towards the adapting level (A). This produces a shift in the appearance of all faces in the direction indicated by the arrows. (c) A split-range code in which the norm is represented explicitly by a null point between the responses of two channels that each respond to signals higher or lower than the norm (e.g. because each is formed by the half-wave rectified differences between the input channels at top left). Adapting to a new stimulus level (A) shifts the inputs and thus the null point. (b) If both the stimulus and the channels are narrowband, then adaptation will reduce the response at the adapting level and skew the responses of other stimuli away from the adapting level. In this case there is not a unique norm. (d) However, if the stimulus is broadband, then a norm is again implicitly represented by equal responses across the set of channels, and again renormalizes from the neutral stimulus (N) when the adapting stimulus is biased (A).

Figure 6.

Norms and channel coding along two perceptual dimensions (gender and ethnicity). (a) The two dimensions represented by two independent channels. Adapting to stimuli that covary in gender and ethnicity (e.g. female Asian versus male Caucasian) should produce independent response changes along each axis and thus will not lead to contingent after-effects. (b) If the stimulus dimensions are instead encoded by multiple mechanisms selective for different combinations of the two attributes, then adapting to covarying stimuli will tilt the appearance of intermediate axes away from the adapting axis, resulting in contingent after-effects.

Figure 7.

The six basic expressions (top row) and their anti-expressions (bottom row), corresponding to faces with the opposite configuration.