Abstract

Many psychological experiments would benefit from a method to measure the similarity between photographs. Contrary to low-level features (e.g. size or hue), which we can measure directly, photographs are too complex. Convolutional neural networks provide not only a method for image categorization, but we can use the resulting image features to evaluate image similarity. In the talk, we will describe one potential method to measure visual similarity, which is relatively easy to use and does not require training a new neural network. We will show how scientists can use the deep feature information about similarity in memory or change-detection experiments. In particular, we demonstrate how the proximity to other stimuli in deep feature space affects memory performance and how manipulating distractors based on this measure can affect false alarm errors. We will also discuss this approach's limitations and the extent to which the measure is sufficiently sensitive to predict performance in change-detection experiments.