Bennett King-Nyberg

Presenter: Bennett King-Nyberg

Title: Automating Justice: A Computational Method for Creating Photo Lineups

Abstract:

The perceptual similarity of individuals within an eyewitness lineup plays a key role in identification accuracy. When lineup members are highly similar to one another, witnesses may struggle to distinguish the perpetrator from innocent fillers, reducing discriminability and increasing the risk of misidentification. Conversely, when similarity is too low, the suspect may stand out, making the task artificially easy and compromising lineup fairness. Maintaining an appropriate level of similarity among lineup members is therefore critical to ensuring both accuracy and fairness in eyewitness identification.
Traditionally, lineup similarity has been assessed subjectively by lineup constructors, relying on intuition or informal judgment rather than standardized metrics. This introduces variability and potential bias, underscoring the need for more objective, data-driven methods of lineup construction.

To this end, we employed PsiZ, a state-of-the-art machine learning framework that transforms behavioural similarity judgments into psychologically meaningful multidimensional embeddings. These embeddings provide a nuanced representation of perceived similarity between faces, making them well suited for examining how lineup composition influences identification accuracy. Participants completed a similarity judgment task using faces drawn from existing eyewitness lineup datasets. We then used PsiZ to model these judgments, yielding continuous similarity metrics for each pair of stimuli. Our central aim was to evaluate whether these PsiZ-derived similarity values could predict lineup identification accuracy, offering a more precise and interpretable account of how perceptual similarity shapes eyewitness performance.

Finally, we developed a program that uses pairwise similarity values to algorithmically construct lineups with controlled levels of perceptual similarity between members.