A Multi-Objective Optimization Framework for Deep Content Models in Low-Latency Networks

Xiaoying Liu
Chongqing Electronic Information College, China

Abstract:

In this paper, we propose a multi-objective optimisation model for low-latency network environments, fusing the three major metrics of deep content recognition accuracy, inference delay and resource energy consumption, constructing an improved hybrid model based on the LSTM backbone structure, and introducing the feature selection and lightweighting modules to adapt to edge computing scenarios. In the optimisation method, a fusion of reinforcement learning and non-dominated sorting strategy is adopted to combine task classification, dynamic network state and resource load information for scheduling decision. Evaluated on three task datasets, namely Yahoo recommendation, LibriSpeech speech recognition and UCF101 video inference, the results show that the model in this paper has an average inference accuracy of 92.6, an inference latency of 17.5 ma, and a unit energy consumption of 43.7 mj, which is significantly better than the benchmark methods such as Logistic Regression, Random Forest ResNet18 and other benchmark methods, and maintains robustness and high resource utilisation under multiple network loads. Experiments show that the method has good deployment flexibility and engineering practicality, and is suitable for low-latency and high-density task scenarios.