Two Factors in Face Recognition: Whether You Know the Person's Face and Whether You Share the Person's Race.

One of the best-known phenomena in face recognition is the other-race effect, the observation that own-race faces are better remembered than other-race faces. However, previous studies have not put the magnitude of other-race effect in the context of other influences on face recognition. Here, we compared the effects of (a) a race manipulation (own-race/other-race face) and (b) a familiarity manipulation (familiar/unfamiliar face) in a 2 × 2 factorial design. We found that the familiarity effect was several times larger than the race effect in all performance measures. However, participants expected race to have a larger effect on others than it actually did. Face recognition accuracy depends much more on whether you know the person's face than whether you share the same race.

How many faces do people know?

Over our species history, humans have typically lived in small groups of under a hundred individuals. However, our face recognition abilities appear to equip us to recognize very many individuals, perhaps thousands. Modern society provides access to huge numbers of faces, but no one has established how many faces people actually know. Here, we describe a method for estimating this number. By combining separate measures of recall and recognition, we show that people know about 5000 faces on average and that individual differences are large. Our findings offer a possible explanation for large variation in identification performance. They also provide constraints on understanding the qualitative differences between perception of familiar and unfamiliar faces-a distinction that underlies all current theories of face recognition.

Children and adults recall the names of highly familiar faces faster than semantic information.

Adults find it harder to remember the names of familiar people than other biographical information such as occupation or nationality. It has been suggested that the opposite effect occurs in children (Scanlan & Johnston, 1997). We failed to replicate the effects found by Scanlan and Johnston and instead found that children were slower to match a name than an occupation to a famous face (Experiment 1). In Experiments 2 and 3, however, we show a temporal advantage for names in both adults and children when highly familiar faces are used. This is the case for famous and personally known faces. These results show that the speed of name retrieval is influenced by familiarity in the same way in both children and adults and indicate that children do not represent knowledge for familiar people differently from adults. The implications of these results for current models of name retrieval difficulties are discussed.

Understanding covert recognition.

An implementation of Bruce and Young's (1986) functional model of face recognition is used to examine patterns of covert face recognition previously reported in a prosopagnosic patient, PH. Although PH is unable to recognize overly the faces of people known to him, he shows normal patterns of face processing when tested indirectly. A simple manipulation of one set of connections in the implemented model induces behaviour consistent with patterns of results from PH obtained in semantic priming and interference tasks. We compare this account with previous explanations of covert recognition and demonstrate that the implemented model provides the most natural and parsimonious account available. Two further patients are discussed who show deficits in person perception. The first (MS) is prosopagnosic but shows no covert recognition. The second (ME) is not prosopagnosic, but cannot access semantic information relating to familiar people. The model provides an account of recognition impairments which is sufficiently general also to be useful in describing these patients.

Lighting direction affects recognition of untextured faces in photographic positive and negative.

Face recognition in photographic positive and negative was examined in a same/different matching task in five lighting direction conditions using untextured 3-D laser-scanned faces. The lighting directions were +60, +30, 0, -30 and -60 degrees, where negative values represent bottom lighting and positive values represent top lighting. Recognition performance was better for faces in positive than in negative when lighting directions were at +60 degrees. In one experiment, the same effect was also found at +30 degrees. However, faces in negative were recognized better than positive when the direction was -60 degrees. There was no difference in recognition performance when the lighting direction was 0 and -30 degrees. These results confirm that the effect of lighting direction can be a determinant of the photographic negative effect. Positive faces, which normally appear to be top-lit, may be difficult to recognize in negative partly because of the accompanying change in apparent lighting direction to bottom-lit.

A principal component analysis of facial expressions.

Pictures of facial expressions from the Ekman and Friesen set (Ekman, P., Friesen, W. V., (1976). Pictures of facial affect. Palo Alto, California: Consulting Psychologists Press) were submitted to a principal component analysis (PCA) of their pixel intensities. The output of the PCA was submitted to a series of linear discriminant analyses which revealed three principal findings: (1) a PCA-based system can support facial expression recognition, (2) continuous two-dimensional models of emotion (e.g. Russell, J. A. (1980). A circumplex model of affect. Journal of Personality and Social Psychology, 39, 1161-1178) are reflected in the statistical structure of the Ekman and Friesen facial expressions, and (3) components for coding facial expression information are largely different to components for facial identity information. The implications for models of face processing are discussed.

