Initial analysis document on the AI engine

Initial analysis document on the AI engine

Zhao Li

1. What is the purpose of AI engine development? What is its function? What are the characteristics?

We hope to develop a complete set of artificial intelligence software development platforms to solve the smart problems that cannot be handled by the current software technology, so as to gradually form standards in this regard. I think the future technological changes in software must be in the field of artificial intelligence.

2. Data andCodeThe close combination of storage and computing, traditional programming technology data and code separation is too powerful, so that software can obviously be data-intensive andAlgorithmIntensive classification, resulting in a lack of intelligence for data-intensive software, while a lack of self-learning capabilities for algorithm-intensive software.

3. The separation of features during software design and runtime leads to a lack of intelligence in software adaptability. ========================================================== ================================

Preliminary analysis of general Pattern Recognition System

1. Target capturing algorithm, Motion Object capturing, and static object capturing

Moving objects can be determined by comparing the color difference between the two adjacent images (motion can be used as a means of system self-learning)

Static objects establish their independence through edge continuity.

2. Target Recognition Algorithms, plane object recognition, and three-dimensional object recognition

You can form a sparse matrix composed of image edges and then perform pattern matching.

Recognition Algorithms for plane objects (with only one side), which identify problems when a plane object rotates on the vertical plane (the pattern of different angles can be derived from the rotating edge image)

Main extracted features: edge shape and color distribution (stability under different lights must be taken into account). edge features are mainly determined by the ratio between edge sampling points and the Target Center.

Exclusion of Environmental Interference (such as light sources): the exclusion of a colorless light source is mainly based on the fact that it only produces light and shade changes without color distortion, the main reason for the exclusion of colored light sources is that the color difference in each area is relatively stable when the illumination is uniform.

Recognition Algorithm for a three-dimensional object (with multiple sides and angles): feature extraction can be performed by capturing a set of moving targets (for horizontal rotation) as plane targets respectively, then, the largest matching among the features is used as the matching for identification.

3. in pattern recognition, the priority of "Attention" is: moving objects, static objects. The static object recognition algorithm can be used to identify large to small partitions, or quickly match and recognize objects based on edge features.

Target Center Location Selection, fine-grained edge elimination, and blank area Elimination

4. Recognition Method: Specify the target recognition (single-mode recognition) and list the identified targets (multi-mode recognition ).

5. Identify abstract targets, for example, how to identify whether there is a TV set in the screen, but the specific TV set has different models and brands, which are very different, how to identify an abstract TV set.

========================================================== ================================

Preliminary Analysis of Neural Networks

1. What is the purpose of AI engine development? What is its function? What are the characteristics?

We hope to develop a complete set of artificial intelligence software development platforms to solve the smart problems that are hard to handle in the current software technology.

2. What is the philosophical basis? My "layer-level" Theory

3. The close combination of data and code, storage and operation. The separation of data and code in traditional programming technologies is too powerful, so that software can be clearly classified by data-intensive and algorithm-intensive types.

4. Implementation of the new concept of neural networks, used to deal with perceptual problems (inductive problems ).

Excitement refers to the degree of excitement of a single neuron.

Connectivity, the degree of association between excited neurons and target neurons (that is, the degree of consolidation of conditioned reflex)

Formula for occurrence of conditional reflection: Degree of excitement * degree of connection = degree of excitement of the target

Formula for conditioned reflection: Connection = excitement 1 * excitement 2 * simultaneously excited

The relativity of excitement and suppression, the algorithm searches for a few neurons with the highest degree of excitement

The degree of excitement decreases naturally with time (The step size is controlled by a third party and can be used as a set parameter). A benchmark clock frequency can be used as the time latitude, you can also use the calculation cycle as a clock cycle.

5. Conceptual programming mode for solving rational Problems

6. How to connect to emotional and rational Problems

7. Data Structure: Neurons can use an array as the data structure. Multiple Neurons form a fully connected neural network. The first element in the neuron array is used to store the excitement of the neuron, each other element is independent of the neural Unit and other neurons. The initial value is 0. Therefore, if the neural network contains 100 neurons, therefore, each neuron must have an array of 100 elements. Therefore, the storage scale of the fully connected neural network with N neurons is N * n. To save the storage space of the neural network, there is no need to establish too many fully connected neural networks. You can establish a multi-level network composed of multiple independent networks. To facilitate storage, each independent network can be a continuous storage space.

8. Multi-layer network connection and self-organization issues.