The Project
The FIDO Lab at Georgia Tech researches novel methods of dog-computer interaction. Dogs in this lab have been trained to interact with vests and screens embedded with sensors to respond to emergencies, such as a diabetic seizure. A service dog could pull on a rope on their vest when their human fell, calling 911 and getting them critical aid. In Spring 2018, I pitched a project to study the extent to which Gestalt principles apply to dogs, with the goal of improving our understanding of not only how dogs perceive the visual world, but which aspects of the world are more important to them as well. Over the course of the semester, I worked with another student to research similar perception studies, design and code a touchscreen task for a dog to complete, and work with a trainer and dog to test our hypothesis.
Our first goal of the semester was to conduct a literature analysis. We needed to understand the background of the Gestalt principles in humans, any research into dog perception beyond their color spectrum, and understand study design for animals (which is very different from humans!). We completed study training administered by the Georgia Tech Institutional Animal Care and Use Committee (IACUC). Gestalt principles refer to a set of principles developed in the 1920s to explain oddities and tricks in the human vision system (read more here). In our research, we found the area of Gestalt principles in animals was fairly new; this meant our research would lay the groundwork for future studies and more advanced understanding. One of the most important things we learned about dog study design is dogs are extremely sensitive to subconscious cues from their handlers, known as the "Clever Hans" phenomenon.
For our study, we chose to isolate the Gestalt principle of similarity because it translated the best to the dog vision system and would give us the clearest results when tested. We set up two targets on a touchscreen that were the same size and shape and varied their position. If our study dog favored color, then he would be able to discriminate when the circles changed position, and touch them in the correct order. If he favors shape, it will be more difficult for him to associate the command with the color. See the video below for the study in action.
The Results
While we did not have time to repeat the study with multiple dogs, the results for this dog, Sky, in particular were fascinating. Sky seemed to naturally cue on position: he would always check where the blue dot was previously when we moved it around the screen. When three circles were arranged in a line he would perform a swiping motion, despite being trained for distinct taps, indicating that he perhaps perceived the circles as a pattern (also known as the common fate Gestalt principle).
Again, these results may be specific to Sky, and more testing would be required to understand how to generalize dogs in general. If position were truly more important, keeping a particular action tied to a particular location would help the dog more easily train on that action, and of course, more reliably perform it.
The full paper was submitted to the Animal Computer Interface conference and can be found at this link.
