Knowledge Distillation for Reservoir-based Classifier: Human Activity Recognition

The paper introduces PatchEchoClassifier, an energy-efficient model for human activity recognition using time-series sensor data, leveraging Echo State Networks and a tokenizer for patch-level features. It employs knowledge distillation from a high-capacity teacher, achieving over 80% accuracy with significantly lower computational costsabout one-sixth of DeepConvLSTMmaking it suitable for real-time edge applications.