Core Concepts
This paper proposes a Cloud-RAN based collaborative edge AI inference architecture that jointly optimizes transmit precoding, receive beamforming, and quantization to maximize the inference accuracy by leveraging the discriminant gain metric.
Abstract
The paper proposes a Cloud-RAN based collaborative edge AI inference system with the following key components:
Edge Devices:
Geographically distributed devices capture real-time noise-corrupted sensory data and extract local feature vectors.
The local feature vectors are transmitted to the remote radio heads (RRHs) via over-the-air aggregation using AirComp technique.
RRHs:
RRHs receive the aggregated local feature vectors from devices, quantize them, and forward them to the central processor (CP) through capacity-limited fronthaul links.
Central Processor (CP):
The CP further aggregates the received quantized feature vectors from RRHs and performs the downstream inference task.
The key challenges lie in simultaneously suppressing the sensing noise, AirComp distortion, and quantization error. To address these, the paper proposes a joint optimization of transmit precoding, receive beamforming, and quantization error control to maximize the inference accuracy based on the discriminant gain metric.
The discriminant gain metric measures the discernibility of different classes in the feature space and serves as a surrogate for the inference accuracy. By maximizing the discriminant gain, the most significant feature elements can be well received at the CP, leading to enhanced inference performance.
The proposed optimization problem is non-convex, so an alternating optimization approach is developed to solve it efficiently. Numerical experiments demonstrate the effectiveness and superiority of the proposed method compared to various baselines.
Stats
The paper does not provide any specific numerical data or statistics. It focuses on the system model and optimization framework for the collaborative edge AI inference over Cloud-RAN.
Quotes
There are no direct quotes from the content that are particularly striking or support the key logics.