Core Viewpoint - The article discusses a novel approach to machine learning deployment at the edge, introducing a disaggregated computing model that utilizes radio frequency (RF) for computation, thereby addressing bandwidth and privacy issues associated with traditional cloud-based models [5][6][11]. Group 1: Traditional Approaches - Traditional machine learning inference methods involve either uploading model inputs to the cloud for processing or broadcasting models to edge devices, both of which have significant drawbacks such as bandwidth consumption and privacy concerns [5][6][7]. Group 2: Disaggregated Computing Model - The proposed disaggregated computing model broadcasts the machine learning model via RF, allowing edge devices to modulate inputs onto RF signals, with all computations performed in the RF domain using frequency mixers [8][11][14]. - This model eliminates the need for local storage of models on edge devices, reducing storage overhead and energy consumption [11][30]. Group 3: Experimental Validation - Experiments were conducted using a software-defined radio testbed, demonstrating the feasibility of broadcasting models to multiple edge devices, achieving a maximum vector inner product of 32,768 points with energy consumption at the femtojoule level, significantly lower than traditional digital computations [17][23][27]. Group 4: Performance Metrics - In tests on the MNIST dataset, a single-layer model achieved 95.7% accuracy with an energy consumption of 6.03 fJ/MAC, while a three-layer model maintained 98.1% accuracy using traditional methods [27]. - For the AudioMNIST dataset, the proposed method achieved 97.2% accuracy with energy consumption reduced to 2.8 fJ/MAC [28]. Group 5: Innovations and Implications - Key innovations include the ability to broadcast models wirelessly for simultaneous inference across multiple edge devices, and the use of existing RF components to perform computations without additional hardware modifications [29][30][31]. - The approach allows for scalable neural network computations, supporting modern deep learning model requirements without the need for specialized AI chips [31].