作为神经科学领域的一种仿生模型，预测编码因能够对大脑视觉皮层的组织结构和感知推断功能做出合理性解释而受到广泛关注。由于假定神经网络的隐层表示同时受低层和高层神经元表示的影响，且隐层表示没有显式计算公式，预测编码的推断功能必须通过迭代计算的优化过程实现，导致计算复杂度随网络层数超线性增长，无法应用于实际问题。申请人前期提出通过直接学习从输入到隐层表示的回归映射，可将单隐层预测编码模型的推断速度提高65-77倍。本项目将以预测编码的推断问题为切入点，进一步研究可进行高效推断的深度预测编码网络结构，研究从确定性推断模式到概率型推断模式的拓展方法，研究可提高所提模型判别能力和重构能力的高效学习算法。研究结果有望构建更加接近于生物神经系统的网络模型，进一步提升深度结构模型的特征提取能力和表示推断效率，拓展其应用范围。

As a biomimetic model, predictive coding (PC) has become increasingly popular for explaining a range of neural inference functions and many aspects of brain organization. The basic PC assumes that the latent representation of neural networks is affected by both the lower- and higher-level representations, and that there is no explicit formulation for computing the latent representation. Therefore, the PC has to perform inference through an iterative optimization procedure, which results in the superlinear increasing of computational complexity with the number of layers, and hindering its applications. The applicant has previously proposed that by learning a regression mapping from the input to the representation, the inference process of the improved model would be 65-77 times faster than that of the original PC model. Aim to solve the problem of inference inefficiency in PC, this project will further study the deep PC network with the efficient inference process, the deterministic and probabilistic inference methods, and the associated efficient learning algorithms to improve the model's performance on discriminative and reconstruction tasks. The goal of this project is to construct new network models inspired by the biological neural systems, and improve the efficiency of feature learning and inference, as well as exploiting the models' practical applications.