Rapid Unsupervised Encoding of Object Files for Visual Reasoning

Visual thinking plays a central role in human cognition, yet we know little about the algorithmic operations that make it possible. Starting with outputs of a JIM-like model of shape perception, we present a model that generates object file-like representations that can be stored in memory for future recognition, and can be used by a LISA-like inference engine to reason about those objects. The model encodes structural representations of objects on the fly, stores them in long term memory, and simultaneously compares them to previously stored representations in order to identify candidate source analogs for inference. Preliminary simulation results suggest that the representations afford the flexibility necessary for visual thinking. The model provides a starting point for simulating not only object recognition, but also reasoning about the form and function of objects.