It is often said that machine learning requires a strong background in statistics. While it is true that statistics is a must for ML engineers, but machine learning itself is not only about statistics.
A machine learning expert requires a basic understanding of statistics and that is because of the the data. Statistics is a branch of mathematics that deals with analysis, whereas ML is a subsidiary of computer science. Both contribute to data science at large.
Why Are Statistics And ML Thought To Be The Same
People often ignore the blurred borders between the two fields. One major reason is that both use data to solve problems and both have an application in mathematics. Also, both ML and statistics work on models to solve problems. There are many things that ML cannot do alone and has to rely on statistics. Both the spheres here deal with model and both deal with predictions, and are therefore under a constant debate.
ML and statistics have many similar terms, which is why a lot of people think it to be the same. Larry Wasserman, a Canadian Statistician explains in his blog:
The Fuzzy Line Between The Two
1. Rules: ML relies on learning from the past data with no strict rules. ML learns from data and does not have to rely on strict rules or standard programming practices like object oriented design. Most of the work is done on computing and so it provides strong predictive ability with minimal human efforts. Statistics, on the other hand, has a set of rules to follow and it does not involve any learning from data. It strictly follows the program algorithm and gives a result. Statistics does provide the best estimate but with a more human effort as it demands the understanding of the relation that the variable has on an equation.
2. Applications: We need to identify what factors to consider in case of statistics. But in ML models, we only need to have more data and the model learns from it. More the data, more will be the accuracy of the model. So, all that the model needs is data. ML works with large sets of data and its models are applied to high dimensional datasets. Statistical modelling is a lot about studying the relationships between variables in the form of mathematical equations for the purpose of predicting relevant outcomes. Statistical models are suitable to low dimensional datasets and it takes fewer attributes. ML is a very new subject compared to Statistics, which is a work originated in 1760s, and hence it has better applications as it is an advanced branch.
3. Assumptions: Statistics involves many assumptions. ML models do not need assumptions. It only learns from the input data. In ML fewer assumptions are made due to a better accuracy from the predictive models in comparison to statistical models which is more mathematically based.
4. Building block: ML is governed by algorithms and Statistical models are fundamentals of mathematical equations, which is obvious from the fact that the former is a branch of computer science and later that of Mathematics.
5. Result: One of the most fundamental difference between the two is that, Statistics gives us inferences based on the data that it receives, while ML gives us predictions based on the data.
John Rauser’s gave a very good example in his keynote Statistics Without the Agonising Pain. He gave a brief example to establish that if one can program a computer, he has direct access to the deepest, most fundamental ideas in Statistics. It in itself shows that the two fields are very closely related, but calling machine learning is all about glorified statistics would be completely incorrect.
It is indeed true that ML, in a lot of ways, is an inspiration from statistics but it is not all about statistics. It takes help from statistics right from understanding the data to modelling. But it has used statistics as its support, took all the best outcomes of it and and progressed its own way. Statistical models are needed to work through an ML predictive model. It provides the basis and compliments the ML models. Statisticians and ML engineers have different paths to travel to conquer their goals and even are suitable for different work a lot of times. But it is also true that the two worlds overlap in many applications and compliment each other in remarkable ways.
Try deep learning using MATLAB