Auxiliary-Model Training for Resource-Limited Acoustic Scene Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing model-ensemble techniques for improving artificial neural network performance increase model complexity, which is not feasible in resource-constrained environments like mobile devices.
Innovation Solution
A method involving an auxiliary model to train an acoustic scene classification model by updating weights based on a reflection ratio determined by a weight coefficient, using an exponential moving average, and calculating a combined cost function to enhance performance without increasing model complexity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If model-ensemble technique is used to improve acoustic scene classification performance, then classification accuracy is improved, but model complexity increases
Solution Approach 1:
An auxiliary model is introduced as an intermediary component that generates predictions to guide the training of the main acoustic scene classification model. The auxiliary model does not need to be deployed at inference time, allowing the ensemble effect to be achieved during training without permanently increasing model complexity. This mediator enables knowledge transfer from multiple model perspectives while maintaining a single deployable model structure.
Solution Approach 2:
The auxiliary model performs preliminary predictions during the training phase, generating target predictions that are used to update the main model's weights. This preliminary action allows the system to benefit from ensemble-like training signals before the actual inference occurs, improving classification performance without requiring multiple models to be maintained during deployment.
2Measurement precision
If multiple models are combined in model-ensemble technique, then inference accuracy is improved, but computational resources required increase
Solution Approach 1:
The inference capability of the auxiliary model is extracted and discarded after training completion. Only the main acoustic scene classification model is retained for actual inference operations. This extraction approach allows the system to benefit from ensemble training signals while eliminating the computational overhead of maintaining and running multiple models during deployment, thus reducing energy consumption and computational resource requirements.
3Reliability
If model-ensemble technique is applied in mobile environment, then classification performance is improved, but device resource constraints are violated
Solution Approach 1:
The training process is segmented into distinct phases: the auxiliary model operates during training to provide ensemble signals, while only the main model is deployed for inference. This segmentation allows mobile devices to benefit from ensemble training methodology during the offline training phase without exceeding resource constraints during the online inference phase, making the solution adaptable to mobile environment limitations.
Data Source
AI summary
Disclosed are methods of training an acoustic scene classification model and classifying an acoustic scene and an electronic device for performing the methods. The training method of an acoustic scene classification model includes inputting training data labeled as an acoustic scene to the acoustic scene classification model that is repeatedly trained by using the training data and outputting a first result predicting the acoustic scene, updating the weight of the auxiliary model configured to induce training of the acoustic scene classification model, based on a weight of the acoustic scene classification model and a weight of an auxiliary model in a previous epoch, inputting the training data to the auxiliary model and outputting a second result, calculating a cost function, based on the first result, the second result, and labeling of acoustic data, and updating the weight of the acoustic scene classification model, based on the cost function.