EFRI-COPN Deep Learning in the Mammalian Visual Cortex

EFRI-COPN 哺乳动物视觉皮层深度学习

• 批准号: 0835878
• 负责人: Andrew Ng
• 金额: $ 200万
• 依托单位: Stanford University
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2008
• 资助国家: 美国
• 起止时间: 2008-10-01 至 2014-09-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=0835878&HistoricalAwards=false
• 关键词: EFRI; COPN; Deep; Learning; Mammalian

This project will start to integrate what we know about the challenging task of recognizing objects from visual inputs, by drawing on the highest-performing artificial neural network systems, new models of deep belief learning from cognitive science, and new experiments on the visual cortex. The most transformative aspect of this work is that it will aim at decisive experiments which challenge traditional assumptions about purely local feedback in the learning system as such, assumptions which are prevalent in today,s mathematical models of learning in neural circuits. Many engineers and more classical systems neuroscientists believe that these assumptions are obviously false, but a decisive set of experiments would be crucial in encouraging new types of computational models of the brain, including models which fit with what actually works in image processing in technology. On the other hand, if the experiments begin to show how such learning capabilities are actually possible within the traditional paradigm, that would be equally transformative. Brain-like capabilities in image processing are an additional goal of this work; image processing is a large and growing part of the nation's cyberinfrastructure.

该项目将开始整合我们对从视觉输入中识别物体这一具有挑战性的任务的了解，方法是利用性能最高的人工神经网络系统、认知科学中的深度信念学习新模型以及视觉皮层的新实验。这项工作最具变革性的方面是，它将致力于决定性的实验，这些实验挑战了关于学习系统中纯粹局部反馈的传统假设，这些假设在今天的神经回路学习数学模型中很普遍。许多工程师和更经典的系统神经科学家认为，这些假设显然是错误的，但一组决定性的实验对于鼓励新型的大脑计算模型至关重要，包括符合技术中图像处理实际工作的模型。另一方面，如果实验开始显示这种学习能力在传统范式中实际上是可能的，那将同样具有变革性。图像处理中的类脑功能是这项工作的另一个目标;图像处理是国家网络基础设施中一个重要且不断增长的部分。