Human and automatic face recognition: a comparison across image formats.

Human subjects perform poorly at matching different images of unfamiliar faces. When images are taken by different capture devices (cameras), matching is difficult for human perceivers and also for automatic systems. We test an automatic face recognition system based on principal components analysis (PCA) and compare its performance with that of human subjects tested on the same set of images. A number of variants of the PCA system are compared, using different matching metrics and different numbers of components. PCA performance critically depends on the choice of distance metric, with a Mahalanobis metric consistently outperforming a Euclidean metric. Under optimal conditions, the automatic PCA system exceeds human performance on the same images. We hypothesise that unfamiliar face recognition may be mediated by processes corresponding to rather simple functions of the inputs.

Two loci of repetition priming in the recognition of familiar faces.

Four experiments examining the repetition priming of familiar face recognition are reported. The experiments showed that the speed of deciding whether a face is familiar was facilitated by prior presentation of the face, but not by reading the written name or by producing the name in response to a definition. In contrast, reading names and producing names to definitions both primed subsequent naming of the corresponding faces (Experiments 1 and 2). Face naming was primed more by face naming than by either familiarity decisions or naming from description (Experiments 3 and 4). The authors propose that repetition priming of familiar face recognition occurs at 2 distinct loci. The first involves the perceptual recognition of a face as familiar and is domain-specific. The second involves name retrieval and is susceptible to both within- and cross-domain priming.

Matching identities of familiar and unfamiliar faces caught on CCTV images.

People can be inaccurate at matching unfamiliar faces shown in high-quality video images, even when viewpoint and facial expressions are closely matched. However, identification of highly familiar faces appears good, even when video quality is poor. Experiment 1 reported a direct comparison between familiar and unfamiliar faces. Participants who were personally familiar with target items appearing on video were highly accurate at a verification task. Unfamiliar participants doing the same task performed very inaccurately. Familiarity affected discriminability, but not bias. Experiments 2 and 3 showed that brief periods of familiarization have little beneficial effect unless "deep" or "social" processing is encouraged. The results show that video evidence can be used effectively as a probe to identity when the faces shown are highly familiar to observers, but caution should be used where images of unfamiliar people are being compared.

I recognize your face but I can't remember your name: a simple explanation?

When shown the faces of familiar people, subjects are typically slower and less accurate at retrieving names than other semantic information. This finding, along with converging evidence from neuropsychological studies, has influenced most theoretical accounts of face recognition (e.g. Bruce & Young, 1986). These accounts propose that names are stored separately from semantic information, and that they may not be retrieved in the absence of other information. Here we show that it is possible to account for empirical findings without positing a separate store for names. The account is based on an implemented simulation with an interactive activation and competition architecture. We demonstrate that the fact that most names are unique leads naturally to the patterns of recall found in experimental studies.

Understanding face recognition with an interactive activation model.

In this paper we describe how the microstructure of the Bruce & Young (1986) functional model of face recognition may be explored and extended using an interactive activation implementation. A simulation of the recognition of familiarity of individuals is developed which accounts for a range of published findings on the effects of semantic priming, repetition priming and distinctiveness. Finally, we offer some speculative predictions made by the model, and point to an empirical programme of research which it suggests.

Sequence learning by action and observation: evidence for separate mechanisms.

