Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster A in Poster Session A: Tuesday, August 4, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Biasing optimization to find more informative model metamers

William F. Broderick¹, Daniel Herrera-Esposito², Erica Shook³, Eero P Simoncelli¹; ¹Flatiron Institute, ²University of Pennsylvania, ³Columbia University

Presenter: William F. Broderick

Sensory neuroscientists must carefully choose or design experimental stimuli in order to test hypotheses. Recently, scientists have leveraged computational models and automatic differentiation to synthesize model-optimized stimuli, such as model metamers -- stimuli that are physically distinct but that produce identical model responses. However, there exist many different stimuli which satisfy the constraints of the stimulus-generation process, but which may give rise to different scientific interpretations. Here, we propose adding a penalty term to the objective function used to generate the stimuli. This biases the synthesis procedure, allowing researchers to preferentially search for stimuli with certain properties. We demonstrate the use of several penalty functions on a simple LGN-inspired model to increase perceptual diversity among synthesized model metamers. By carefully choosing their penalty functions, researchers can better design stimulus sets to address their scientific question.

Topic Area: Methods, Tools, Theory & Neural Coding