- Email: [email protected]
Recognition of unfamiliar faces: three kinds of effects
Update Meetings Max Coltheart (Macquarie Centre for Cognitive Science, Sydney, Australia) who introduced the notion of proximal and distal causes of cognitive disorders, and encouraged researchers to focus on the proximal causes; that is, whatever is abnormal about the cognitive system required to carry out a certain cognitive task. Professor Coltheart closed his presentation with a summary that questioned the role of cognitive neuroscience in general – is it to concern itself with the exploration of cognition itself or its neural substrates? A day of presentations on language processing would not have been complete without some discussion of neural networks and Kim Plunkett (University of Oxford, UK) provided such a discourse. In an elegant demonstration of a seemingly counterintuitive contention, a double-dissociation within a single-route connectionist model of regular and irregular noun and verb morphology was explained. Professor Plunkett's
didacticism ensured that the non-specialists in the audience maintained interest as he then went on to describe the importance of convergent data from neural network models, developmental disorders and adult cognitive neuropsychology in the study of all language processes. The final day of the school was dedicated to a series of presentations on fMRI. This topic has appeared in previous autumn schools and, given its importance for cognitive neuroscience, will probably appear in future events too. Some of the morning presentations described the principal components of the MRI signal, thus allowing a brief respite from the standard ‘blobology’ approach – what some would consider as inferring too much from activation patterns in brain images. The afternoon session saw the presentation of a number of fMRI studies examining both normal and clinical populations. Key to these discussions was the notion that fMRI signals can only imply involvement of a
particular region of the brain in the function under study – it does not allow us to say that that brain region is responsible for that function; it is all too easy to make gross assumptions about structure/function relationships. Certainly a combination of approaches (e.g. TMS and fMRI) such as described at this year’s autumn school would seem to be the best route for functional elucidation of a given brain area. All four days of talks, coupled with the luxurious accommodation at Wadham College, led to a thoroughly enjoyable experience. The organisers, Sally Harte and Edmund Rolls, should be commended for staging such an event, which we are sure will be well attended in years to come.
Carl Senior and Tamara Russell Institute of Psychiatry, King’s College London, 103 Denmark Hill, De Crespigny Park, London, UK SE5 8AF. e-mail: [email protected]
Correspondence
Recognition of unfamiliar faces: three kinds of effects An intriguing phenomenon in face recognition that has been demonstrated repeatedly is that although people are remarkably good at recognizing familiar faces, they are surprisingly poor and error prone with unfamiliar ones. A recent review by Hancock et al. describes how recognition of unfamiliar faces can be easily disrupted by variations such as lighting, viewpoint, facial expression, external features and image quality1. However, recognition of familiar faces is often robust and effortless under such conditions. This discrepancy raises important issues regarding the mechanisms underlying familiar versus unfamiliar face recognition. To resolve this issue, it is first necessary to uncover the exact factors that are responsible for the fallible nature of unfamiliar face recognition. Here, we suggest three possible determinants: a so-called ‘memory factor’, a ‘perceptual factor’, and an ‘object-related factor’. The failure to recognize a face can be due to a lack of compatibility between the images of the face used at encoding and at retrieval (memory factor), deterioration of image qualities, such as insufficient luminance contrast (perceptual factor), or disruption of configural information, as with inversions (object-related factor). The effects that arise from these different factors might be explained by some general principles found within these domains. The memory factor can be explained by the principle of encoding specificity,
which predicts performance from the degree of congruity between stimulus information presented at the time of encoding and retrieval; the perceptual factor can be explained by a minimal requirement of contrast for pattern vision; and the object-related factor can be explained by a preferred (upright) orientation required for effective processing of configural or holistic information in faces (or other types of homogenous objects). There has been little effort at examining how these factors differentially affect recognition of unfamiliar faces. The advantage of separating these factors is obvious. Because each factor is likely to be determined by different underlying mechanisms, distinction between them would provide insight into how such mechanisms adapt to handling face-recognition tasks more effectively as a face becomes increasingly familiar. A first step towards understanding the role of these factors might be to investigate their relative contributions to the known behavioral effects. For example, we can identify two factors in the photographic negative effect, where faces are harder to recognize in negative contrast than in positive. One pertains to the contrast reversal itself in the negative image and the other to the incongruity in contrast polarities between the learned and test face stimuli. The contribution of each can be assessed by running a condition in which faces are both learned and tested
in negative or positive contrast (congruent stimulus condition) and comparing with a condition where contrast polarity is switched from positive to negative or vice versa between learned and test faces (incongruent stimulus condition). Using this paradigm, it has been shown that recognition performance in the congruent condition was far superior, suggesting that the photographic negative effect was due mainly to incongruent stimulus information2. In this example, the main reason for recognition failure can be explained by the principle of encoding specificity, whereas an object-related factor (contrast polarity) played a less important role. Encoding specificity might be able to explain many other effects found in recognition of unfamiliar faces, for example, the effect of spatial-frequency overlap between learned and test face images. It has been shown that even a slight variation in frequency overlap can affect recognition performance3. However, in the presence of perfect overlap, recognition performance becomes highly robust for a broad range of object frequencies, even when filtered images fall outside the range predicted to be recognizable by the theory of critical frequencies (around 8–16 cycles per face). It therefore appears that image congruity, as shown in this example by frequency overlap, is a better predictor of recognition performance than the theory of critical frequencies.
