The success of machine learning is a generalized algorithm which picks up patterns in the data. Instead of hand crafting programs to make the software intelligent, the computer is able to learn this on its own. The resulting algorithms are only as good as the data (garbage in, garbage out). The training dataset needs to adequately represent the inputs the model will see out in the wild. DataComp is a data-centric competition which flips the traditional way of incrementally improving AI algorithms on its head: instead of fixing the training dataset and iterating on the model architecture, participants are given a fixed model and find ways to improve the training dataset. We scraped vast portions of the internet to create a large pool of images and captions that participants can try their filtering methods on, as well as bringing their own data. This was a large scale project with lots of collaborators from UW, Apple, Google, AI2, Stability.AI and more. Check out the website, code, and paper to learn more. 

One approach to making computers "see" like humans is to leverage our interdisciplinary understanding of the human vision system (psychology, biology, and optics) to design systems that mimic properties that we observe in humans. For example, I have trained a neural network to classify textures as a human would. Additionally, we have created a system that responds to optical illusions by "filling in gaps" similar to way a human does. Check out the website, paper (BMVC 2021), and presentation below.