Layer-Stabilizing Deep Learning

摘要

Even though stability is a crucial property of dynamical systems and has been recognized as advantageous for learning, it has been often ignored in machine learning tasks. This paper presents a deep neural network learning approach that enforces layer stability during the learning process. Instead of solving the corresponding constrained optimization problem, the stability constraints are approximated based on a structured layer weight modification, and incorporated into the cost. Building upon existing learning approaches, an algorithm is provided that guarantees layer stability. The proposed method yields an improved learning progress and overall system performance compared to baseline approaches, as shown in a reinforcement-learning-based cart-pole stabilization, and a supervised-learning-based system for predicting the steering angle of an automated driving vehicle.