深度结构和深度学习在机器学习、数据挖掘、大脑建模等领域起着愈来愈重要的作用。心理行为分析、认知学研究和神经生理实验的大量证据表明，多层级联组织是大脑的一个关键特征。在这种多层级联深度结构中，层与层之间前向连接和后向连接在组织结构和运作方式上存在明显的非对称差异，这种差异对于大脑的有效运作具有不可或缺的重要作用，而现有的深度结构和深度学习方法并未有效考虑这一特征。本项目试图借助近年来提出和发展的三类描述大脑运作机制的神经计算原理和理论：自由能原理、贝叶斯脑假设和有效编码原理，探索在深度机器学习模型中引入前向连接和后向连接不对称性结构，研究后向连接的组织方式和约束条件，构建以双层网为基础学习单元的深度学习方法，研究利用表示神经元和误差神经元分别表述的同层异类网络模型，研究将权重作为随机变量的后验分布学习方法及抽样运作机制。所获结果有望进一步提升深度学习算法性能，拓展其应用范围。

Deep architecture and deep learning play a very important role in machine learning, data mining, and brain modeling. There is strong behavioral, cognitive and physiological evidence that the hierarchical organization is a key architectural principle. In this hierarchical deep architecture, the forward and backward connections show both a microstructure and functional asymmetrical distinction. Such an asymmetrical distinction is very important for brain work and is not involved in the existed artificial deep learning model. In this project, We will use three neural computational theories, free-energy principle, Bayesian brain hypothesis, and the principle of efficient coding, that describe the working mechanisms of the brain, to construct a deep learning machine with asymmetrical forward and backward connections, explore different constraints and connection forms on backward connections, construct double-level network and design its fast training algorithms. We will also consider introducing different kind of neurons such as error neurons and representation neurons in the same level, computing a posterior distribution over connection weights and designing the sampling methods. The proposed models and learning algorithms will improve the performance and expand the application scopes of deep learning machines.