Click on " Xiaobai Learning Vision " above , and choose to add " Star " or " Top "
重磅干货，第一时间送达
Editor's Recommendation
All machine learning models are best fits, and in this sense, all models are wrong. However, it is not to say that a model is worthless just because it is wrong. Beginners often equate the accuracy of the model with the value of the model, but this is not the case.
Reprinted from丨Big Data Analysis and Mining
It was raining during the May Day holiday, and I couldn’t go anywhere because of the epidemic, so I went online at home. I read an article "When Not to Use Neural Networks" on towards-data-science. I feel that there are several viewpoints that are not bad. The translation excerpts are as follows.
Not long ago, a friend asked me about how to choose an ML/AI model. At that time, most of the ML/AI model knowledge he accepted were neural networks or their variants. In his cognition, neural networks were powerful enough to solve any problem. Whenever, choose neural networks Always the right choice.
1. All models are wrong, but some are useful
Anderson, a pioneering modeling physicist, said at the 1977 Nobel Prize award ceremony: The art of modeling is to remove the irrelevant parts of reality, and both modelers and users face certain risks. Modelers may miss crucial factors; users may ignore that the model is only approximate, intended to reveal a certain possibility, and understand and use the specific result samples of experiments or calculations too bluntly.
All machine learning models are best fits, and in this sense, all models are wrong. However, it is not to say that a model is worthless just because it is wrong.
Beginners often equate the accuracy of the model with the value of the model. but it is not the truth. Towards-data-science recommends more than 20 articles to me every day, and only one or two of them may be of real interest to me and I will read them. The accuracy rate is equivalent to only 10%. But this is enough, this 10% can guarantee my visit to the website every day, and the recommendation has guaranteed the traffic value in business. Accuracy is only an estimate of one part of a model's value.
2. The value of the model
The value of a model is closely related to these four aspects.
(1) Interpretability: Can explain how the model solves the problem step by step
(2) Explainability: Can explain the physical meaning behind the model output
(3) Flexibility: the ability of the model to describe complex things
(4) Complexity: the cost of model training and reasoning
With a decision tree model, it is easy for us to understand how it reasoned to get the final result (Interpretability), and it is easy for us to make business decisions based on the output of the model (Explainability). The deeper the decision tree, the better the ability to describe complex things ( Flexibility), but the cost is higher (Complexity)
A neural network model is weak in Interpretability and Explainability, strong in Flexibility, and usually not low in Complexity.
The value of a model is for specific problems. Some problems are intuitively obvious but difficult to model, such as face recognition. The problem-solving model should focus on Flexibility and Complexity. At this time, the neural network is more suitable, and there is no need to provide too strong Interpretability and Explainability (both of which are the weaknesses of the neural network model)
Other questions, such as credit approval, are less intuitive but logical. The problem-solving model should focus on Interpretability and Explainability.
3. Simple models solve complex problems, and complex models solve simple problems
(1) Simple model: Simple models include some very traditional machine learning models, such as linear classifiers, decision trees, and K-nearest neighbors. Even if you write your own code to implement it, you don't need too advanced mathematical knowledge.
(2) Complex problems: Complex problems are problems that humans need to receive professional training and training, and require human thinking to solve. such as credit approval
(3) Complex models: Complex models include some models that require complex numerical calculations and mathematical theoretical proofs, such as SVM, neural networks, kernel methods, and gradient boosting methods. It is generally difficult to write code to achieve it yourself.
(4) Simple questions: Simple questions are those that are in line with human intuition and that we do naturally every day. For example, distinguish a cat from a dog, recognize a person, and read a paragraph of text.
Machines are often bad at what humans are good at. Conversely, humans are not necessarily good at what machines are good at. This is easy to understand. So let simple models solve complex problems, and complex models solve simple problems.
4. Model complexity VS interpretability
The more complex the model, the less interpretable it is.
(1) Low-level complexity models: linear regression, logistic regression, decision tree, K-nearest neighbor.
(2) Moderate complexity models: boosting trees and forests, naive Bayesian, Gaussian process
(3) Advanced complexity models: SVM, neural network, Bayesian method
In practical applications, it depends on whether accuracy is the priority or interpretability is the priority.
This article is only for academic sharing. If there is any infringement, please contact to delete the article.
下载1：OpenCV-Contrib扩展模块中文版教程
在「小白学视觉」公众号后台回复：扩展模块中文教程，即可下载全网第一份OpenCV扩展模块教程中文版，涵盖扩展模块安装、SFM算法、立体视觉、目标跟踪、生物视觉、超分辨率处理等二十多章内容。
下载2：Python视觉实战项目52讲
在「小白学视觉」公众号后台回复：Python视觉实战项目，即可下载包括图像分割、口罩检测、车道线检测、车辆计数、添加眼线、车牌识别、字符识别、情绪检测、文本内容提取、面部识别等31个视觉实战项目，助力快速学校计算机视觉。
下载3：OpenCV实战项目20讲
在「小白学视觉」公众号后台回复：OpenCV实战项目20讲，即可下载含有20个基于OpenCV实现20个实战项目，实现OpenCV学习进阶。
交流群
欢迎加入公众号读者群一起和同行交流，目前有SLAM、三维视觉、传感器、自动驾驶、计算摄影、检测、分割、识别、医学影像、GAN、算法竞赛等微信群（以后会逐渐细分），请扫描下面微信号加群，备注：”昵称+学校/公司+研究方向“，例如：”张三 + 上海交大 + 视觉SLAM“。请按照格式备注，否则不予通过。添加成功后会根据研究方向邀请进入相关微信群。请勿在群内发送广告，否则会请出群，谢谢理解~