Types of machine learning and how they work

Machine Learning

Machine Learning is a subfield of Artificial Intelligence, with the goal of developing algorithms and models that enable machines to learn from data and make predictions without explicitly programming their logic.

How does machine learning differ from regular programming?

In traditional programming, the developer (programmer) explicitly maps out, for example, the business logic based on which a decision is to be made. That means the programmer receives a business requirement (input) and writes a program that does exactly what the requirement specifies.

In the case of artificial intelligence, we do not predefine the criteria for making a business decision. Instead, we provide data to a learning algorithm, which then learns the underlying business logic from this data. Therefore, the AI developer (Data Scientist) uses data as input, and the result of this process is a trained model."

What are the types of machine learning?

Supervised learning

During the learning phase, the algorithm receives data for which the class of each observation is known (target variable). For example, whether a banking transaction was fraudulent or not. This allows the algorithm to learn how to infer the value of the target variable from the other available data (input variables).

In business applications, this is one of the most commonly used methods. Its significant advantage is that we can test the trained model’s accuracy using historical data.

The most commonly used methods within supervised learning are:

- Regression: Our goal is to estimate a continuous numerical value, such as predicting a stock price.

- Classification: Our goal is to categorize observations into different classes, such as:

- "Is a particular customer interested in a specific service or not?"

- "Will a particular medication be effective for a specific disease or not?"

Fine tune

In this case, we are building upon models that have already undergone extensive training, often by multiple sources. A prime example is the field of image recognition, where numerous pre-trained models exist with a high degree of proficiency in object identification. These models, however, can be fine-tuned and further optimized to meet more specialized or nuanced requirements.

Unsupervised learning

One of the most commonly applied examples of this learning method in the business world is clustering, also known as grouping (for instance, segmenting customers for marketing purposes). In this process, the algorithm itself attempts to group the data so that similar customers are placed in the same cluster, while distinct customers are categorized into different clusters.

Reinforcement learning

In reinforcement learning, the algorithm continuously receives new data, which it integrates into its knowledge, effectively 'improving' itself. This method is frequently used in robotics development and in the creation of large language models (LLMs)

It is essential to highlight recommendation engines, which typically integrate multiple learning methods from those listed above to implement a sophisticated and comprehensive recommendation logic.

How do we know that the AI has learned the task correctly?

One of the most crucial aspects of training artificial intelligence models is the verification and validation of the trained model. Broadly speaking, this process consists of two key components:

Statistical validation: This involves using various statistical methods to assess how accurately the model performs its tasks. Depending on the type of problem the model has been trained to solve, different methods must be applied to evaluate the accuracy of its performance.

Business validation: Naturally, not everything can be verified with statistics alone, making business-oriented validation critically important. During this phase, we seek answers to questions such as, 'What impact will incorporating the developed model into my decision-making process have on my overall business operations?

What are the main steps in the development of AI model?

Developing AI models can fundamentally be divided into two main parts:

1) Development of Learning Algorithms: This is more research-oriented work, where someone tries to create a completely new approach in the field of machine learning. It could involve developing an entirely new type of algorithm or creating a novel variant of an existing approach. This is less common in the business world.

2) Application of Existing Learning Algorithms: This involves using and fine-tuning an existing learning algorithm with a dataset that is relevant to a specific business problem. The result is a trained model that can be applied to address the particular business issue

Main Steps in Training Machine Learning Models:

1. Problem definition
2. Data collection and data processing
3. Creating training and test datasets
4. Selecting a Machine Learning algorithm
5. Training the model
6. Evaluation of the model
7. Model deployment and monitoring

In what business challenges can machine learning be successfully applied?

Without claiming to be exhaustive, here are some application areas:

- Healthcare: Supporting diagnosis, prognosis, and the development of treatment plans. Assisting in the evaluation of X-ray, CT, and MRI images.

- Finance: Fraud detection, credit application evaluation systems, algorithmic trading.

- Telecommunication: Cross-selling, network analysis, fraud detection.

- Retail: Personalized marketing, pricing, inventory planning.

- Manufacturing: Improving efficiency, quality assurance, failure prediction.

Overall, it can be said that in nearly every industry, there are areas where machine learning can significantly aid in optimizing processes.

What tools can we use for developing AI models?

For developing machine learning models, Python is currently the most widely used programming language. Almost every major provider—whether in the cloud or on-premises—offers an environment where ML models can be developed using Python.

In addition, there are now numerous tools on the market that allow for model development through graphical interfaces based on low-code / no-code principles: