The paper proposes using techniques from control theory to update Deep Neural Networks (DNN) parameters online. It formulates the fully-connected feedforward DNN as a continuous-time dynamical system. The authors present novel last-layer update laws that guarantee error convergence under various conditions. The paper shows that training the DNN under spectral normalization controls the upper bound of the error trajectories of the online DNN predictions. The proposed online DNN adaptation laws are validated in simulation to learn the dynamics of the Van der Pol system under domain shift.
Publication date: 1 Feb 2024
Project Page: https://arxiv.org/abs/2402.00761
Paper: https://arxiv.org/pdf/2402.00761
