What is the difference between Artificial Intelligence, Machine Learning, and Data Science?

In the global universal sense, AI is a very broad term. It includes both scientific theories and specific technological practices for developing programs that attempt to imitate human intelligence.

Machine Learning is a field of AI, that is actively applied in practice. Today, when it comes to using AI in business or manufacturing, most often we mean Machine Learning.

Normally, ML-algorithms work on the principle of self-learning mathematical model that performs analysis using a large amount of data, while conclusions are drawn without following rigidly defined rules.

The most common type of task in machine learning is learning with a teacher. To solve this kind of problem, training is used on an array of data for which the answer is known in advance (see below).

Data Science is the theoretical and practical approach to the analysis of large amounts of data using various mathematical methods, including machine learning, as well as solving related tasks related to the collection, storage, and processing of data arrays.

Data Scientists are data processing experts, who analyze data using Machine Learning as one of the methods.

Let’s consider the application of ML on the example of banking scoring. The Bank has data for existing customers. The Bank knows if someone is overdue on loan payments. The task is to determine whether a new potential customer will make payments on time. For each client, The Bank has a combination of certain traits/characteristics: gender, age, monthly income, occupation, place of residence, education, etc.

Among these characteristics can be poorly defined parameters, such as data from social networks or purchase history. Besides, the data can be enhanced with the information from external sources: currency exchange rates, credit score data, etc.

A computer sees any client as a set of characteristics X1, X2, X3, ..., Xn, where, for example, X1 – age, X2 – income, X3 – the number of photos of expensive purchases per month (in practice, as a part of a similar task, Data Scientist could use more than a hundred of such characteristics). Each client has another variable Y with two possible outcomes: 1 (there are late payments) or 0 (no late payments).

The totality of all these data X and Y is the Data Set. Using these data, Data Scientist creates a model F by selecting and developing the Machine Learning algorithm.

In this case, the analysis model looks like this:

Machine Learning algorithms imply that model responses approach to true answers (which are known in advance in the training Data Set) with multiple iterations. This is so-called training with a teacher with a particular data sample.

In practice, the computer learns only on a part of the data array (80%), using the remainder (20%) to verify the correctness of the selected algorithm. For example, a system can be trained on an array from which the data for a couple of regions are excluded, and the accuracy of the model is verified later using the data for these regions.

Now, when a new client arrives at the bank, whom the bank does not know Y yet, the system will assess the new client reliability based on the data available for the new client F (X1, X2, X3, ..., Xn) = Y.

However, learning with a teacher is not the only type of problem that ML can solve.

Also, ML-algorithms are applied to solve such tasks as support specialist communication modeling, or creating works of art indistinguishable from human creations (for example, neural networks paint pictures).

The new and popular area of application is reinforcement training, which takes place in a limited environment that evaluates the actions of participating agents (for example, using this algorithm, AlphaGo was created that defeated a human in the Go checker game).

Neural network is one of the Machine Learning methods. This algorithm was inspired by the structure of the human brain, which is based on neurons and the connections between them. In the learning process, the connections between neurons are adjusted in such a way as to minimize errors of the entire network.

A specific feature of neural networks is the architectures suitable for almost any data format: convolutional neural networks for analyzing pictures, recurrent neural networks for analyzing texts and sequences, auto-encoders for data compression, generative neural networks for creating new objects, etc.

At the same time, almost all neural networks have a significant limitation - a large amount of data is needed for their training (orders of magnitude greater than the number of connections between neurons in a given network). Since recently the volumes of data ready for analysis have grown significantly, the area of this method application is also growing. For example, today, neural networks solve such image recognition tasks as determining the age and the gender of a person from a video or determining if a worker wears a helmet.

Result Interpretation is the Data Science field, which explains the reasons for choosing one or another solution by the ML model.

There are two main areas of research in this field:

• Study of the model as a black box. Analyzing the provided examples, the algorithm compares the features of these examples and the results of the algorithm, making conclusions about the priority of the provided features. The black box is usually used for neural networks.
• Study the properties of the model itself. The study of the characteristics that the model uses to determine the degree of their importance. The method is most often applied to algorithms based on the decision tree method.

For example, when forecasting defects in a production line, the characteristics of objects are machine setting data, raw material chemical composition, sensor readings, video from the conveyor, etc. And the question to answer is whether the defective product will be produced or not.

Naturally, a company is interested not only in the forecast of the defective product production but also in the interpretation of the results, i.e., the reasons for the defects and their subsequent elimination. The reasons could be long periods between machine maintenance, the quality of the raw materials, or simply abnormal readings of some sensors, which the technologist should pay attention to.

Therefore, in the framework of the production defect forecast, it is not enough just to create an ML-model, but the interpretation of the model results has to be performed to identify factors that affect production defects.

Machine Learning is applied when there is a large set of statistical data, but it is impossible or very laborious to find correlations using expert or classical mathematical methods.

So, if there are more than a thousand parameters at the input (which include numerical and text entries, as well as video, audio, and pictures), then it is impossible to determine the correlation between the results and the input data without a machine.

For example, in addition to the substances themselves entering into the interaction, a chemical reaction is affected by many parameters: temperature, humidity, the material of the container in which it occurs, etc.

It is difficult for a chemist to take into account all these parameters to accurately calculate the reaction time. Most likely, he will take into account several key parameters and will base his estimations on his experience. At the same time, based on the data from previous reactions, the Machine Learning algorithm will be able to take into account all applicable parameters and give a more accurate estimation.