RI: Small: A Hierarchical Approach to Unsupervised Feature Discovery

与大众媒体的一些描述相反，人类在理解周围的视觉世界方面仍然比任何计算机程序都要好得多。如果我们了解了视觉系统是如何做到这一点的，也许就可以建立更好的人工视觉系统。这个项目的目标是了解大脑是如何表现视觉世界的，以及为什么。遵循理查德·费曼（Richard Feynman）的一句名言——我不能创造的东西，我不理解——这个项目的方法是创建一个从自然输入（图像、视频）中学习的计算机系统，假设视觉系统的运作目标是有效地代表世界。然后将这些表征与视觉神经元的测量结果进行比较。长期目标是了解早期视觉处理层在人类视觉通路中的功能作用。该模型基于有效编码假设，其中早期视觉通路通过在其输出中有效编码视觉信息来捕获其视觉输入的统计结构。大多数遵循这一假设的计算模型只关注于一个或两个视觉层的建模。在这个项目中，Cottrell的团队提出了一个分层信息处理模型，该模型与高效编码假说一致，但提供了迄今为止最完整的早期视觉处理层描述。在该模型中，首先使用稀疏主成分分析（SPCA）对视觉输入进行压缩以降低噪声，然后使用过完全稀疏编码（overcomplete Sparse Coding）扩展数据维度以捕获视觉输入的统计结构。然后一个非线性激活函数为下一层格式化这一层的输出，整个过程重复进行。初步研究表明，所得到的分层模型可以学习具有早期视觉通路神经元接受野特性的视觉特征，包括视网膜神经节细胞、LGN、V1简单和复杂细胞以及V2细胞。

Contrary to some depictions in popular media, humans are still far better than any computer program at understanding the visual world around them. If we understood how the visual system does this, perhaps better artificial vision systems could be built. The goal of this project is to understand how the brain represents the visual world and why. Following a mantra famously credited to Richard Feynman -- What I cannot create, I do not understand -- this project's approach is to create a computer system that learns from natural input (images, videos), assuming that the visual system operates with the goal of efficiently representing the world. These representations will then be compared to measurements of visual neurons. The long term goal is to understand the functional roles of the early visual processing layers in the human visual pathway.The model is based on the efficient coding hypothesis, in which the early visual pathway serves to capture the statistical structure of its visual inputs by efficiently coding visual information in its outputs. Most computational models following this hypothesis have focused on modeling only one or two visual layers. In this project, Cottrell's group proposes a hierarchical information processing model, which concurs with the efficient coding hypothesis, yet provides the most complete description so far of the early visual processing layers. In this model, the visual inputs are first compressed to reduce noise using Sparse Principal Components Analysis (SPCA), then the data dimensions are expanded to capture the statistical structure of the visual inputs using overcomplete Sparse Coding. A nonlinear activation function then formats the outputs of this layer for the next layer up, and the whole process is repeated. Preliminary work shows that the resulting hierarchical model can learn visual features exhibiting the receptive field properties of neurons in the early visual pathway, including retinal ganglion cells, LGN, V1 simple and complex cells, and V2 cells.