Abstract
Deep learning models deployed on edge devices often struggle with real-time data distribution shifts. Traditional on-device fine-tuning is computationally expensive, slow, and prone to overfitting. We present Persona, a novel framework that utilizes a cloud-based Parameter Editor to generate weight-editing matrices. This approach enables backpropagation-free, millisecond-level adaptation, achieving superior accuracy and efficiency compared to traditional retraining methods.
Introduction: The Challenge of Dynamic Edge Data
Models deployed on mobile devices or autonomous systems are typically static, trained on global datasets and fixed upon deployment. However, real-world data is inherently dynamic—user preferences evolve, and environments change. This "real-time data distribution shift" renders static models obsolete. Existing solutions rely on on-device fine-tuning, which is hindered by hardware resource constraints, local data scarcity, and high latency that fails to meet real-time requirements.
The Persona Framework
Persona addresses these hurdles by splitting the adaptation workload between the cloud and the device, treating model adaptation as a parameter generation task rather than an optimization problem.
Parameter Editor (PE) and Editing Matrices
The core of Persona is the Parameter Editor located in the cloud. It processes real-time features from the device to generate a Parameter Editing Matrix. This matrix is transmitted to the device, where it "edits" the weights of specific adaptive layers through simple addition. This bypasses the need for expensive gradient descent on the edge device.
Multi-Prototype Strategy and Dynamic Assignment
To ensure stability and generalization, Persona clusters devices into "Prototypes" representing distinct data distributions. Instead of a single global model, devices are initialized from a prototype model that is distributionally "closer" to their local data. If a user's behavior shifts significantly (e.g., moving from electronics to beauty interests), the system dynamically reassigns them to a more appropriate prototype.
Cross-Layer Knowledge Transfer (CLKT)
To maintain architectural consistency, Persona employs a sequence modeling module for Cross-Layer Knowledge Transfer. This ensures that parameter edits across different layers are synchronized and do not counteract each other, providing a unified adaptation strategy across the neural network.
Experimental Results
Persona was evaluated across recommendation datasets (Amazon, Douban, MovieLens) and computer vision tasks (CK+ facial expression recognition).
- Efficiency: Persona reduces adaptation latency to approximately 1 second, whereas traditional fine-tuning often exceeds 10 seconds on similar hardware.
- Recommendation Performance: On the MovieLens dataset, Persona (multi-prototype) achieved an NDCG@5 of 0.5036, significantly higher than the base GRU4Rec model (0.4903) and the fine-tuned version (0.4942).
- Vision Performance: On the CK+ dataset using MobileNetV3-Small, Persona reached 73.74% accuracy, outperforming Test-Time Adaptation (TTA) (70.69%) and cloud-collaborative DUET (67.68%).
Conclusion
Persona demonstrates a scalable and efficient solution for real-time edge adaptation. By leveraging cloud-based editing and prototype clustering, it provides a backpropagation-free method to handle dynamic data shifts. Future work will focus on reducing communication overhead for uploading features and exploring the adaptation of deeper backbone layers without increasing computational complexity.
Summary updated