Reviews: Shape and Material from Sound

This paper contains a lot of ideas and methods packed into a tightly wound package - less would have been certainly more: The general aim is to build a system that can mimick the human ability to recognise shape and material objects from their sounds. Certainly a nice idea to explore within a NIPS community. To this end a generative model is defined, and a process for inference is proposed ("Physics-Based Audio Engine") that uses a physics simulation, this physics simulation is coupled (somehow) to a sound generation algorithm. This appears to use (somehow) the physics simulation engine representation of the vibrating surfaces of the objects to render the sounds. This engine is then used to create training stimuli for 4 variants of machine learning algorithms to learn appropriate representations.