Learning Sensor Sample-Reweighting for Dynamic Early-Exit Activity Recognition Via Meta Learning.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Dynamic early-exit strategies improve deep neural network efficiency by allowing simpler samples to exit at earlier layers.
• Existing methods often ignore the dynamic nature of early exits during training, creating a mismatch with test-time behavior.
• Human Activity Recognition (HAR) on wearable devices requires efficient models due to computational constraints.

Purpose of the Study:

• To develop a novel sample-reweighting approach for efficient activity inference in HAR.
• To address the training-test mismatch in dynamic early-exit strategies.
• To enhance the accuracy-efficiency trade-off for HAR on resource-constrained wearable devices.

Main Methods:

• Introduced a sample-reweighting approach using a weight-predicting network to dynamically adjust training loss contributions for each sample at different exits.
• Designed a meta-learning-based optimization objective to jointly train the weight-predicting network and the backbone network.
• Evaluated the method on UCI-HAR, WISDM, and UniMiB-SHAR datasets.

Main Results:

• The proposed method consistently improved accuracy-efficiency trade-offs under budgeted batch classification and anytime prediction scenarios.
• Demonstrated improved performance by incorporating test-time early-exit behavior into the training pipeline.
• Showcased a natural advantage in handling class-imbalanced HAR problems.

Conclusions:

• The sample-reweighting approach effectively bridges the training-test gap in dynamic early-exit HAR.
• This method offers a significant improvement in computational efficiency and accuracy for wearable HAR systems.
• The approach is robust and adaptable, particularly for imbalanced datasets and real-world deployment scenarios.