The idea of model compression is quite exciting to me. It could bring difficult classification tasks such as scene analysis and large scale classification of visual or auditory concepts to devices with limited computation and memory resources. Think smart phones, but robots as well.