1364-6613/00/$ – see front matter © 2000 Elsevier Science Ltd. All rights reserved.
PII: S1364-6613(00)01558-8
Trends in Cognitive Sciences – Vol. 4, No. 12,
December 2000
445
Update Monitor
Other effects that are accounted for by encoding specificity include variations in a number of other image parameters, such as image size, colour, shading, texture and temporal resolution. The impairment in recognition of unfamiliar faces is often caused by incongruent image parameters between learning and test conditions. A combination of discrepancies in several image parameters, such as incongruent resolution plus incongruent image size, can have a multiplicative effect. This implies that better performance should be expected in recognition of unfamiliar faces when efforts are made to equate the image parameters of test image and the learning image. For example, Burton et al.4 have shown that facial identification from a poor quality video was severely impaired when the video clips were matched with high quality photographs. This result can be explained by a number of discrepancies between the two types of stimuli. First, because the poor video image did not contain high-spatial-frequency facial information, its presence in the high quality photographs more probably served as noise that interfered with the detection of information contained in the low spatial frequencies. Second, the face stimuli contained motion at encoding whereas only static images were used in testing. And third, there was an incongruity in image size between learning and
test conditions. If these stimulus parameters were equalized – for example, by resizing and defocusing the test photographs – it is likely that recognition performance would have improved even if the quality of the video images had remained the same. It is possible that our true ability to recognize unfamiliar faces is better than we currently believe if other conditions that cause generic memory-related deficits are eliminated. For example, the number of distractors is known to correlate inversely with memory performance. A striking effect of multiple distractors in a face matching task has been demonstrated by Bruce et al., whose subjects had considerable difficulty in identifying unfamiliar faces even when the target was shown simultaneously among a number of distractors5. This effect suggests that, unlike recognition of familiar faces, unfamiliar face recognition is highly sensitive to signal-tonoise ratios. When the number of distractors is minimized, recognition of unfamiliar faces can be remarkably good, provided that the image parameters remain congruent. For example, using a two-alternative forced choice task, Leder has shown that subjects’ matching performance reached a maximum even when the photographs of faces were presented in different views (frontal and three-quarter views)6. Given these considerations, we suggest that
efforts be made to identify the factors that contribute to the reduced performance in unfamiliar face recognition tasks. Hopefully, this will eventually enable us to determine how the mechanisms for the different factors accommodate variations in facial stimuli when they become more and more familiar.
Summaries of recently published papers of interest to cognitive scientists. Readers who would like to contribute to this section, by identifying appropriate papers and writing short summaries, should contact the Editor ([email protected]).
infants are capable of integrating multiple cues when learning new words 3. The question remains whether non-human primates have the same ability. To determine the degree to which language depends on mechanisms unique to humans, we need to seek answers to such questions in comparative studies, such as the one by Ramus et al., rather than revert to a priori assumptions about what makes human language special.
