A Deep Learning Account of Shape and Color Biases in Categorization

• In Event: 3-036 - 30 Years of Shape Bias Research: Lessons for Word Learning, Atypical Development, Categorization and Beyond

Integrative Statement

Do children’s biases to extend concepts according to shape arise from accumulated experience, a maturational change, or a hardwired tendency? Modern neural networks, namely deep convolutional neural networks (DCNNs), provide a testbed to evaluate these theoretical possibilities. Reflecting the accumulation of experience account, these models can learn to name photographs of real-world objects at adult levels of proficiency after being trained extensively. Reflecting the hardwired tendency account, these models’ basic structure is inspired by the ventral stream in the brain for processing visual information with later layers in the model relating to later regions in the brain (e.g., Guest & Love, 2017).

Here, we specify an exemplar model of categorization that uses representations from a DCNN layer (see Figure 1). By comparing models that use representations from different layers, we can determine what information resides at each layer in the DCNN, as well as characterize the visual representations used by human and non-human animals. In modelling the shape bias, the human tendency to prefer matches based on shape over other features like color or texture, we find, using the triplet task from Ritter et al. (2017), that human performance is captured by representations at advanced DCNN layers. Indeed, color bias dominates in models using early layers (see Figure 2). In a follow-up, we systematically investigated at which level of representation information about size, hue, luminance, and shape resides in the DCNN. Finally, we used this modelling approach to determine the basis of human and pigeon performance in a complex categorization task, classifying medical cardiac images as healthy or unhealthy.

These results suggest that aspects of the shape bias in children fall out of natural image statistics, which are reflected in the training set of the DCNN. However, the developmental literature points to a more complex pattern of results in which vocabulary, recent training, and task setting (see Perry et al., 2014 for an example) all conspire to influence shape and material biases. Moreover, the initial structure of the model, patterned after the human visual system, is critical in allowing it to acquire a shape bias, which points to an interplay between the learning mechanism and experience. One challenge is extending DCNN accounts beyond basic image statistics to capture this full range of developmental phenomena.

Authors

• Bradley C Love, University College London
  Presenting Author

• Olivia Guest, University College London
  Non-Presenting Author