The nature of holistic processing is a key channel of enquiry and this review focuses on three papers which each examine a different facet. These findings demonstrate that upright and inverted faces are processed differently and that there is a link between processing features with their surroundings and the face as a whole.
Keywords: Holistic processing, Face recognition, Learning, Composite Task
Unlike other objects, we tend to view faces entirely at once (Tanaka & Farah, 1993). There seems to exist some specialisation which allows faces to be viewed as more than the sum of its features. For us to improve our models of computer vision, it is necessary to learn the extent of skilled facial recognition and what this comprises of.
There are two main theories which explain the reliance on configuration: The expertise hypothesis suggests that as we view faces often, we become habituated with them and learn to recognise them quicker. The innate hypothesis however says is that we are born with the ability and it cannot be learned. There is some strong evidence for this such as babies preferring top-heavy, face like shapes to others (Simion, Cassia, Turati, & Valenza, 2001).
Young, Hellawell and Hay (1987) demonstrated for the first time that configural information only affects recognition when the face is portrait. To test this, they created composite photos of the top and bottom half of famous British males which were either aligned to create a standard face shape, or unaligned so that the bottom half was shifted 50% right. Subjects were then asked to name the identity of the top or bottom half of each face. Subjects on average spent an extra 212ms recognising aligned faces.
In the second experiment the same task was performed, but half of the stimulus was presented upside down. Interestingly the composite effect disappeared with only 9ms on average between recognising either composite or non-composite.
In the third experiment, Young et al. tested subjects exclusively trained on the top half of unfamiliar faces which allows us to see if the effect is only due to our familiarity with seeing the whole face. The result was the same as the original trial and subjects were slower at recognising composite faces (P < 0.01).
These experiments enforce the importance of configural information and conclude that this is only perceivable in the upright position. It may be the case that this is evidence of the precedence of global facial cues which we must ignore to identify only the top half of the face.
Following this, Robbins and McKone (2003) attempted to study whether holistic processing was innate or if it could be possible to learn to recognise inverted faces as a whole rather than by parts. To remove as much variance as possible two sets of identical twins were used. Subjects were shown over 1100 images of the twins, but no one recognised the inverted faces without relying on individual features. The only individuals who could successfully identify either twin used differences in eyebrows or eyelashes to distinguish them. In comparison, everyone who could distinguish them in the upright position failed to notice any individual feature which separated them.
They additionally tested the original group with the Composite Test pioneered by Young et al. (1987). As expected from previous research, a composite effect was perceived in the upright condition (P < 0.05) but not in the converse (P > 0.05). If a composite effect was evident in both conditions, it would be evidence that some level of holistic processing had been learned.
It is not possible from this study alone to reject the expertise hypothesis, but we can conclude that face recognition does not adapt to other orientations even after extended practise. This is contrary to the processing of other objects (McKone & Grenfell, 1999) and indicates that face recognition is unique.
Degutis et al. (2013) enquired into the method of analysis that previous studies had used to find if there was any correlation between the Cambridge Face Memory Test (CFMT), Part-Whole (PW) and Composite Test (CT). The CFMT requires subjects to pick a face out of a group of three in different lighting and positions. The PW experiment is derived from Tanaka & Farah (1993). This test asks the subjects to identify a feature either alone or a part of a face. The CT was the same as Young et al. (1987).
The default data analysis calculated by the sources they refer to is a subtraction method which, as the name implies, contains the difference between the control and the condition of interest (CI) times. However, a low subtraction score could arise from a poor control time or an above average CI time. Degutis et al. (2013) propose that using regression would instead compare each person’s CI performance to the CI performance of a typical subject with the same control performance. This instead means that a low regression score is separated from the variance in the control. The result of this is that we can draw stronger effects as we have less noise in our data. This means that we can use smaller sample sizes with the same reliability.
The results show that it is the aligned congruency effect (from the CT) rather than the PW effect that is key to the relationship between the two tests. When calculating PW and CT via regression, together they explained 21% of the variance in face recognition ability. This is significant as this gives us a clue into how large an effect holistic processing has on facial recognition.
Although it was not the focus of the study, it is also apparent from the analysis that the part and misaligned trials were significantly predictive of CFMT performance which implies featural processing does aid facial recognition.
From these three studies it is evident that holistic face processing is key to understanding facial recognition. Even when only the upper half is learned, the brain cannot view it separately from the lower half. This implies the face must be viewed in terms of configuration initially. This is due to faces being largely static, but features can be altered in terms of pigmentation and form. For example, we still recognise people when they have makeup, beards or pulling an odd expression.
McKone and Robbins (2003) research confirms that it is not possible to process inverted faces holistically through practice alone. The only way in which subjects could distinguish inverted faces was through finding featural clues.
Finally, DeGutis et al. (2013) lead us to a singular theory of facial processing, which takes the interdependence of facial features hand in hand with the viewing of the face as a whole. This correlation in turn aids face recognition ability. Results were also found which displayed that non-holistic processing is necessary for skilled facial recognition.
Further research could examine whether it is possible for non-experts to learn to recognise inverted stimulus at a similar level to experts. This may provide evidence that there is a blocking effect from their habituation with upright stimulus. Another avenue of research would be to investigate the scale of non-holistic processing by displaying features for short periods of time in sequence. It may also be illuminating to use inverted videos of subjects to see if motion could aid mental images - which could then transfer to still face processing.
Tanaka, J. W., & Farah, M. J. (1993). Parts and wholes in face recognition. The Quarterly Journal of Experimental Psychology A: Human Experimental Psychology, 46A(2), 225–245. https://doi.org/10.1080/14640749308401045
Simion, F., Cassia, V. M., Turati, C., & Valenza, E. (2001). The origins of face perception: Specific versus non-specific mechanisms. Infant and Child Development, 10(1-2), 59–65. https://doi.org/10.1002/icd.247
Young, A. W., Hellawell, D., & Hay, D. C. (1987). Configurational Information in Face Perception. Perception, 16(6), 747–759. https://doi.org/10.1068/p160747
Piepers DW, Robbins RA. A Review and Clarification of the Terms “holistic,” “configural,” and “relational” in the Face Perception Literature. Front Psychol. 2012;3:559. Published 2012 Dec 17. https://doi.org/10.3389/fpsyg.2012.00559
DeGutis J, Mercado RJ, Wilmer J, Rosenblatt A (2013) Individual Differences in Holistic Processing Predict the Own-Race Advantage in Recognition Memory. PLOS ONE 8(4): e58253. https://doi.org/10.1371/journal.pone.0058253