Chang Hong Liu and Avi Chaudhuri Department of Psychology, McGill University, 1205 Dr. Penfield Ave, Montréal, Québec, Canada H3A 1B1. tel: +1 514 398 6151 fax: +1 514 398 4896 e-mail: [email protected]
References 1 Hancock, P.J.B. et al. (2000) Recognition of unfamiliar faces. Trends Cognit. Sci. 4, 330–337 2 Liu, C.H. and Chaudhuri, A. (1997) Face recognition with multi-tone and two-tone photographic negatives. Perception 26, 1289–1296 3 Liu, C.H. et al. (2000) The effects of spatial frequency overlap on face recognition. J. Exp. Psychol. Hum. Percept. Perform. 26, 956–979 4 Burton, A.M. et al. (1999) Face recognition in poor quality video: evidence from security surveillance. Psychol. Sci. 10, 243–248 5 Bruce, V. et al. (1999) Verification of face identities from images captured on video. J. Exp. Psychol. Appl. 5, 339–360 6 Leder, H. (1999) Matching person identity from facial line drawings. Perception 28, 1171–1175
Monitor
Monkey hears, monkey does – like infants? To what degree is human language dependent on species-specific mechanisms? It seems clear that at least some of our language ability depends on mechanisms for auditory perception shared with other species. For example, it is known that chinchillas can perceive the same phonetic distinctions as human infants. A new study by Ramus et al. suggests that the degree of species-specificity in the mechanisms of language processing is even less than thought previously1. They used two equivalent habituation–dishabituation procedures to compare the ability of newborn human infants and cotton-top tamarin monkeys to distinguish between Dutch and Japanese sentences. The sentences were played either forwards or backwards, and the results showed that tamarins, like human infants, can distinguish between Dutch and Japanese when the sentences are played forwards (as in normal speech), but not
446
when played backwards. Thus, the tamarins showed sensitivity to the linguistic properties that distinguish Dutch and Japanese – properties that are presumably eliminated when the speech stimuli are played backwards. This finding suggests that a significant part of the mechanism or mechanisms used for speech perception might be shared with other nonhuman primates. However, no non-human primate appears to have a communication system that comes close to human language in terms of flexibility and complexity. So, what sets us apart? In an accompanying commentary 2, Werker and Vouloumanos point to the intriguing possibility that what is unique to human language is not basic auditory processing but the ability to integrate different properties of speech in just the right way for the acquisition and processing of language. We know that human
1364-6613/00/$ – see front matter © 2000 Elsevier Science Ltd. All rights reserved. Trends in Cognitive Sciences – Vol. 4, No. 12,
December 2000
References 1 Ramus, F. et al. (2000) Language discrimination by human newborns and by cotton-top tamarin monkeys. Science 288, 349–351 2 Werker, J.F. and Vouloumanos, A. (2000) Who’s got rhythm? Science 288, 280–281 3 Mattys, S.L. et al. (1999) Phonotactic and prosodic effects on word segmentation in infants. Cognit. Psychol. 38, 465–494
Monitor contributors We thank this month’s contributors, who are:
Morten Christiansen If you would like to contribute to Monitor, please contact the Editor at [email protected]. Regular contributors receive a free personal subscription to the journal.
didacticism ensured that the non-specialists in the audience maintained interest as he then went on to describe the importance of convergent data from neural network models, developmental disorders and adult cognitive neuropsychology in the study of all language processes. The final day of the school was dedicated to a series of presentations on fMRI. This topic has appeared in previous autumn schools and, given its importance for cognitive neuroscience, will probably appear in future events too. Some of the morning presentations described the principal components of the MRI signal, thus allowing a brief respite from the standard ‘blobology’ approach – what some would consider as inferring too much from activation patterns in brain images. The afternoon session saw the presentation of a number of fMRI studies examining both normal and clinical populations. Key to these discussions was the notion that fMRI signals can only imply involvement of a
particular region of the brain in the function under study – it does not allow us to say that that brain region is responsible for that function; it is all too easy to make gross assumptions about structure/function relationships. Certainly a combination of approaches (e.g. TMS and fMRI) such as described at this year’s autumn school would seem to be the best route for functional elucidation of a given brain area. All four days of talks, coupled with the luxurious accommodation at Wadham College, led to a thoroughly enjoyable experience. The organisers, Sally Harte and Edmund Rolls, should be commended for staging such an event, which we are sure will be well attended in years to come.
