Neural Network Verification
When
Deep learning and neural networks are being used to develop softwares to do complex computing tasks. However, lack of guarantees about their behavior limits their use in safety-critical and security critical systems. It is unrealistic to provide functional correctness of a deep learning neural network model. Researchers rather try to verify some of its desirable or undesirable properties. Verifying these properties helps to prove a model to be robust, safe and consistent. Thus, it is crucial to verify a model before using. Two main approaches are studied for verification. One approach is constraint-based where the verification problem is encoded as a set of constraints. Then the problem is solved using a Satisfiability Modulo Theories (SMT) solver or a Mixed Integer Linear Program (MILP) solver. Another approach is abstraction-based where the input is abstracted to sets of domains and verification is done within these domains. The former is complete but not scalable, the later is scalable but incomplete.
Place: Math, 402 and Zoom: https://arizona.zoom.us/j/82075792519 Password: 150721