Deep Networks as Models of Human and Animal Categorization

Introduction Convolutional neural networks (CNNs) trained as classifiers learn by associating visual inputs (e.g., photographs of objects) with appropriate output labels (e.g., “crow”, “dog”, “car”). These complex models, which contain millions of weights, are the state-of-the art in machine vision, rivaling humans in object recognition tasks (LeCun, Bengio, & Hinton, 2015; Krizhevsky, Sutskever, & Hinton, 2012). What these networks learn displays some commonalities with human learning (Kubilius, Bracci, & de Beeck, 2016; Lake, Zaremba, Fergus, & Gureckis, 2015). Furthermore, the layers in these networks have been related to neural activity along the ventral stream (Khaligh-Razavi & Kriegeskorte, 2014; Yamins & DiCarlo, 2016) The similarity spaces created by these models at various network layers allow us to draw parallels with the brain’s neural coding schemes (Guest & Love, 2017). At earlier layers, networks display similarity spaces that reflect the high-level categories found in the input space, e.g., lions and tigers are more similar to one another than to mopeds. At the more advanced layers, similarity structure tends to break down such that representations of different object categories become orthogonal. Can these networks also shed light on how non-human animals categorize? CNNs can be used to determine at what level of representation (i.e., what network layer) animals are coding similarities between images. For example, are animals learning regularities at a very low level, close to the pixels in the image, or are they seizing upon more abstract shape features? In this contribution, we address this question by examining data from pigeons trained to categorize images of cardiograms as normal or abnormal. Pigeons are excellent at classifying visual stimuli (Bhatt, Wasserman, Reynolds, & Knauss, 1988). For example, pigeons trained to discriminate between medical images of nora) b)