外文翻译--场景分析中的神经网络视觉计划表示(有word版）.pdfVIP

Representing Visual Schemas in Neural Networks for Scene Analysis Wee Kheng Leow and Risto Miikkulainen Department of Computer Sciences, University of Texas at Austin, Austin, Texas 78712, USA leow@cs.utexas.edu, risto@cs.utexas.edu Abstract| Using ob ject recognition in simple Consider the rst problem: limited pro cessing re- scenes as the task, this research fo cuses on two fun- sources. In practice, it is only p ossible to construct a damental problems in neural network systems: (1) neural network with a xed numb er of input units and pro cessing large amounts of input with limited re- internal pro cessing units. The weights and activities have sources, and (2) the representation and use of struc- nite precision and are b ounded within certain ranges of tured knowledge. The rst problem arises b ecause no values. The numb er of input units may b e smaller than practical neural network can pro cess all the visual in- put simultaneously and eciently. The solution is to the size of the scene (in pixels). Even if the network can pro cess a small amount of the input in parallel, and capture a large part of the scene at once, it may not b e successively fo cus on other parts of the input. This able to pro cess all the information in parallel unless it has strategy requires that the system maintains structured an exp onential amount of units and connections [1]. The knowledge for describing and interpreting successively only viable option is to pro cess a small amount of visual gathered information.