Using Image Saliency to Reduce Runtime of Image Classification in Deep Neural Networks
Object recognition runtime is a major limiting factor in computer vision because the tasks are so complex particularly in real-world situations. This project aims to improve object recognition runtime and accuracy by inputting salient regions of images° the most informative parts of the images as determined by humans or algorithms that approximate human fixations° along with the full images. We will use trained CNNs and enhance them with functions that determine and feed in salient regions of input images and compare our findings with controls.

I am studying both computer science and neuroscience and I’ve always found a range of topics in math and natural science interesting. I have had experience applying machine learning to behavioral data and now hope to learn more about how my two fields of study interface. This project excites me because it is computer vision a more involved application of my interests.