Self-trainable and adaptive sensor intelligence for selective data generation

Arghavan Rezvani, Wenjun Huang, Hanning Chen, Yang Ni, Mohsen ImaniFrontiers Artif. Intell 2024 · OpenReview

TL;DR

This paper introduces a novel framework for "intelligent sensing" in IoT devices that significantly reduces energy consumption and data transmission by learning to identify and transmit only relevant data frames. It achieves this through a unique combination of online learning, active learning, and knowledge distillation, allowing near-sensor models to adapt dynamically to their environment without extensive pre-training or manual labeling.

What Problem Does This Paper Address?

The proliferation of Internet of Things (IoT) devices generates an overwhelming amount of data, posing significant challenges for energy consumption and data transmission, especially for edge devices with limited computational power. While many IoT applications rely on complex machine learning (ML) models for data analysis, deploying these models directly on edge devices is often infeasible due to resource constraints. Offloading computations to servers can be inefficient, as it often involves transmitting large volumes of redundant data. Existing "intelligent sensing" approaches, which use lightweight models near sensors to filter data, typically lack adaptability, requiring extensive pre-training and struggling with dynamic environmental changes. This paper addresses the need for an adaptive, resource-efficient, and self-trainable intelligent sensing framework for edge devices.

Methods

The proposed framework builds upon existing intelligent sensing concepts but introduces several key innovations:

  • Online Learning: The near-sensor model is trained after deployment using pseudo-ground-truth labels generated by a more powerful server-side model. This eliminates the need for costly manual labeling and allows the model to adapt dynamically to changing data distributions in real-time.
  • Active Learning (AL): To manage storage and computational constraints, AL is employed to intelligently select the most informative data points for retraining the near-sensor model. This is achieved by maintaining a fixed-size buffer and prioritizing uncertain or misclassified samples.
  • Knowledge Distillation (KD): A compact "student" model (near-sensor) is trained to mimic the behavior of a larger "teacher" model (server-side). This allows for significant model compression while retaining high performance, a crucial aspect for edge deployment. The paper specifically adapts KD techniques for object detection tasks, which differ from standard classification KD.
  • Modified Loss Function: The authors utilize a customized loss function for the student model that emphasizes the objectness score, a technique shown to improve performance in this context.

The framework operates in alternating "training" and "execution" phases. During training, all data is sent to the server for model refinement. During execution, the near-sensor model filters data, and misdetections are collected for subsequent retraining.

Datasets

The primary dataset used for training and evaluation is the Microsoft Common Objects in Context (MS COCO) dataset. The authors specifically selected and re-labeled images from this dataset to focus on animal detection, categorizing frames as either "Frames of Interest" (FOI) (animals) or background.

Key Insights and Ideas

  • Self-Sufficiency for Edge Intelligence: The core idea is to enable edge devices to become "intelligent" without constant human intervention or extensive pre-configuration. The online learning approach, leveraging server-side predictions as labels, is a significant step towards this goal.
  • Adaptability is Key: The framework's ability to adapt to environmental shifts (e.g., new object types appearing) is crucial for real-world IoT deployments. The online learning mechanism directly addresses this, preventing performance degradation over time.
  • Balancing Performance and Resources: The combination of KD and AL demonstrates a sophisticated approach to achieving high performance on resource-constrained devices. KD compresses the model, while AL ensures that the limited training data is used most effectively.
  • The Power of Misdetections: Misdetections, often seen as errors, are reframed as valuable learning signals for the near-sensor model, highlighting areas where it needs improvement.

Key Results

  • Significant Energy and Data Reduction: The intelligent sensing framework shows substantial improvements in energy consumption and data transmission compared to conventional sensing, especially as the ratio of background to FOI increases (Figure 6A).
  • Performance Convergence: The near-sensor model's performance (AUC) rapidly approaches that of a model trained on the full dataset, even when starting with a small percentage of data and relying on subsequent misdetections for training (Figure 3A).
  • Robustness to Environmental Shifts: The adaptive framework with online learning significantly outperforms a non-adaptive intelligent sensing approach when faced with a shift in data distribution (e.g., introduction of new animal types) (Figure 4).
  • Effective Model Compression: Knowledge distillation successfully compresses larger models into smaller ones (e.g., YOLOv5n to YOLOv5ns) while maintaining competitive accuracy (Figures 5A, B, E).
  • Reduced Training Data Needs: Active learning reduces the volume of training data required by over 80% with minimal AUC loss (<1%), leading to substantial reductions in training time and energy consumption (Figures 6B, C, D, E, F).

Limitations and Future Work

  • Environmental Changes: While the framework adapts to environmental changes, the paper notes that the most influential training occurs during the initial phases. Occasional updates are still required.
  • Scalability: The authors suggest future work to explore the scalability of their framework.
  • Algorithm Sophistication: Extending the approach to more sophisticated ML algorithms is also proposed.
  • Broader Applications: The framework's applicability to a wider range of sensing applications is an area for future exploration.
  • Quantization Impact: While fp16 and int8 quantization had no impact, int5 and int4 quantization showed a slight degradation in AUC, indicating potential trade-offs at very low bit precisions (Table 1).

Summary updated

BibTeX
@inproceedings{rezvani2024self,
  author = {Arghavan Rezvani and Wenjun Huang and Hanning Chen and Yang Ni and Mohsen Imani},
  title = {{Self-trainable and adaptive sensor intelligence for selective data generation}},
  booktitle = {Frontiers Artif. Intell},
  year = {2024},
  url = {https://openreview.net/forum?id=EsnbDEmQ6g},
}
Figure