Vabljeni na na gostujoče predavanje prof. dr. Jie Yang, ki prihaja s Shanghai Jiao Tong University. Predavala bo na temo "Researchers on the defenses and out-of-distribution detection in trustworthy deep learning". 

Predavanje bo v petek, 24. julija 2026, ob 9:45 v predavalnici P19 na FRI.

Povzetek:

The rapid advancement of deep learning has had a transformative effect on the development of technology and society across a multitude of sectors. In safety-critical contexts, the potential for neural network models to produce unreliable outputs in response to “malicious” or “unanticipated” inputs poses a severe risk. This talk delves into the output reliability from neural network models within the domain of trustworthy deep learning. 1) Inputs that involve pixel perturbations, exemplified by adversarial examples,w.r.t the task of adversarial and certified robustness;  2) Inputs that represent distribution shifts, exemplified by Out-of-Distribution (OoD) data, w.r.t the task of out-of-distribution detection. We introduce a novel strategy of model augmentation, adopt a multi-head neural network structure, and pose diversity constraints related to adversarial robustness into the model parameters.We adopt a multi-head neural network structure, use the ensemble of multiple heads in place of the ensemble of multiple neural networks,which significantly reduces the computational load in both training and certification phases. We propose that the non-linearity in InD and OoD data hinders PCA from learning a subspace that fully embodies their diversities. We propose a mode ensemble method that not only enhances detection performance but also significantly reduces the performance variance among independent modes.We propose performing linear dimension reduction on the gradient using a designated subspace that comprises principal components.