Chapter One Introduction
Mainly for the review of exams and the annotation of some key content, the content of the book has been deleted, and my own notes and reflections have been added.
Note: This series refers to the model classification of the Machinery Industry Press and the watermelon book of Teacher Zhou Zhihua. It stands to reason that it should be submitted for reprint, but I gave up because I didn't know how to write the original link, but here I still have to explain that the type is [reprint], and the same is true for the subsequent chapters, so I won't repeat the description.
0.1 Pattern Recognition System
Sensors
such as camera and microphone arrays, temperature sensors and other input data, images, text, speech signals, etc. such as weather monitoring instruments or flux meters, etc. may need to consider bandwidth, resolution, sensitivity, distortion, signal-to-noise ratio, delay etc.
Segmentation and organization
For example, object and background segmentation in images, phoneme segmentation in speech recognition. The problem at hand is how to identify or organize the different parts of a composite object (similar to the problem of division and splicing).
Feature extraction
A good fish classifier has no effect on fingerprint recognition or cell recognition. In general, a special feature extractor is needed for the research object, so it is more dependent on specific problems and specific fields.
Classifier
The role of the classifier in the system is to assign a class label to a measured object according to the feature vector obtained by the feature extractor. The ease of classification depends on 1. the fluctuation of eigenvalues between different individuals from the same class (the fluctuation may come from the complexity and noise of the problem); 2. the difference between the eigenvalues of samples belonging to different classes.
Noise: If a perceived pattern property is not derived from a model of the true pattern, but rather from some randomness in the environment or a flaw in sensor performance, then it is noise.
Postprocessing
Conceptually, the simplest measure of classifier performance is the classification error rate, which is the percentage of new patterns that are flagged as the wrong class. So a general approach is to seek the classifier with the lowest classification error rate, and a better approach is to recommend an action that reduces the overall cost (risk).
0.2 Design Cycle
Designing a pattern recognition system usually involves the repetition of several different steps: data collection , model selection , training , evaluation , and computational complexity .
These steps all face common problems, such as data collection, how to know that enough representative data has been collected? In feature selection we hope to find feature sets that are easy to extract, insensitive to irrelevant deformations, insensitive to noise, and effective at distinguishing patterns from different classes. For model selection, how do we know that one type of model should be rejected in favor of another? How can the model be expected to improve? Training, the process of using sample data to determine a classifier is training a classifier. Experiments and experience over the past 25 years have shown that "sample-based learning" is the most effective method. What exactly is this method? The evaluation, in the case of fish classification, went from a single feature to two features on the basis that the evaluation of the classification error rate for a single feature was not good enough. An overly complex system can achieve perfect performance on the training sample set, but overfitting will occur. So is there a principled way to determine a classifier with optimal complexity? What is the trade-off between computational simplicity and classification performance in computational complexity? What are the functions of feature dimension, number of patterns, and number of categories used?
0.3 Learning Algorithms
Supervised learning
In supervised learning, there is a teacher signal, which can provide a class label and a classification cost for each input sample in the training set sample (equivalent to a decision maker to tag each sample), and find a way to reduce the overall The direction of the cost (e.g. gradient descent).
Unsupervised Learning
In unsupervised learning there is no teacher signal. The system automatically forms clusters or natural organizations for the input samples. The rules of clustering are determined by explicit or implicit criteria employed by the clustering system.
Reinforcement learning
does not require a teacher signal to specify the target category, but only needs to give correct or incorrect feedback on the completion of this classification task. That is, given an input sample, calculate its output class, compare it with known classes, and improve the performance of the classifier based on the difference.