Until relatively recently, most theorists (e.g., Biederman, 1987) studying object recognition emphasised bottom-up processes. Apparent support for this emphasis can be found in the hierarchical nature of visual processing – during the course of visual processing, neurons higher up the hierarchy respond to increasingly complex stimuli. As Yardley et al. (2012, p. 4) pointed out:
Traditionally, visual object recognition has been taken as mediated by a hierarchical, bottom-up stream that processes an image by systematically analysing its individual elements, and relaying this information to the next areas until the overall form and identity are determined.
The traditional account focuses on a feedforward hierarchy of processing stages progressing from early visual cortex through to inferotemporal cortex.
However, anatomical evidence suggests this is a considerable oversimplification. There are approximately equal numbers of forward and backward projecting neurons throughout most of the visual system (Wyatte et al., 2012; Gilbert & Li, 2013). In essence, backward projecting neurons are associated with top-down processing.
There is an important issue concerning when top-down processes have their effects. Top-down processes may occur only after object recognition and may relate to semantic processing of already recognised objects. Alternatively (and of more theoretical interest), top-down processes (perhaps involving the prefrontal cortex) may occur prior to object recognition and may be necessary for recognition to occur. Before we discuss the relevant research findings, note that top-down processes are more likely to have a major impact on object recognition when bottom-up processes are relatively uninformative (e.g., degraded stimuli, briefly presented stimuli).
Findings
Evidence for the involvement of top-down processes in visual perception has been obtained from research on ambiguous figures having at least two different interpretations. Goolkasian and Woodberry (2010) presented participants with ambiguous figures immediately preceded by primes relevant to one interpretation (see Figure 3.9). The key finding was that the primes systematically biased the interpretation of the ambiguous figures via top-down processes.
Bar et al. (2006) presented participants briefly with drawings of objects that were then masked to make them hard to recognise. Activation in orbitofrontal cortex (part of the prefrontal cortex) occurred 50 ms before activation in recognition-related regions of the temporal cortex. This orbitofrontal
activation predicted successful object recognition and so seemed important for object recognition to occur. There was less involvement of orbitofrontal cortex in object recognition when recognition was easy (longer, unmasked presentation).
Figure 3.9
Ambiguous figures (e.g., Eskimo/Indian, Liar/Face) were preceded by primes (e.g., Winter Scene, Tomahawk) relevant to one interpretation of the following figure.
From Goolkasian and Woodberry (2010). Reprinted with permission from the Psychonomic Society 2010.
Figure 3.10
In this modified version of Bar et al.’s (2006) theory, it is assumed that object recognition involves two different routes: (1) a top-down route in which information proceeds rapidly to the orbitofrontal cortex, which is involved in generating predictions about the object’s identity; (2) a bottom-up route using the slower ventral visual stream.
From Yardley et al. (2012). Reprinted with permission from Springer.
Bar et al. (2006) concluded that top-down processes in orbitofrontal cortex are more important when recognition is difficult than when it is easy. They put forward a model in which object recognition depends on top-down processes involving the orbitofrontal cortex and bottom-up processes involving the ventral visual stream (see Figure 3.10).
Stronger evidence that top-down processes in the prefrontal cortex play a direct role in object recognition was reported by Viggiano et al. (2008).
Participants viewed blurred or non-blurred photographs of living and non-living objects under four conditions: (1) repetitive transcranial magnetic stimulation (rTMS; see Glossary) applied to the left dorsolateral prefrontal cortex; (2) rTMS applied to the right dorsolateral prefrontal cortex; (3) sham rTMS (there was no magnetic field); and (4) baseline (no rTMS at all).
What did Viggiano et al. (2008) find? First, rTMS applied to the left or right dorsolateral prefrontal cortex slowed down object-recognition time.
Second, rTMS had no effect on object-recognition time with non-blurred photographs. These findings suggest top-down processes are directly involved in object recognition when the sensory information available to bottom-up processes is limited.
Suppose we presented photographs of objects (e.g., mailbox, tractor) to one eye while at the same time presenting high-contrast noise patterns to the other eye to suppress visual awareness of the object. Suppose also that each photograph was preceded by a valid verbal cue (i.e., indicating which object was to be presented), an invalid verbal cue (i.e., indicating an incorrect object was to be presented) or no cue. We might not be very impressed if observers were best at judging which object had been presented when valid cues were used – the valid cues might simply have influenced their judgements about their visual experience rather than perception itself.
An alternative approach could in principle provide more striking evidence that top-down processes influence basic perceptual processes. In essence, a suppressed stimulus is presented on some trials but there is no stimulus on others. Observers decide whether or not a stimulus was presented. The key prediction is that valid cues should lead to superior stimulus-detection performance than invalid cues.
Lupyan and Ward (2013) carried out a study based on the ideas discussed above. Observers were told the suppressed stimulus would be a circle or square or there was no cue. What was presented was a circle, a square, a shape intermediate between a circle and a square or no stimulus. Observers’
performance was significantly better when valid (rather than invalid) cues were used (see Figure 3.11). Thus, top-down processes triggered by the verbal labels activated shape information and influenced basic visual detection.
Figure 3.11
(a) hit rates and (b) detection sensitivity for detecting geometric shapes as a function of shape (square, intermediate, circle) and cue (circle, no cue, square). Performance was best when the cue was valid (matched the subsequent shape).
From Lupyan and Ward (2013). © National Academy of Sciences. Reproduced with permission.
Evaluation
Top-down processes often influence visual perception and object recognition and not simply post-perceptual judgements. As Yardley et al. (2012, p. 1) pointed out, “During our attempts to interpret the world around us, perception relies on existing knowledge as much as it does on incoming information.”
Note, however, that the influence of top-down processes is generally greater when visual stimuli are degraded.
So far as the future is concerned, it is of central importance to understand more fully how bottom-up and top-down processes interact with each other.
These interactions probably occur at several levels within the visual system. However, the factors determining whether and where they occur are largely unknown.
FACE RECOGNITION
There are several reasons for devoting a separate section to face recognition. First, recognising faces is of enormous importance to us. We generally identify people from their faces. Form a visual image of someone important in your life. Your image probably contains fairly detailed information about their faces and its special features rather than their physique.
Second, face recognition seems to differ in important ways from other forms of object recognition. As a result, theories of object recognition are of
only limited value in explaining face recognition and theoretical approaches specifically devoted to face recognition are needed.
Third, we now have a good understanding of the processes involved in face recognition. One reason for this is the diversity of research – it includes behavioural studies, studies on brain-damaged patients and neuroimaging studies.