Machine Learning, a heavy industry trend 4.0 with growing applications (data analysis, image processing, Analytics, etc.), now mainly uses the Python language.
We have already covered AI (Artificial Intelligence) in previous articles on MetalBlog by reviewing the different technologies (expert system, digital twin, Deep Learning) and giving application cases. In this article, we will detail the different methods of Machine Learning (ML) and the contribution of the Python language by illustrating certain points on the basis of a data file.
The importance of data:
Numerical simulation versus Machine Learning and the importance of data. In Machine Learning, the quality and completeness of data is essential for learning. IT professionals illustrate this with the formula “Garbage in, Garbage Out” which means that if the input data is corrupted, false, … If the calculation is incomplete, the result will also be marred by a potentially significant error.
This is even more true in the field of ML where model learning is carried out entirely from data and no longer from physical laws as in conventional numerical simulation (thermal, fluidic, etc.).
Preparation and cleaning of data known as “data-cleaning” (removal of blank boxes, invalid data, etc. ) and their visualization thus becomes a stage in itself before the data processing itself. This step can take up to 80% of the working time.
On large files, many boxes are empty (not filled in). Several strategies can be applied; delete the data line, impute the missing value by the average of the values of the variable or, for time series, impute the missing value by linear interpolation (between previous and next value).
Linear regression, the simplest tool in Machine Learning:
Linear regression between variables is the simplest tool. This is Machine Learning, since from existing data, we will model a behavior that will be used to predict data not yet known on this regression line.
Data labelled and not labelled: Regression and classification:
In ML, there are two main types of data: labeled and unlabelled data. They differ by the fact that a labeled data is tagged («label») by a significant value or quantity, in general, the result to be predicted. Labeled data will make it possible to make supervised learning (regression and classification) and non labeled data of unsupervised learning such as Clustering (identification of population with similar characteristics). regression algorithms will be applicable (linear regression, KNN, etc.). In the engineering sciences, where a lot of numerical data is available, the Supervised Machine Learning and regression algorithms are mostly used.
The different algorithms of Machine Learning:
Supervised learning is by far the most popular and widespread. It requires labeled data. The supervised ML algorithms are essentially six large families: linear regression, the neighbouring k-nearest method (or KNN for K-Nearest-Neighbors), the Naïves Bayes method, decision trees (Decision Tree), support vector machines (or SVM for Support Vector Machines) and finally neural networks. These methods each have their advantages and limitations.
If neural networks are the most powerful in terms of prediction, they are also the most complex to use and require much more computing power in the training phase, requiring graphics cards or running on the cloud in some cases.
We will come back to neural networks in a future article.
The Python language for Machine Learning:
Two languages have gradually become established in data science, the R code and the Python. Created in 1989, the open-source code Python now largely dominates computing because while being simple, it is relatively powerful and has more than many libraries that allow calling pre-programmed functions.
The Python developer community is very active. You can also find many codes (or parts of code) in open-source, tutorials in French, explanatory videos (YouTube) and many online help on this language.
Python as a language for controlling commercial codes:
In addition, there are Python interfaces for some non open-source commercial codes. ThermoCalc (thermodynamic prediction for alloys) has a TC-Python interface that allows to develop much more elaborate commands than «Console» mode standard and will search for information directly in ThermoCalc alloy databases, bringing more flexibility in the use of ThermoCalc.
The Matlab software also allows, for example, to integrate programs developed in Python.
keywords: machine learning, machine learning is, python machine learning,machine learning modeling, andrew ng machine learning , ai learning , aws machine learning, supervised learning ,unsupervised learning, ai ml, deep learning ai, tensorflow, data analytics, master's in data science, online master's data science, data analytics degrees, data science degrees, certified data scientist, master's in data analytics online , ms in data science, datascience berkeley ,uc berkeley data science, data science for managers, data science for beginners, certified data scientist, data science for all, big data analyst, r for data science, pandas, keras,tensorflowjs,hands on machine learning.
crossorigin="anonymous">