学　术

　　为加强我校各学科之间的学术交流，搭建教师学术交流平台，促进教师学术水平提升和跨学科合作，教师发展中心开展跨学科学术沙龙活动。

　　本次活动教师发展中心特别邀请美国神经形态有限责任公司首席技术总监Marwan A. Jabri博士，与我校师生共同探讨基于生物信息学的无监督学习方法来处理高阶视觉信息特征，具体安排如下，欢迎感兴趣的教师和同学参加。

　　一、主　题：Biologically-Inspired Unsupervised Learning of Higher Order Visual Features 

　　二、主讲人：Marwan A. Jabri博士

　　三、时　间：2016年7月15日（周五）上午10:30

　　四、地　点：清水河校区经管楼宾诺咖啡

　　五、内容简介：

  An important aspect of machine learning for visual pattern recognition is the understanding of how higher order visual feature detectors (tuned processing elements) develop. Understanding how cortical areas such as V4, posterior inferotemporal (PIT), and anterior inferotemporal cortices (AIT), could help shed some light.

　　We present an architecture and unsupervised learning algorithms inspired from the primate visual neural processing pathway. The architecture includes a V1 (simple and complex layers), and layers representing V4, PIT and AIT equipped with lateral excitatory and inhibitory projections. The V4 layer consists of two sublayers, integration and pooling. Hebbian learning occurs in the V4 integration layer, PIT and AIT. We show how complex visual features detectors can form in these higher cortical areas, in particular, face like tunings are observed after learning in the subsystem representing AIT on images of faces.

　　We apply the architecture and learning algorithms to the task of face recognition from the LFW and proprietary datasets. The output of the AIT layer is used as input features to a two-layer multi-layer perceptron trained for labelling. We obtain very encouraging results on fairly challenging recognition conditions, which include multiple facial poses, illuminations/brightness, and face rotations, with over 89% success rate on LFW datasets and 95% on a proprietary dataset.

　　六、主办单位：人力资源部教师发展中心

　　　　承办单位：计算机科学与工程学院

　　联系人：袁老师   电话：61830230

　　　　　　　　　　　　　　　　　　人力资源部教师发展中心

　　　　　　　　　　　　　　　　　　　　2016年7月12日