Carl Senior and Tamara Russell Institute of Psychiatry, King’s College London, 103 Denmark Hill, De Crespigny Park, London, UK SE5 8AF. e-mail: [email protected]
Correspondence
Recognition of unfamiliar faces: three kinds of effects An intriguing phenomenon in face recognition that has been demonstrated repeatedly is that although people are remarkably good at recognizing familiar faces, they are surprisingly poor and error prone with unfamiliar ones. A recent review by Hancock et al. describes how recognition of unfamiliar faces can be easily disrupted by variations such as lighting, viewpoint, facial expression, external features and image quality1. However, recognition of familiar faces is often robust and effortless under such conditions. This discrepancy raises important issues regarding the mechanisms underlying familiar versus unfamiliar face recognition. To resolve this issue, it is first necessary to uncover the exact factors that are responsible for the fallible nature of unfamiliar face recognition. Here, we suggest three possible determinants: a so-called ‘memory factor’, a ‘perceptual factor’, and an ‘object-related factor’. The failure to recognize a face can be due to a lack of compatibility between the images of the face used at encoding and at retrieval (memory factor), deterioration of image qualities, such as insufficient luminance contrast (perceptual factor), or disruption of configural information, as with inversions (object-related factor). The effects that arise from these different factors might be explained by some general principles found within these domains. The memory factor can be explained by the principle of encoding specificity,
which predicts performance from the degree of congruity between stimulus information presented at the time of encoding and retrieval; the perceptual factor can be explained by a minimal requirement of contrast for pattern vision; and the object-related factor can be explained by a preferred (upright) orientation required for effective processing of configural or holistic information in faces (or other types of homogenous objects). There has been little effort at examining how these factors differentially affect recognition of unfamiliar faces. The advantage of separating these factors is obvious. Because each factor is likely to be determined by different underlying mechanisms, distinction between them would provide insight into how such mechanisms adapt to handling face-recognition tasks more effectively as a face becomes increasingly familiar. A first step towards understanding the role of these factors might be to investigate their relative contributions to the known behavioral effects. For example, we can identify two factors in the photographic negative effect, where faces are harder to recognize in negative contrast than in positive. One pertains to the contrast reversal itself in the negative image and the other to the incongruity in contrast polarities between the learned and test face stimuli. The contribution of each can be assessed by running a condition in which faces are both learned and tested
in negative or positive contrast (congruent stimulus condition) and comparing with a condition where contrast polarity is switched from positive to negative or vice versa between learned and test faces (incongruent stimulus condition). Using this paradigm, it has been shown that recognition performance in the congruent condition was far superior, suggesting that the photographic negative effect was due mainly to incongruent stimulus information2. In this example, the main reason for recognition failure can be explained by the principle of encoding specificity, whereas an object-related factor (contrast polarity) played a less important role. Encoding specificity might be able to explain many other effects found in recognition of unfamiliar faces, for example, the effect of spatial-frequency overlap between learned and test face images. It has been shown that even a slight variation in frequency overlap can affect recognition performance3. However, in the presence of perfect overlap, recognition performance becomes highly robust for a broad range of object frequencies, even when filtered images fall outside the range predicted to be recognizable by the theory of critical frequencies (around 8–16 cycles per face). It therefore appears that image congruity, as shown in this example by frequency overlap, is a better predictor of recognition performance than the theory of critical frequencies.
1364-6613/00/$ – see front matter © 2000 Elsevier Science Ltd. All rights reserved.
PII: S1364-6613(00)01558-8
Trends in Cognitive Sciences – Vol. 4, No. 12,
December 2000
445
Update Monitor
Other effects that are accounted for by encoding specificity include variations in a number of other image parameters, such as image size, colour, shading, texture and temporal resolution. The impairment in recognition of unfamiliar faces is often caused by incongruent image parameters between learning and test conditions. A combination of discrepancies in several image parameters, such as incongruent resolution plus incongruent image size, can have a multiplicative effect. This implies that better performance should be expected in recognition of unfamiliar faces when efforts are made to equate the image parameters of test image and the learning image. For example, Burton et al.4 have shown that facial identification from a poor quality video was severely impaired when the video clips were matched with high quality photographs. This result can be explained by a number of discrepancies between the two types of stimuli. First, because the poor video image did not contain high-spatial-frequency facial information, its presence in the high quality photographs more probably served as noise that interfered with the detection of information contained in the low spatial frequencies. Second, the face stimuli contained motion at encoding whereas only static images were used in testing. And third, there was an incongruity in image size between learning and
test conditions. If these stimulus parameters were equalized – for example, by resizing and defocusing the test photographs – it is likely that recognition performance would have improved even if the quality of the video images had remained the same. It is possible that our true ability to recognize unfamiliar faces is better than we currently believe if other conditions that cause generic memory-related deficits are eliminated. For example, the number of distractors is known to correlate inversely with memory performance. A striking effect of multiple distractors in a face matching task has been demonstrated by Bruce et al., whose subjects had considerable difficulty in identifying unfamiliar faces even when the target was shown simultaneously among a number of distractors5. This effect suggests that, unlike recognition of familiar faces, unfamiliar face recognition is highly sensitive to signal-tonoise ratios. When the number of distractors is minimized, recognition of unfamiliar faces can be remarkably good, provided that the image parameters remain congruent. For example, using a two-alternative forced choice task, Leder has shown that subjects’ matching performance reached a maximum even when the photographs of faces were presented in different views (frontal and three-quarter views)6. Given these considerations, we suggest that
efforts be made to identify the factors that contribute to the reduced performance in unfamiliar face recognition tasks. Hopefully, this will eventually enable us to determine how the mechanisms for the different factors accommodate variations in facial stimuli when they become more and more familiar.
