网范文：“Sparse visual models ” 哺乳动物的大脑皮层致力于视觉处理。猕猴至少有50%的大脑皮层的直接参与。视觉皮层的功能依赖于组织的连接、突触形成的类型和突触后的神经元如何应对，英语毕业论文，以及整合突触输入。这篇生物范文讲述了视觉形成模式。视觉皮层分为五个独立的区域,V1、V2、V3,V4和V5 。这些区域进一步细分和发送信息到任何大脑的其他区域。生理和解剖探讨表明组织在视觉皮层形成一个视觉世界的表示。
这种表示措施是通过模块化略论形成平行的层次。这个总体分为两条平行的途径。这些细胞特征通过通路，英语论文网站，穿过腹侧流,一个人可以想象层次结构神经元，感受稳步增加大小。下面的范文进行讲述。
Abstract
Much of the mammal cortex is devoted to visual processing. In the macaque monkey at least 50% of the neocortex appears to be directly involved in vision. The function of visual cortex is dependent on the organization of its connections, the types of synapses they form, and how postsynaptic neurons respond to and integrate synaptic inputs. Roughly, the visual cortex is divided into 5 separate areas, V1, V2, V3, V4, and V5/MT (Zeki 1999). Each of these areas is further subdivided and sends information to any of 20 or more other areas of the brain that process visual information (Hubel 1995). Physiological and anatomical studies suggest that organizing principles in visual cortex is forming an economical representation of the visual world. This representation is formed through a modular analysis that is both parallel and hierarchical. This general arrangement is subdivided into two parallel pathways as Fig 2. Cells in dorsal MST are particularly sensitive to small moving objects or the moving edge of large objects. These cellular characteristics make the dorsal pathway especially able to quickly detect novel or moving stimuli.
Moving through the ventral stream, one can conceive a hierarchy of neurons with the steady increase of receptive field sizes. At corresponding eccentricities near the fovea receptive fields in V2 are (in linear dimensions) 23 times larger than in V1; in V4 perhaps 5-6 times larger; cells in IT (inferotemporal cortex) have receptive fields that can include the entire central visual field, on both sides of the vertical midline (Lennie 1998). These large receptive fields are presumably necessary to recognize large complex objects and may mediate the ability to recognize objects of any size as the same, regardless of their retinal location.
Hierarchical network architecture
An essential behavior of animals is the visual recognition of objects that are important for their survival. Human activity, for instance, relies heavily on the classification or identification of a large variety of visual objects (Logothetis and Sheinberg 1996). One of the major problems which must be solved by a visual system for object recognition is the building of a representation of visual information which allows recognition to occur relatively independent of size, contrast, spatial frequency, position on the retina, and angle of view, etc (Ullman, Vidal-Naquet et al. 2017).
This requires that features extracted by the visual pathway create a rather complete representation of the current sensory scene using the principle of sparse coding, which means that at any one time only a small selection of all the units is active, yet this small number firing in combination suffices to represent the scene effectively. Hubel and Wiesel proposed a model in which V1 simple cells with neighboring receptive fields feed into the complex cell with same receptive-field orientation and roughly the same positions, thereby endowing that complex cell with phase and shift invariant features.
Following this, visual processing in cortex is classically modeled as a hierarchy of increasingly sophisticated representation (Fukushima 1980; Marr 1982; Biederman 1987; Poggio and Edelman 1990; LeCun, Boser et al. 1992; Riesenhuber and Poggio 1999). Here we present a hierarchical network architecture (see Fig. 3) with sparse coding constraint to extract low level features (such as edges, orientations, spatial frequencies, and contours) for further processing in the ventral pathway, such as part-based shape representation in cortex V4 (Desimone, Schein et al. 1985; Schiller 1995; Pasupathy and Connor 1999; Pasupathy and Connor 2017), and object recognition in the inferotemporal cortex (IT) (Kobatake and Tanaka 1994; Tanaka 1996).
Complex cells model
V1 is located in the occipital lobe at the back of the brain. Nearly all visual information reaches the cortex via V1. The receptive fields of V1 simple cells are localized in space and time, have band-pass characteristics in spatial and temporal frequency domains, are oriented, and often sensitive to the direction of motion of a stimulus. This sort of properties encourages the notion that the purpose of the neurons in V1 is to construct economical description of the images. Independent Component Analysis (ICA) on natural images produces receptive fields like those of simple cells (Olshausen and Field 1996; Bell and Sejnowski 1997; Olshausen and Field 1997; Lee 1998; Hyvarinen, Karhunen et al. 2017).
End-stopped cells model
In most respects, cells in V1 and V2 are not remarkably different. V2 is strongly reciprocally connected with V1, and end stopping seems to be more prevalent there, particularly in the pale strips. An ordinary simple cell or complex cell usually shows length summation: the longer the stimulus line, the better is the response, until the line is as long as the receptive field; making the line still longer has no effect. For an end-stopped cell, lengthening the line improves the response up to some limit, but exceeding that limit in one or both direction results in a weaker response (Hubel 1995).
Conclusion
Our approach is related to Hoyer’s contour coding network (Hoyer and Hyvarinen 2017). However, Hoyer computed the complex cell responses by a simple energy model, therefore the receptive fields in his V1 layer are fixed, or precalculated. In contrast, our approach uses the end-to-end learnt receptive fields, and thus represents the natural image sparsely and sufficiently (see Fig. 4). Also, the property of the receptive fields and their sizes in our architecture are richer and more in-line with the diversity known from biology. Note that repeating Hoyer’s experiments using 100,000 image patches and 100 iterations took 2 days on the same computer mentioned above, the selective resulting basis patterns are shown in Fig. 7. Practically, using the responses of V2 cells in our architecture, we have trained the Figure 7.
A selective set of basis function learned by Hoyer’s network. V4 layer and obtained some interesting results, such as object parts. However, the responses of V2 cells produced by Hoyer’s model are too weak to be further used in high layers. Our study is also related to the predictive coding model of (Rao and Ballard 1999), in which, the feedback connections from a higher- to a lower- order visual area carry predictions of lower-level neural activities. The feedforward connections carry the residual errors between the predictions and the actual lower-level activities. They proposed that endstopping cell stopped responding when the stimulus length was increased because then it could be predicted and there were no residual errors.()
网站原创范文除特殊说明外一切图文作品权归所有；未经官方授权谢绝任何用途转载或刊发于媒体。如发生侵犯作品权现象，保留一切法学追诉权。()
更多范文欢迎访问我们主页 当然有需求可以和我们 联系交流。-X()
|
