智能计算新范式将无线通信融入神经网络

Core Insights - Nanjing University's Brain-like Intelligent Technology Research Center has proposed a novel "edge-cloud integration" intelligent computing paradigm that incorporates wireless communication into neural networks, significantly reducing power consumption of wireless communication modules while maintaining high-precision inference capabilities [1][2]. Group 1: Research Findings - The new paradigm addresses the challenge of high energy costs associated with transmitting massive data to cloud devices, which has hindered the large-scale application of edge intelligence [1]. - The research team developed a neural network inference system and a wireless communication system using a self-developed analog in-memory computing chip, enabling efficient processing of neural network models and wireless transmission of computation results [1]. Group 2: Training Methodology - The team introduced a communication-aware training approach that optimizes the "algorithm-hardware" collaboration, significantly reducing energy consumption and hardware costs while enhancing system robustness [2]. - By integrating wireless communication into neural network optimization training, the system learns to manage energy expenditure for data transmission, allowing for high-precision inference tasks at lower energy costs across various wireless environments [3]. Group 3: Implications for Industry - This research breaks traditional design conventions of edge-cloud collaborative computing systems, proposing a new task-centered, end-to-end collaborative optimization paradigm for efficient intelligent computing in large-scale terminal devices [3].