复杂电商环境下自主高效可泛化的神经网络模型研究

针对浙江省数字经济快速发展的需求，为实现对海量电商数据的智能理解、抽象和推理，本项目重点开展面向复杂电商环境下的神经网络模型研究，拟从学习机理、模型安全和可解释推理等角度切入，有效提升神经网络的效率和自主可泛化能力。基于此，本项目主要研究：（1）多任务小样本元学习，通过构建多任务元学习框架，实现神经网络建模能力和泛化能力的跃升；（2）基于结构搜索和动态路由的紧致模型学习，通过挖掘数据驱动的语义拓扑结构，实现神经网络自动结构搜索和自适应路由推理；（3）复杂动态环境下去中心化加密的分布式智能联邦学习，探索基于智能合约的联邦学习构建方法，将安全合并与差分隐私相结合而实现隐私保护；（4）面向数字经济场景的可解释性因果推理，建立群智涌现的可计算模型。本项目从神经网络学习机理出发，建立面向复杂电商环境的自主高效可泛化的神经网络模型，促使数字经济向更加智能和安全的方向发展。

In response to the rapid development of the digital economy in Zhejiang Province, this project focuses on the research of neural network models for complex e-commerce environments, in order to achieve intelligent understanding, abstraction and inference of massive e-commerce data, greatly enhancing both the efficiency and the autonomous generalization capabilities of neural networks. Based on this, the main research points of this project is: (1)Multi-task and few-shot meta-learning. A leap in modeling capability and generalization capability of neural network is achieved by building a multi-task meta-learning framework. (2)Compact model learning based on architecture search and dynamic routing inference. Automatic architecture search and adaptive routing inference of neural network is achieved using data-driven semantic topology.(3)Decentralized, encrypted and distributed federal learning in complex dynamic environment. Exploring the construction of federal learning based on smart contract, privacy protection is achieved by combining security merging and differential privacy. (4)Interpretable causal inference in digital economy scenarios to establish a collective-intelligent computing model. Starting from the neural network learning mechanism, this project will eventually establish efficient and autonomously generalizable neural network models for complex e-commerce environments, promoting the development of the digital economy in a more intelligent and secure direction.