Carlo Tomasi, Iris Einheuser Distinguished Professor of computer science at Duke University, has won an Amazon Research Award, and will receive gift funding and Amazon Web Services (AWS) promotional credits to support his project on “Deep neural network classifiers with margins in input space.” Tomasi said, “I am grateful to Amazon for their generous support. Their forward-looking bet on our new idea provides a crucial seed towards sustaining my students’ research on this topic.”
Deep learning predictors are used in a variety of applications involving images, from authentication to car-driving assistance and medical diagnostics. These predictors are very flexible and often accurate, but they are also brittle: A malicious actor can modify nearly any input image so slightly that the change is invisible to the human eye and yet it can fool the predictor into reporting a wrong result with potentially catastrophic results. For instance, a pedestrian can be classified as a clear tract of road or the wrong person can gain access to a phone or another resource. Deep learning predictors can be similarly sensitive to non-malicious perturbations.
The Amazon award allows Tomasi to further advance his research to thwart image perturbations, adversarial or benign, by designing predictors with wider decision margins. “Our unique angle is to induce these wider decision margins through new parameterizations of the predictor, rather than by using special training techniques,” said Tomasi. “In this way, resilience to image perturbations can be achieved by merely swapping the standard implementation of certain predictor components with ours, while still using familiar and well-studied training algorithms.”
Amazon Research Awards provide unrestricted funds and AWS promotional credits to academic researchers investigating various research topics in multiple disciplines. Tomasi is among 79 2022 fall award recipients who represent 54 universities in 14 countries.
Recipients have access to more than 300 Amazon public datasets and can use AWS AI/ML services and tools through their AWS Promotional Credits. Recipients also are assigned an Amazon research contact who offers consultation and advice, along with opportunities to participate in Amazon events and training sessions. Amazon also encourages the publication of research results, presentations of research at Amazon offices worldwide, and the release of related code under open-source licenses.
Tomasi's research is at the intersection of computer vision, machine learning, and applied mathematics. Tomasi's current projects include image motion analysis (funded by NSF), satellite image interpretation (funded by IARPA), computer-assisted diagnosis, and object recognition. He is an ACM Fellow and has won the IEEE Computer Society Helmholtz Prize twice.