In the Serial Reaction Time (SRT) task, participants respond to a set of stimuli the order of which is apparently random, but which consists of repeating sub-sequences. Participants can become sensitive to this regularity, as measured by an indirect test of reaction time, but can remain apparently unaware of the sequence, as measured by direct tests of prediction or recognition. Some researchers have claimed that this learning may take place by observation alone. We suggest that observational learning may be due to explicit acquired knowledge of the sequence, and is not mediated by the same processes which give rise to learning by action. In Expt 1, we show that it is very difficult to acquire explicit sequence knowledge under dual task conditions, even when participants are told that a regular sequence exists. In Expt 2, we use the same conditions to compare actors, who respond to the sequence during learning, and observers, who merely watch the stimuli. Furthermore, we manipulate the salience of the sequence, in order to encourage learning. There is no evidence of observational learning in these conditions, despite the usual effects of learning being demonstrated by actors. In Expt 3, we show that observational learning does occur, but only when observers have no secondary task and even then only reliably for a sequence which has been made salient by chunking subcomponents. We conclude that sequence learning by observation is mediated by explicit processes, and is eliminated under conditions which support learning by action, but make it difficult to acquire explicit knowledge.

Priming the access to names of famous faces.

There is a continuing controversy in models of face identification concerning the level of access to names relative to semantic information. In order to determine whether names are accessed sequentially after or in parallel to semantic information, we studied participants' speeded decisions about famous faces that were primed by partial semantic or partial name information. Decisions that required the access to the celebrity's name (one or more forename syllables, Expt 1) were significantly primed by partial name primes (initials or name fragments). However, at variance with sequential stage models, no reliable priming was observed by partial semantic primes (information about nationality, occupation, or whether a person was dead or alive). Moreover, there was a clear and consistent priming effect by partial semantic primes if the task was a nationality (British or American) decision that required the access to semantic information (Expt 2), demonstrating the effectiveness of these primes. The effects of partial name primes on nationality decisions were less consistent, with a significant effect for name fragments but not initials. However, effects of name primes were generally greater for syllable decisions than nationality decisions, and effects of semantic primes were generally greater for nationality decisions than syllable decisions. Taken together, these results favour a model of parallel rather than sequential access and suggest some degree of independence in the access to personal semantics and names.

An artificial neural network system for diagnosis of acute myocardial infarction (AMI) in the accident and emergency department: evaluation and comparison with serum myoglobin measurements.

Recent studies have confirmed that artificial neural networks (ANNs) are adept at recognising patterns in sets of clinical data. The diagnosis of acute myocardial infarction (AMI) in patients presenting with chest pain remains one of the greatest challenges in emergency medicine. The aim of this study was to evaluate the performance of an ANN trained to analyse clinical data from chest pain patients. The ANN was compared with serum myoglobin measurements--cardiac damage is associated with increased circulating myoglobin levels, and this is widely used as an early marker for evolving AMI. We used 39 items of clinical and ECG data from the time of presentation to derive 53 binary inputs to a back propagation network. On test data (200 cases), overall accuracy, sensitivity, specificity and positive predictive value (PPV) of the ANN were 91.8, 91.2, 90.2 and 84.9% respectively. Corresponding figures using linear discriminant analysis were 81.0, 77.9, 82.6 and 69.7% (P < 0.01). Using a further test set from a different centre (91 cases), the accuracy, sensitivity, specificity and PPV for the admitting physicians were 65.1, 28.5, 76.9 and 28.6% respectively compared with 73.6, 52.4, 80.0 and 44.0% for the ANN. Although myoglobin at presentation was highly specific, it was only 38.0% sensitive, compared with 85.7% at 3 h. Simple strategies to combine clinical opinion, ANN output and myoglobin at presentation could greatly improve sensitivity and specificity of AMI diagnosis. The ideal support for emergency room physicians may come from a combination of computer-aided analysis of clinical factors and biochemical markers such as myoglobin. This study demonstrates that the two approaches could be usefully combined, the major benefit of the decision support system being in the first 3 h before biochemical markers have become abnormal.

100% accuracy in automatic face recognition.

Accurate face recognition is critical for many security applications. Current automatic face-recognition systems are defeated by natural changes in lighting and pose, which often affect face images more profoundly than changes in identity. The only system that can reliably cope with such variability is a human observer who is familiar with the faces concerned. We modeled human familiarity by using image averaging to derive stable face representations from naturally varying photographs. This simple procedure increased the accuracy of an industry standard face-recognition algorithm from 54% to 100%, bringing the robust performance of a familiar human to an automated system.