All articles

Data Science, Machine Learning, Deep Learning, Statistics and AI – What’s The Difference?

October 20, 2020

In our time, the amount of data that is created every year exceeds the amount of data generated throughout all of the previous centuries. Humanity has been gathering and analyzing data from its surroundings since the dawn of civilization. Nowadays, however, we not only gather and analyze data with the help of technology, but we can also produce and access tremendous volumes of it, and use it for various purposes, such as making business decisions and even saving people’s lives. 

The problems of gathering, preprocessing and “making sense” of data are tackled by data science – an unusual mixture of business, mathematics and computer science. In this article, we’re going to explain what data science is and unravel the differences between the terms that are usually used interchangeably yet don’t mean the same: data science, statistics, machine and deep learning.

Data Science 

There’s a famous saying that a data scientist is someone who “knows statistics better than a programmer and knows programming better than a statistician”. While it is true to a certain extent, data science is much more manifold than only statistics mixed with software engineering. 

Data science is a field of study that combines domain expertise, programming skills, and knowledge of mathematics and statistics. Its main objective is to retrieve meaningful insights from data and make decisions based on it.

A typical data science project consists of such parts as:

  • Gathering data
  • Data wrangling/munging
  • Exploratory data analysis
  • Feature engineering
  • Predictive modeling (different AI methods, such as specific ML/DL algorithms)
  • Presenting the results

Two data scientists working on the same project can do different things and possess diverse expertise, yet still be considered to be data scientists. For example:

  • The first one might be an expert in experimental design, forecasting, modeling, statistical inference, or other things related to statistics, yet have little experience in coding. 
  • The second one can also have some background in statistics, yet be experienced in software development.  

In addition to that, data scientists are often actively participating in decision- and strategy-making, what makes it necessary for them to understand (to a certain extent) the business domain they are working in. 

It’s no surprise that this “jack-of-all-trades” role is often mistaken for some narrower roles from fields like machine learning and deep learning. However, all the fields mentioned above (i.e. machine learning, deep learning, AI, statistics) are just sources of various tools for a data scientist, and data science itself is a broad term. Let’s explore these tools in detail further below.

Statistics

Statistics is a purely mathematical discipline that encompasses various means of working with data, describing and evaluating it and its parameters. Statistics studies and describes data, while data science is more focused on making decisions based on the extracted knowledge, as well as creating prediction models. 

Statistics works with small amounts of data. Data science, on the other hand, often works with big data – using statistical methods only is not enough for its purposes. 

To sum it up, statistics comes in handy for data science and serves as one of its instruments, but other methods also exist.

Machine Learning

Retrieving knowledge and searching for some patterns in data can be tricky, hard, energy- and time-consuming. We did it with our own computer that Mother Nature has kindly gifted us with, i.e. our brain, for almost all of our history as species. A few centuries ago, we invented statistics and began working with data. And less than 100 years ago we started to develop algorithms that can extract patterns from data themselves – that’s when machine learning was born.

Machine learning is a branch of Artificial Intelligence. It focuses on algorithms that are able to search for patterns in data. Unlike “classical” algorithms, they’re not just a set of instructions that can be executed for obtaining the result. In comparison to the “classical” ones, machine learning algorithms are able to optimize their output in the process of doing some similar tasks, i.e. to “learn”. Such algorithms do not learn and understand patterns exactly the same way humans do (although they do it in a similar fashion) – but they can be very good at approximating them. This approach allows us to use them on data that we produce to approximate patterns, and store the acquired knowledge in the form of a predictive model.

Pure theoretical machine learning is pure mathematics: it is based on probability theory, statistics, linear algebra, calculus, etc. In its applied form, however, machine learning is more about creating and maintaining software that uses ML algorithms. Machine learning engineers might not deeply understand all the mathematical background behind the hood, but they definitely know how, when, where and what kinds of ML algorithms and methods they should use and apply. Their technical skill set usually consists of strong Python knowledge, experience in applying different ML, scientific and data-analytical libraries such as Sklearn, Pandas, NumPy, etc., and a background in maths and computer science.

If you would like to focus more on theoretical research and fully dive into mathematics, as well as create new algorithms and optimize existing ones, then ML researcher/AI scientist might be more suitable for you. For such a position it would be preferred for you to have a Masters or even a PhD degree. 

While a data scientist can fulfill the role of a MLE, he wouldn’t be very suitable for the role of a ML researcher, though it depends on the company. 

Deep Learning 

The data that surrounds us doesn’t always come in simple forms and shapes. A table of an online shop’s customers with a few columns is much easier to analyze than a few gigabytes of images, video, or text. Such multimedia requires its own means of preprocessing and gathering insights from. Fortunately, there is a “silver bullet” – a method that performs very well in such tasks and is even able to outperform humans (for example, in recognizing tumors on medical images) – neural networks. It is a type of a machine learning algorithm that is suitable for automatically extracting patterns from complex and highly-dimensional data such as digital images, video, text and other. 

The subfield of machine learning that works with no algorithm other than neural networks, its applications (including software ones) and architectures is called deep learning. The reason why this is a whole different subfield is not only because it’s good in dealing with certain types of data, but also because of the complexity of the algorithm itself. Designing a proper neural network architecture requires a profound understanding of the method, certain mathematical intuition and a bit of luck. To solve all the required tasks a deep learning engineer needs a solid mathematical background, Python or C++ knowledge, and also experience in applying certain deep learning frameworks such as Tensorflow or PyTorch.

Neural networks come with their own disadvantages: they are very slow to train and very expensive to use in terms of time and computing power. It is essential to mention that you probably don’t need to use them if you’re not dealing with multimedia or other high-dimensional data. Although they are not used for each and every machine learning task, their use cases are fascinating because they are closest to what a human can do: image and speech recognition and generation, natural language processing, etc.

The diagram above shows how AI, machine learning and deep learning are related to each other. 

To sum it up, deep learning is a more narrow subfield in comparison to machine learning. While machine learning is about the broader spectrum of algorithms and methods, deep learning mainly works with neural networks and highly-dimensional data, studying and developing new efficient architectures. 

Data is the oil of the 21-st century. The same way you cannot run a car without fuel, it is impossible to create knowledge without data. You can’t make something out of nothing.

Data science encompasses all the steps of creating and utilizing knowledge, starting from the process of gathering data and ending with making decisions based on predictive modeling. It’s a relatively new but very rapidly developing apparatus that found its way to the various industries and introduced tremendously successful applications that have already changed and will continue to change our world. 

by Stanislav Bychkovsky

;

Stanislav is a Machine Learning Engineer with a Bachelor’s Degree in Applied Mathematics and Computer Science. Passionate about Deep Learning, Neural Networks and AI, firmly believes in their ability to make the world a better place.