深度认知神经网络理论与方法研究

Deep neural networks have attracted much attention recently in the field of artificial intelligence. The variant networks such as convolutional neural networks and recurrent neural networks have achieved great success in the modelling of perception, which can effectively implement the nonlinear mappings from information perception to primary decision. However, when human beings perceive the world, they do not only depend on the nonlinear mappings, but include more cognition schemes, e.g., attention, memory and learning. Existing deep neural networks cannot well model these cognition processes, and thus cannot implement human-like intelligence systems. Inspired by the human visual cognitive mechanism, this project aims to study the computational modelling of the key cognitive schemes such as attention and memory, and their corresponding models. Then, we can implement the cognitive functions in terms of attention and memory based on deep neural networks, called deep cognitive neural networks. After that, we will study their learning and inference algorithms, and experimental validation. Finally, we will apply the deep cognitive neural networks to practical image and video applications. Inspired by human brain cognitive mechanism and built on deep neural networks, this project aims to develop human-like deep cognitive neural networks. This project has very strong theory, as well practicality by applying deep cognitive neural networks to visual pattern analysis.

深度神经网络是当前人工智能领域的热点研究问题，以卷积前馈网络、时序递归网络等为代表的深度网络在感知层面取得了一系列成功，有效地实现了感觉信息到初级决策的非线性映射。然而人类在认识外部世界的过程中绝不只是简单的映射关系，它还包括注意、记忆、学习等若干认知活动。现有的深度神经网络并不能有效地建模这些认知过程，因此不能实现具有类人特性的智能系统。本项目拟从人类视觉认知机理出发，研究注意、记忆等关键认知过程的计算机制，建立相应的计算模型，并基于深度神经网络实现注意、记忆等认知功能，从而提出深度认知神经网络。在此基础上进一步研究深度认知神经网络的学习与推理算法，并将深度认知神经网络应用于图像和视频分析等具体问题。本项目借鉴人脑认知机理，以深度神经网络为载体，建立具有类人特性的深度认知神经网络，具有较强的理论性和科学意义。同时，将认知神经网络应用于视觉分析等问题，具有较强的实用性和现实意义。

本项目按照计划，调研了人脑认知机理与计算机制，开展了深度认知神经网络理论与方法研究。在多认知机制计算建模方面，受注意、记忆、推理等机制启发，提出了多种新型深度认知神经网络模型；在人体运动分析方面，围绕姿态估计、运动建模等一系列任务展开研究，分别在骨架行为识别、远距离步态识别等方面取得了关键技术突破；在场景语义理解方面，从知识建模、语义概念和顺序学习等角度分别提出了多种新模型，在多个任务上取得了当时国际领先的性能。在实际应用方面，相关技术成果与华为等公司开展了较多项目合作。在数据库建设方面，建设了2个大规模视觉数据库用于验证算法。在系统建设方面，开发了2套系统平台。总体上，通过团队五年的研究，项目成果在相关领域产生了积极的学术影响，创造了良好的社会和经济效益。