Match me if you can: Evidence for a domain-general visual comparison ability

Measuring object recognition ability: Reliability, validity, and the aggregate z-score approach

Measurement of domain-general object recognition ability (o) requires minimization of domain-specific variance. One approach is to model o as a latent variable explaining performance on a battery of tests which differ in task demands and stimuli; however, time and sample requirements may be prohibitive. Alternatively, an aggregate measure of o can be obtained by averaging z-scores across tests. Using data from Sunday et al., Journal of Experimental Psychology: General, 151, 676-694, (2022), we demonstrated that aggregate scores from just two such object recognition tests provide a good approximation (r = .79) of factor scores calculated from a model using a much larger set of tests. Some test combinations produced correlations of up to r = .87 with factor scores. We then revised these tests to reduce testing time, and developed an odd one out task, using a unique object category on nearly every trial, to increase task and stimuli diversity. To validate our measures, 163 participants completed the object recognition tests on two occasions, one month apart. Providing the first evidence that o is stable over time, our short aggregate o measure demonstrated good test-retest reliability (r = .77). The stability of o could not be completely accounted for by intelligence, perceptual speed, and early visual ability. Structural equation modeling suggested that our tests load significantly onto the same latent variable, and revealed that as a latent variable, o is highly stable (r = .93). Aggregation is an efficient method for estimating o, allowing investigation of individual differences in object recognition ability to be more accessible in future studies.

Manipulating and measuring variation in deep neural network (DNN) representations of objects

We explore how DNNs can be used to develop a computational understanding of individual differences in high-level visual cognition given their ability to generate rich meaningful object representations informed by their architecture, experience, and training protocols. As a first step to quantifying individual differences in DNN representations, we systematically explored the robustness of a variety of representational similarity measures: Representational Similarity Analysis (RSA), Centered Kernel Alignment (CKA), and Projection-Weighted Canonical Correlation Analysis (PWCCA), with an eye to how these measures are used in cognitive science, cognitive neuroscience, and vision science. To manipulate object representations, we next created a large set of models varying in random initial weights and random training image order, training image frequencies, training category frequencies, and model size and architecture and measured the representational variation caused by each manipulation. We examined both small (All-CNN-C) and commonly-used large (VGG and ResNet) DNN architectures. To provide a comparison for the magnitude of representational differences, we established a baseline based on the representational variation caused by image-augmentation techniques used to train those DNNs. We found that variation in model randomization and model size never exceeded baseline. By contrast, differences in training image frequency and training category frequencies caused representational variation that exceeded baseline, with training category frequency manipulations exceeding baseline earlier in the networks. These findings provide insights into the magnitude of representational variations that can be expected with a range of manipulations and provide a springboard for further exploration of systematic model variations aimed at modeling individual differences in high-level visual cognition.

The novel object-matching test (NOM Test): A psychometric measure of visual comparison ability

This paper presents a new test of object-matching ability: the Novel Object-Matching Test (NOM Test). Object-matching (or visual comparison) is a complex cognitive and perceptual visual comparison task undertaken by forensic scientists - yet no openly available, standardised and psychometrically validated test of object-matching ability exists. This is in contrast to other visual comparison domains like face-matching where many tests are widely available. In this paper, we present the development and psychometric validation of the first openly available object-matching test where people view two complex artificial visual patterns side-by-side and decide if they are from the same source or different sources. We provide normative data and psychometric properties for two long-form and two short-form versions of the test, and two additional versions designed to identify high and low-performers. We also provide evidence of discriminant validity and convergent validity that demonstrates the NOM Test correlates strongly with other object-matching tasks like fingerprint-matching - but not other tasks requiring cognitive-perceptual skill (e.g., visual intelligence). The NOM Test is free for research use with acknowledgment and is available at https://osf.io/pv6ye/ .

Evidence for an amodal domain-general object recognition ability

A general object recognition ability predicts performance across a variety of high-level visual tests, categories, and performance in haptic recognition. Does this ability extend to auditory recognition? Vision and haptics tap into similar representations of shape and texture. In contrast, features of auditory perception like pitch, timbre, or loudness do not readily translate into shape percepts related to edges, surfaces, or spatial arrangement of parts. We find that an auditory object recognition ability correlates highly with a visual object recognition ability after controlling for general intelligence, perceptual speed, low-level visual ability, and memory ability. Auditory object recognition was a stronger predictor of visual object recognition than all control measures across two experiments, even though those control variables were also tested visually. These results point towards a single high-level ability used in both vision and audition. Much work highlights how the integration of visual and auditory information is important in specific domains (e.g., speech, music), with evidence for some overlap of visual and auditory neural representations. Our results are the first to reveal a domain-general ability, o, that predicts object recognition performance in both visual and auditory tests. Because o is domain-general, it reveals mechanisms that apply across a wide range of situations, independent of experience and knowledge. As o is distinct from general intelligence, it is well positioned to potentially add predictive validity when explaining individual differences in a variety of tasks, above and beyond measures of common cognitive abilities like general intelligence and working memory.

Diverse types of expertise in facial recognition

Facial recognition errors can 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 deep neural networks disagreed on the similarity of faces, pointing to differences in their representations of faces. Our findings therefore reveal multiple types of facial recognition expertise, with each type lending itself to particular facial recognition roles in operational settings. Finally, we show that harnessing the diversity between individual experts provides a robust method of maximizing facial recognition accuracy. This can be achieved either via collaboration between experts in forensic laboratories, or most promisingly, by statistical fusion of match scores provided by different types of expert.

