Pattern recognition-Reading Notes 2

1,Features are the key to determining similarity and classification. After the purpose of classification is determined, how to find suitable features becomes the core issue of cognition and recognition.

Notes:

What is the current approach to extract features for specific fields? For example, in the vehicle track detection of large jobs, parameters such as the extraction speed, direction, and rotation degree can be used as the feature vector component for subsequent analysis. Is there a general feature extraction method? What is our human brain doing? Before learning a new thing, it is impossible for the brain to know in advance what are the characteristics of this thing? So how do we learn it?

 

2,The process of analyzing animal sensory information processing can be seen that there are neurons in the animal and human brain that respond to specific features and form a structure from simple to complex layer-by-layer feature extraction.

Notes: Does this mean that at the initial layer of the receptor, the data is collected in simple but large quantities and redundant, and then abstracted layer by layer? But how does our human body normalize this information? Whether the stimulation received from the retina and the stimulation received by the skin are in the same "format" after the conversion of the receptor, then, if we can simulate the conversion of the eye retina to the "format" of specific nerve stimulation, can we use a camera instead of the eyes to regain light for blind patients?

 

3,It is extremely difficult to calculate the weight of a Chair directly from the image of the Chair on the retina, but we can break down this difficulty into several layers of simpler problems. For example, the weight can be calculated by the weight and volume, and the weight can be determined by the material. The material can be known from the texture and color. The volume can be calculated by the shape, and the three-dimensional shape can be inferred by the two-dimensional shape, the two-dimensional shape can be calculated from the edge, and the edge can be obtained from the lower-layer feature-local direction.

In order to obtain new features through learning, we need to have adequate low-level features. If low-level features are rich enough, and new features can be obtained through selection and simple operations, it is not difficult to generate new features.

Notes: From the first section, we can see that the completion of a complex task is accomplished by splitting it into several simpler tasks. If the human brain is working in this way, how does it know how to "divide and conquer" the problem? In the second section, since it is a "new" feature, it at least indicates that the feature did not exist in the brain before. It is easy to understand the existing features from low-level features, however, because of the existing low-level features, it seems that a new branch is added to the ultimate switch case statement, so where does our human brain put this new branch?

I used to think that the brain is a dynamic storage and computing space, and both storage and computing are completed by the same computing. In other words, our memories exist in computing. That is to say, the memory of the brain is not written somewhere like a hard disk, but directly linked by neurons to create the "by-product" of memory ".

Suppose that at the beginning, the direct connection of neurons was empty. Then, starting from the first moment of thinking, neurons began to explore whether such stimulation had occurred, because they were in the empty state, there must be no branch to choose from. At this time, neuron A pulls neuron B and connects it. The condition that a activates B is the stimulus feature (maybe the voltage size ), in this way, a branch is generated. This branch will be activated when this stimulus occurs again for the second time. In the process of human life, there will be a variety of neural connections, and the weight of the branch that is often selected will increase, which can explain the "Practice makes perfect" issue, the connections that are not often used become weak, but after the connections are generated, they should exist until they are damaged by trauma, otherwise they will not be able to retrieve memories through hypnosis.

After a person is affected by an accident and his brain loses his memory, the memories of the patient can be aroused if he relives things he has experienced. In my opinion, it is to re-experience lost memories to repair the connection of neurons. People will lose or produce memory fragments because the direct connections of neurons break. If the connections between them can be restored, the fragments can be respliced, the human memory is restored.