ACTIVE ADAPTIVE ON-DEVICE MACHINE LEARNING

Smart overview of the Invention

Active adaptive on-device machine learning is a technique designed to allow machine learning models to dynamically adapt to changes in their environment while operating within resource-constrained settings, such as edge networks. This approach leverages a system that includes a memory with instructions and a processor to execute these instructions. The system facilitates the execution of a machine learning model that features an adaptive component capable of outputting classification results and reconstructing data streams for classification purposes. By evaluating these outputs, the system detects environmental changes and updates the adaptive part of the model accordingly if the changes are deemed relevant.

Field and Background

This innovation primarily pertains to machine learning models and artificial intelligence, focusing on their application within resource-constrained environments like edge networks. The rise of edge computing has necessitated bringing computation closer to data generation points, improving response times and reducing bandwidth usage. With the proliferation of IoT devices, there's a pressing need for efficient processing methods that can run on devices with limited resources. TinyML, a subset of machine learning, has been developed to address this need, though it faces challenges related to memory, computational power, and energy consumption.

Technical Approach

The disclosed technique involves a system that can execute a machine learning model with an adaptive part. This model operates in different phases: model inference, active on-device learning, and on-device model adaptation. During the model inference phase, predictions and data reconstructions are generated. In the active learning phase, these outputs are used to dynamically adjust thresholds and determine environmental changes. If relevant changes are detected, the system enters the adaptation phase, updating the model's adaptive part. This approach includes a drift detector using dynamically adjusted thresholds based on reconstruction errors, optimizing resource usage and minimizing overfitting risks.

Applications and Benefits

The system is applicable in various settings, such as manufacturing plants and smart city infrastructures, where it can automate processes and optimize resource usage. By reducing dependency on cloud infrastructure, the system decreases communication overhead and enhances security and privacy. This technique is particularly relevant as more enterprise-generated data is expected to exist outside the cloud, necessitating efficient edge computing solutions. The global edge computing market is poised for significant growth, and innovations like this are crucial for expanding machine learning capabilities in edge environments.

Market and Future Directions

As the edge computing market expands, with projections of it becoming a trillion-dollar industry, the demand for efficient on-device machine learning solutions will increase. TinyML is expected to play a significant role, with billions of devices projected to incorporate TinyML chipsets by 2030. Companies are investing in developing and expanding the TinyML ecosystem, offering specialized libraries and services. Current solutions often involve static models or frequent updates that can lead to overfitting and resource inefficiency. The disclosed technique addresses these challenges by allowing models to adapt dynamically based on environmental changes, optimizing resource usage and enhancing performance.