A method for quantifying individual decision thresholds of latent print examiners

In recent years, 'black box' studies in forensic science have emerged as the preferred way to provide information about the overall validity of forensic disciplines in practice. These studies provide aggregated error rates over many examiners and comparisons, but errors are not equally likely on all comparisons. Furthermore, inconclusive responses are common and vary across examiners and comparisons, but do not fit neatly into the error rate framework. This work introduces Item Response Theory (IRT) and variants for the forensic setting to account for these two issues. In the IRT framework, participant proficiency and item difficulty are estimated directly from the responses, which accounts for the different subsets of items that participants often answer. By incorporating a decision-tree framework into the model, inconclusive responses are treated as a distinct cognitive process, which allows inter-examiner differences to be estimated directly. The IRT-based model achieves superior predictive performance over standard logistic regression techniques, produces item effects that are consistent with common sense and prior work, and demonstrates that most of the variability among fingerprint examiner decisions occurs at the latent print evaluation stage and as a result of differing tendencies to make inconclusive decisions.

The neural correlates of domain-general visual ability

People vary in their general ability to compare, identify, and remember objects. Research using latent variable modeling identifies a domain-general visual recognition ability (called o) that reflects correlations among different visual tasks and categories. We measure associations between a psychometrically-sensitive measure of o and a neurometrically-sensitive measure of visual sensitivity to shape. We report evidence for distributed neural correlates of o using functional and anatomical regions-of-interest (ROIs) as well as whole brain analyses. Neural selectivity to shape is associated with o in several regions of the ventral pathway, as well as additional foci in parietal and premotor cortex. Multivariate analyses suggest the distributed effects in ventral cortex reflect a common mechanism. The network of brain areas where neural selectivity predicts o is similar to that evoked by the most informative features for object recognition in prior work, showing convergence of 2 different approaches on identifying areas that support the best object recognition performance. Because o predicts performance across many visual tasks for both novel and familiar objects, we propose that o could predict the magnitude of neural changes in task-relevant areas following experience with specific task and object category.

Object recognition ability predicts category learning with medical images

We investigated the relationship between category learning and domain-general object recognition ability (o). We assessed this relationship in a radiological context, using a category learning test in which participants judged whether white blood cells were cancerous. In study 1, Bayesian evidence negated a relationship between o and category learning. This lack of correlation occurred despite high reliability in all measurements. However, participants only received feedback on the first 10 of 60 trials. In study 2, we assigned participants to one of two conditions: feedback on only the first 10 trials, or on all 60 trials of the category learning test. We found strong Bayesian evidence for a correlation between o and categorisation accuracy in the full-feedback condition, but not when feedback was limited to early trials. Moderate Bayesian evidence supported a difference between these correlations. Without feedback, participants may stick to simple rules they formulate at the start of category learning, when trials are easier. Feedback may encourage participants to abandon less effective rules and switch to exemplar learning. This work provides the first evidence relating o to a specific learning mechanism, suggesting this ability is more dependent upon exemplar learning mechanisms than rule abstraction. Object-recognition ability could complement other sources of individual differences when predicting accuracy of medical image interpretation.

Selecting police super-recognisers

People vary in their ability to recognise faces. These individual differences are consistent over time, heritable and associated with brain anatomy. This implies that face identity processing can be improved in applied settings by selecting high performers-'super-recognisers' (SRs)-but these selection processes are rarely available for scientific scrutiny. Here we report an 'end-to-end' selection process used to establish an SR 'unit' in a large police force. Australian police officers (n = 1600) completed 3 standardised face identification tests and we recruited 38 SRs from this cohort to complete 10 follow-up tests. As a group, SRs were 20% better than controls in lab-based tests of face memory and matching, and equalled or surpassed accuracy of forensic specialists that currently perform face identification tasks for police. Individually, SR accuracy was variable but this problem was mitigated by adopting strict selection criteria. SRs' superior abilities transferred only partially to body identity decisions where the face was not visible, and they were no better than controls at deciding which visual scene that faces had initially been encountered in. Notwithstanding these important qualifications, we conclude that super-recognisers are an effective solution to improving face identity processing in applied settings.

The low prevalence effect in fingerprint comparison amongst forensic science trainees and novices

The low prevalence effect is a phenomenon whereby target prevalence affects performance in visual search (e.g., baggage screening) and comparison (e.g., fingerprint examination) tasks, such that people more often fail to detect infrequent target stimuli. For example, when exposed to higher base-rates of 'matching' (i.e., from the same person) than 'non-matching' (i.e., from different people) fingerprint pairs, people more often misjudge 'non-matching' pairs as 'matches'-an error that can falsely implicate an innocent person for a crime they did not commit. In this paper, we investigated whether forensic science training may mitigate the low prevalence effect in fingerprint comparison. Forensic science trainees (n = 111) and untrained novices (n = 114) judged 100 fingerprint pairs as 'matches' or 'non-matches' where the matching pair occurrence was either high (90%) or equal (50%). Some participants were also asked to use a novel feature-comparison strategy as a potential attenuation technique for the low prevalence effect. Regardless of strategy, both trainees and novices were susceptible to the effect, such that they more often misjudged non-matching pairs as matches when non-matches were rare. These results support the robust nature of the low prevalence effect in visual comparison and have important applied implications for forensic decision-making in the criminal justice system.