DNN Repair

DNN repair focuses on correcting errors and improving the performance of deep neural networks, addressing issues like vulnerability to adversarial attacks, inaccurate predictions, and performance degradation after quantization. Current research explores various techniques, including retraining, fine-tuning, and direct weight modification at the neuron or block level, often employing optimization algorithms and formal verification methods to ensure provable repairs. This field is crucial for enhancing the reliability and safety of DNNs in critical applications, such as autonomous driving and medical diagnosis, by improving their robustness and accuracy while minimizing unintended consequences.