Smart Device App Discovery and Installation with Lightweight Models

The patent segments the machine learning model into multiple components: a complex model for initial training and a simplified model for deployment. The complex model processes training data to generate predictions, which are then used to train a simpler model that can be efficiently deployed on smart devices. This segmentation allows high accuracy to be achieved through the complex model during training, while the deployed simplified model maintains acceptable accuracy with reduced computational requirements.

The patent introduces an intermediary processing stage where predictions from a complex model are used to train a simplified model. This intermediary step acts as a bridge between the complex analysis capability and the simplified deployment requirement. The complex model serves as a teacher model that guides the training of the student model, allowing the student model to achieve good performance without requiring the full complexity of the teacher model during inference.

The patent implements feedback mechanisms where the system continuously learns from installation outcomes and user behavior. The machine learning models are trained on historical data including successful app installations, appliance identifications, and user interactions. This feedback loop allows the system to correct biases over time by adjusting model predictions based on actual performance data, thereby maintaining automated operation while improving identification accuracy and reducing bias.

The patent performs preliminary actions by pre-training machine learning models on comprehensive datasets that include diverse appliance types, app categories, and installation scenarios. This preliminary training with balanced and representative data helps prevent bias from developing in the first place. The models are pre-adjusted to recognize various appliance and app patterns accurately before deployment, reducing the need for corrective actions later while maintaining automated operation.