Convolutional neural net face recognition works in non-human-like ways

Hancock, Peter; Somai, Rosyl S.; Mileva, Viktoria R

doi:10.1098/rsos.200595

Cited by 21 publications

(23 citation statements)

References 23 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Until the Short-Form of the OFMT is revised, the short-form of the GFMT may be preferable for use when testing time is restricted. Finally, Study 1 established that algorithmic similarity judgements were in accordance with judgements provided by neurotypical participants, and that correlations were of greater magnitude for face images taken from the same person compared to face images of different individuals, replicating previous findings (Hancock, Somai, & Mileva, 2020). The size of these correlations was moderate, supporting the contention that the algorithms are not simply mimicking the strategy used by neurotypical participants when judging similarity.…”

Section: Discussionsupporting

confidence: 72%

The Oxford Face Matching Test: A Non-Biased Test of The Full Range of Individual Differences in Face Perception

Stantić¹,

Brewer²,

Duchaine³

et al. 2021

Preprint

View full text Add to dashboard Cite

Tests of face processing are typically designed to identify individuals performing outside of the typical range; either prosopagnosic individuals who exhibit poor face processing ability, or super recognisers, who have superior face processing abilities. Here we describe the development of the Oxford Face Matching Test (OFMT), designed to identify individual differences in face processing across the full range of performance, from prosopagnosia, through the range of typical performance, to super recognisers. Such a test requires items of varying difficulty, but establishing difficulty is problematic when particular populations (e.g. prosopagnosics, individuals with Autism Spectrum Disorder) may use atypical strategies to process faces. If item difficulty is calibrated on neurotypical individuals, then the test may be poorly calibrated for atypical groups, and vice versa. To obtain items of varying difficulty we used facial recognition algorithms to obtain face pair similarity ratings that are not biased towards specific populations. These face pairs were used as stimuli in the OFMT, and participants required to judge if the face images depicted the same individual or different individuals. Across five studies the OFMT was shown to be sensitive to individual differences in the typical population, and in groups of both prosopagnosic individuals and super recognisers. The test-retest reliability of the task was at least equivalent to the Cambridge Face Memory Test and the Glasgow Face Matching Test. Furthermore, results reveal, at least at the group level, that both face perception and face memory are poor in those with prosopagnosia, and good in super recognisers.

show abstract

Section: Discussionsupporting

confidence: 72%

The Oxford Face Matching Test: A Non-Biased Test of The Full Range of Individual Differences in Face Perception

Stantić¹,

Brewer²,

Duchaine³

et al. 2021

Preprint

View full text Add to dashboard Cite

show abstract

“…However, the extent of computational and representational similarity between DNNs and humans remains unclear, and only a small subset of available DNNs have been compared to humans (e.g., see refs. 26,27,32,33 ).…”

Section: Introductionmentioning

confidence: 99%

Diverse routes to expertise in facial recognition

Towler¹,

Dunn²,

Martínez³

et al. 2021

Preprint

View full text Add to dashboard Cite

Facial recognition errors jeopardize national security, criminal justice, public safety and civil rights. Here, we compare the most accurate humans and facial recognition technology in a detailed lab-based evaluation and international proficiency test for forensic scientists involving 27 forensic departments from 14 countries. We find striking cognitive and perceptual diversity between naturally skilled super-recognizers, trained forensic examiners and deep neural networks, despite them achieving equivalent accuracy. Clear differences emerged in super-recognizers’ and forensic examiners’ perceptual processing, errors, and response patterns: super-recognizers were fast, biased to respond ‘same person’ and misidentified people with extreme confidence, whereas forensic examiners were slow, unbiased and strategically avoided misidentification errors. Further, these human experts and algorithms disagreed on the similarity of faces, pointing to differences in their face representations. Our findings reveal there are multiple types of facial recognition expertise, some of which are better suited to particular real-world facial recognition roles than others.

show abstract

“…Research (2020) 5:59 matches (it is possible that the opposite pattern of results will occur for mismatches). Yet, after observing the performance of this same DNN in unrelated studies (Hancock et al 2020), we also predicted that classification accuracy would remain near ceiling in all mask conditions. As such, the DNN's accuracy should be similar to human observers for familiar faces, and superior for unfamiliar faces.…”

Section: Introductionmentioning

confidence: 85%

Surgical face masks impair human face matching performance for familiar and unfamiliar faces

Carragher¹,

Hancock²

2020

Preprint

View full text Add to dashboard Cite

In response to the COVID-19 pandemic, many governments around the world now recommend, or require, their citizens to cover the lower half of their face in public. Consequently, many people now wear surgical face masks in public. We investigated whether surgical face masks affected the performance of human observers, and a state-of-the-art face recognition system, on tasks of perceptual face matching. Participants judged whether two simultaneously presented face photographs showed the same person or two different people. We superimposed images of surgical masks over the faces, creating three different mask conditions: control (no masks), mixed (one face wearing a mask), and masked (both faces wearing masks). We find that surgical face masks have a large detrimental effect on human face matching performance, and that the degree of impairment is the same regardless of whether one or both faces in each pair are masked. Surprisingly, this impairment is similar in size for both familiar and unfamiliar faces. When matching masked faces, human observers are biased to reject unfamiliar faces as mismatches and to accept familiar faces as matches. Finally, the face recognition system showed very high classification accuracy for control and masked stimuli, even though it had not been trained to recognise masked faces. However, accuracy fell markedly when one face was masked and the other was not. Our findings demonstrate that surgical face masks impair the ability of humans, and naïve face recognition systems, to perform perceptual face matching tasks. Matching decisions for masked faces should be treated with caution.

show abstract

Convolutional neural net face recognition works in non-human-like ways

Cited by 21 publications

References 23 publications

The Oxford Face Matching Test: A Non-Biased Test of The Full Range of Individual Differences in Face Perception

The Oxford Face Matching Test: A Non-Biased Test of The Full Range of Individual Differences in Face Perception

Diverse routes to expertise in facial recognition

Surgical face masks impair human face matching performance for familiar and unfamiliar faces

Contact Info

Product

Resources

About