Decision-centered teamwork in Dynamic Early Exiting Networks

Dynamic early exiting is an effective method for improving the efficiency of deep neural networks. By adding classifiers into intermediate layers of deep neural networks, early exiting allows inference to conclude early for simpler samples. However, usually early exiting networks use the output of the final processed classifier, ignoring the information given by the previous classifiers. To address this, we explore decision-centered teamwork in dynamic networks. We propose a Bayesian updating method to effectively integrate information from earlier classifiers and introduce a conformal prediction-based approach to improve likelihood estimation. Additionally, we propose a logit voting method that, despite using less information, performs well in practical deep learning applications. It provides a practical alternative when accurate likelihoods for Bayesian updating are difficult to obtain. Through synthetic and real-world experiments, we demonstrate the effectiveness of these methods and the significant potential of the Bayesian updating approach.