Abstract
Collaborative intelligence has attracted more and more attention. By properly portioning deep neural networks (DNN) and distributing the DNN calculation to the edge and cloud, we could reduce the prediction delay and power consumption to meet the actual application requirements. Toward this direction, this paper proposes a Cluster-based Edge Compression Method (CECM) to optimize data transmission between edge devices and the cloud, achieving significant data reduction with minimal impact on recognition accuracy.
Introduction
Deep Learning models are powerful but resource-intensive. In the context of electromagnetic object recognition, deploying these models typically involves two extremes: sending raw data to the cloud (consuming high bandwidth and causing latency) or running entire models on edge devices (draining batteries and hitting hardware limits).
"Collaborative intelligence" offers a middle ground by splitting a Deep Neural Network (DNN) between the edge and the cloud. However, a significant hurdle remains: the "data explosion" phenomenon.
The Data Explosion Challenge
A common misconception is that processing data through initial DNN layers reduces its size. In reality, the output of intermediate layers is often significantly larger than the original input data. This increase in data volume at the split point can lead to communication delays that exceed the time required to simply transmit raw data to the cloud.
As shown in Figure 1, the size of output data (dark bars) in early layers like conv1 and relu1 is substantially higher than the initial input. Without compression, collaborative intelligence becomes a bottleneck rather than a solution.
Methods: Cluster-based Edge Compression (CECM)
To address this, the researchers developed the Cluster-based Edge Compression Method (CECM). The workflow involves:
- Model Partitioning: Splitting a pre-trained DNN into edge and cloud components.
- K-means Compression: Instead of transmitting high-dimensional floating-point vectors, the edge device performs K-means clustering on the activation values.
- Efficient Transmission: The edge sends a small "codebook" of cluster centers and a label vector indicating which center each value belongs to.
- Reconstruction: The cloud reconstructs an approximate version of the original vector using the codebook and labels before completing the inference.
Experimental Evaluation
The method was evaluated using the RadioML2016.10a dataset, which contains 220,000 samples across 11 modulation modes with varying Signal-to-Noise Ratios (SNR) from -20dB to +18dB. The baseline model used for comparison was VT-CNN2.
Key Results
The CECM method demonstrated that DNN activations are highly redundant. The researchers found a critical threshold at $k=2$ clusters; beyond this point, recognition accuracy becomes remarkably stable.
As illustrated in Figure 6, when $k \ge 3$, the accuracy of CECM is nearly identical to the uncompressed baseline. At the optimal tradeoff point of $k=4$, the system achieved a compression ratio of 7.81%, transmitting less than 1/12th of the original intermediate data while maintaining high precision.
Conclusion and Future Work
CECM effectively mitigates the communication bottleneck in collaborative intelligence for RF signal recognition. By squashing intermediate data through clustering, it enables real-time AI inference on resource-constrained edge devices. Future research will focus on extending this method to multi-node scenarios and developing dynamic partitioning schemes that adapt to fluctuating network conditions.
Summary updated