Summaries of recently published papers of interest to cognitive scientists. Readers who would like to contribute to this section, by identifying appropriate papers and writing short summaries, should contact the Editor ([email protected]).
infants are capable of integrating multiple cues when learning new words 3. The question remains whether non-human primates have the same ability. To determine the degree to which language depends on mechanisms unique to humans, we need to seek answers to such questions in comparative studies, such as the one by Ramus et al., rather than revert to a priori assumptions about what makes human language special.
Chang Hong Liu and Avi Chaudhuri Department of Psychology, McGill University, 1205 Dr. Penfield Ave, Montréal, Québec, Canada H3A 1B1. tel: +1 514 398 6151 fax: +1 514 398 4896 e-mail: [email protected]
References 1 Hancock, P.J.B. et al. (2000) Recognition of unfamiliar faces. Trends Cognit. Sci. 4, 330–337 2 Liu, C.H. and Chaudhuri, A. (1997) Face recognition with multi-tone and two-tone photographic negatives. Perception 26, 1289–1296 3 Liu, C.H. et al. (2000) The effects of spatial frequency overlap on face recognition. J. Exp. Psychol. Hum. Percept. Perform. 26, 956–979 4 Burton, A.M. et al. (1999) Face recognition in poor quality video: evidence from security surveillance. Psychol. Sci. 10, 243–248 5 Bruce, V. et al. (1999) Verification of face identities from images captured on video. J. Exp. Psychol. Appl. 5, 339–360 6 Leder, H. (1999) Matching person identity from facial line drawings. Perception 28, 1171–1175
Monitor
Monkey hears, monkey does – like infants? To what degree is human language dependent on species-specific mechanisms? It seems clear that at least some of our language ability depends on mechanisms for auditory perception shared with other species. For example, it is known that chinchillas can perceive the same phonetic distinctions as human infants. A new study by Ramus et al. suggests that the degree of species-specificity in the mechanisms of language processing is even less than thought previously1. They used two equivalent habituation–dishabituation procedures to compare the ability of newborn human infants and cotton-top tamarin monkeys to distinguish between Dutch and Japanese sentences. The sentences were played either forwards or backwards, and the results showed that tamarins, like human infants, can distinguish between Dutch and Japanese when the sentences are played forwards (as in normal speech), but not
446
when played backwards. Thus, the tamarins showed sensitivity to the linguistic properties that distinguish Dutch and Japanese – properties that are presumably eliminated when the speech stimuli are played backwards. This finding suggests that a significant part of the mechanism or mechanisms used for speech perception might be shared with other nonhuman primates. However, no non-human primate appears to have a communication system that comes close to human language in terms of flexibility and complexity. So, what sets us apart? In an accompanying commentary 2, Werker and Vouloumanos point to the intriguing possibility that what is unique to human language is not basic auditory processing but the ability to integrate different properties of speech in just the right way for the acquisition and processing of language. We know that human
1364-6613/00/$ – see front matter © 2000 Elsevier Science Ltd. All rights reserved. Trends in Cognitive Sciences – Vol. 4, No. 12,
December 2000
References 1 Ramus, F. et al. (2000) Language discrimination by human newborns and by cotton-top tamarin monkeys. Science 288, 349–351 2 Werker, J.F. and Vouloumanos, A. (2000) Who’s got rhythm? Science 288, 280–281 3 Mattys, S.L. et al. (1999) Phonotactic and prosodic effects on word segmentation in infants. Cognit. Psychol. 38, 465–494
Monitor contributors We thank this month’s contributors, who are:
Morten Christiansen If you would like to contribute to Monitor, please contact the Editor at [email protected]. Regular contributors receive a free personal subscription to the journal.