再一个


所有跟贴·加跟贴·新语丝读书论坛

送交者: 短江学者 于 2018-06-21, 03:50:35:

回答: 一个文章说这个事 由 短江学者 于 2018-06-21, 03:29:17:

引用:

Signal representations suitable for easy extraction of invariants are crucial for many tasks of general intelligence. Our vision system has evolved for this purpose. However, the pixel intensity based representations that underlie many successful image processing algorithms are seemingly inconsistent with our vision functions. For example, our eyes make rapid involuntary movements but our visual perceptions of the environment are stable. In fact, we see many objects in our visual scene simultaneously, and when focusing we see some very clearly. Our two eyes receive different sets of light signals, but we don’t see double. Even in just one eye, we have millions of non-collocated photoreceptors, and yet they work together to provide a seamless perception of the visual scene. How do they collaborate at the algorithmic level? Our hypothesis is that all these signal sensors sample the same set of signal values that characterize the visual scene, thereby helping each other in the spirit of the law of large numbers, where repeated observations of a fixed set of statistical parameters help to improve estimation accuracy.

In our proposed signal representation model, visual (spatial) signals are converted to time signals in order to propagate and convey information to subsequent memory and analogical mechanisms. Such time signals, representing small visual regions, should be able to collaborate with those of neighboring regions in order to represent the larger scene. In Section 2, we develop such a conversion algorithm. The basic algorithmic unit is modeled by a partial differential equation with second order time dynamics. The algorithm aims to exhibit the following necessary features of natural cognitive computation:
• Biological plausibility;
• Accuracy and robustness;
• High fidelity information conversion;
• High information throughput using slow processing units;
• Foundations for high-capacity/fast-access memories, concept abstraction, and analogical thinking.

With the throughput issue addressed in Section 3, our signal model provides the foundations for a memory network comprised of cascaded resonating circuits with a wide range of resonance frequencies. The network nodes are active, in contrast to the usual models for memory networks





所有跟贴:


加跟贴

笔名: 密码: 注册笔名请按这里

标题:

内容: (BBCode使用说明