Early Exit Neural Network
Early-exit neural networks (EENNs) aim to improve the efficiency of deep learning models by allowing predictions to be made at multiple stages of inference, rather than only at the final layer. Research focuses on developing effective decision mechanisms to determine when to exit early, optimizing EENN architectures (including adaptations of ResNets and transformers), and applying them to various tasks and hardware constraints, such as embedded systems and IoT devices. This approach offers significant potential for reducing computational costs and latency in resource-limited environments, impacting applications ranging from image classification and speech recognition to real-time sensor data processing.
Papers
Conditional computation in neural networks: principles and research trends
Simone Scardapane, Alessandro Baiocchi, Alessio Devoto, Valerio Marsocci, Pasquale Minervini, Jary Pomponi
Temporal Decisions: Leveraging Temporal Correlation for Efficient Decisions in Early Exit Neural Networks
Max Sponner, Lorenzo Servadei, Bernd Waschneck, Robert Wille, Akash Kumar
Efficient Post-Training Augmentation for Adaptive Inference in Heterogeneous and Distributed IoT Environments
Max Sponner, Lorenzo Servadei, Bernd Waschneck, Robert Wille, Akash Kumar