Machine Learning: Simply Explained

What is Machine Learning?

Machine learning is a subfield of artificial intelligence and computer science that focuses on creating systems that learn and improve in the absence of explicit rules. So whereas traditional software is typically programmed to perform specific steps with explicit instructions (e.g., with loops, if/else statements, etc.), machine learning systems learn how to perform well and make decisions.

Importantly, machine learning systems learn how to improve their performance by exposure to data examples. This makes machine learning a highly data driven sub-discipline of computing, since data is a key input to machine learning systems, and is central to enhancing their performance. This ability to learn from data enables machine learning systems to solve a wide range of complex problems that are nearly impossible to solve with traditional programming, from forecasting and predicting customer behavior to facial recognition and self-driving cars.

It's important to highlight, however, that although machine learning is closely related to AI, there are differences. AI is a larger field that extends beyond machine learning. Said differently, machine learning is one approach to building intelligent systems.

Machine Learning vs Artificial Intelligence

Many people use the terms "artificial intelligence" (AI) and "machine learning" (ML) interchangeably, but there are important differences.

AI is a broad field that focuses on creating systems with abilities that we typically associate with human intelligence, like decision-making, problem-solving, and reasoning. Therefore, the field of AI includes machine learning, but also extends beyond it, since there are other methods for building "intelligent" systems, like rule-based systems, expert systems, and symbolic reasoning. 

Machine learning, then, is best understood as a sub-discipline of AI that focuses on creating and using algorithms that improve with exposure to data.  Machine learning is one particular approach to AI among many. 

It's worth noting, however, that currently in 2024, machine learning is the most prominent and arguably the most successful approach to building AI systems, since the most popular AI systems at present are machine-learning based AI systems.

The fact that our current AI systems are machine-learning based systems has caused many people and media organizations to blur the terms.  This imprecision in terms arises because generative AI systems have become the dominant type of AI system, so many people believe – mistakenly – that all AI is generative AI, or that generative AI systems can solve all types of problems, when in fact, we often use other types of machine learning systems.

Ultimately, one needs to remember that:

• AI is a broad field that extends beyond machine learning
• Machine learning is one approach to building intelligent AI systems
• Generative AI is one specific type of AI, built largely with machine learning
• Other types of machine learning are better than generative AI for solving certain types of problems

Approaches

Although we can talk about machine learning at a high level as a single discipline, there are actually different "approaches" that we use when building machine learning systems.  These are sometimes called machine learning "paradigms."

Three of the most important machine learning approaches are:

• supervised learning
• unsupervised learning
• reinforcement learning

Supervised Learning

In supervised learning, we train a model on a dataset that has labels. In this approach, we have a dataset that includes both input data and the expected output.

By explicitly providing a machine learning algorithms with pairs of inputs and outputs, the algorithms can "learn" the relationship between typical inputs and outputs. Said differently, the algorithm learns to translate inputs into appropriate outputs. This machine learning approach is well-suited to tasks like classification, where we predict a categorical label like "spam" in an email spam classification task. We also use supervised learning for regression, where we predict a numeric value (like housing prices). Supervised learning requires labeled training data (i.e., data with both inputs and outputs), and is one of the most commonly used approaches in practice.

Unsupervised Learning

In contrast to supervised learning (where we have a labeled dataset), in unsupervised learning we have unlabeled data. Said differently, the data that we use for unsupervised learning only has "input" data, but it lacks a supervising output variable (like a categorical label) that it's trying to predict.

Unsupervised learning operates on the input data, and tries to discover patterns or underlying structure in the data. By discovering hidden structure in the data, unsupervised learning often helps us gain insight into the nature of the data. It can also help us with certain types of data processing and transformation as a precursor to other types of machine learning (e.g., we can often use some types of unsupervised learning on a dataset to transform it, prior to performing supervised learning).

Common algorithms that fall under this ML approach are clustering (which helps us discover segments or groups in a dataset), as well as various dimension reduction techniques like principal component analysis (which helps simplify a dataset, often in preparation for other techniques like visualization or supervised learning).

Reinforcement Learning

Inspired by behavioral psychology, reinforcement learning is a learning approach where a system learns to make decisions by interacting with an environment and getting rewards.

As the system (often called an "agent") interacts with an environment, the system receives rewards and penalties for its actions.  These rewards and penalties are a form of feedback, which enables the system to refine and improve its strategies over time.

This machine learning approach is often applied for problems in game playing (e.g., chess), as well as robotics and autonomous vehicle driving.

Other Approaches

Beyond supervised, unsupervised, and reinforcement learning, there are other approaches.

These additional machine learning approaches include:

• semi-supervised learning, which uses a mixture of labeled and unlabeled data examples
• self-supervised learning, which generates new labels from the data, and then learns how to apply those labels

Common Machine Learning Algorithms

In machine learning, we use a wide range of algorithms, and different algorithms are well-suited to different tasks and types of data.

Linear Regression

Linear regression is arguably one of the most common algorithms that we use in machine learning.

Linear regression predicts numerical values by learning to predict a linear relationship between the input data and the target variable. Effectively, it works by fitting a line (or in higher dimensional data, a hyperplane) to the data in a way that minimizes the error between the predicted values and the actual values.

Logistic Regression

Logistic regression is somewhat similar (and related) to linear regression, but it performs classification. That is, we use logistic regression to predict categorical labels based on input data.

Logistic regression works by using the logistic function to model the probability that the output class is a 0 or 1. Thus, we often use it for binary classification, although there are also ways to apply logistic regression for more than two class labels.

We can use logistic regression for a variety of specific tasks, like spam prediction, disease diagnosis, churn prediction, and a lot more.

Decision Trees

Decision trees make predictions by splitting the data into subsets based on feature values. 

Essentially, a decision tree makes a tree-like path with branches (e.g., "if variable A is greater than 1, go down this path, otherwise go down the other path"), and after multiple branches, the tree makes an output prediction. Decision trees are often easy to understand, easy to visualize, and easy to explain. Moreover, we can use them for both regression and classification tasks. This makes them very useful for a variety of tasks.

However, decision trees do have some weaknesses, such as a tendency to overfit the data (unless overfitting is mitigated with techniques like regularization, random forests, etc.).

Neural Networks

Inspired by networks of neurons in the human brain, the neural network algorithm builds networks of artificial "neurons" that link together to process input data. We create neural networks by combining large numbers of artificial neurons in specific architectures (often in layers), in a way that enables them to predict outputs and identify patterns.

Neural networks often require very large amounts of training data.  But, when trained properly, they work exceptionally well for complex tasks like image recognition, natural language processing and more.

K-Means Clustering

K-means clustering is one of the most common unsupervised learning algorithms. We commonly use k-means to cluster (AKA, segment) data into groups based on the similarity of data points.

To do this, k-means assigns data points to clusters by computing how close every data point is to a set of cluster "centroids". The centroids are initially set randomly in the data space, but are then recomputed as the center of the data points in the cluster.  So k-means iteratively assigns data points to a cluster, recomputes the cluster centroid, and then re-assigns data points to a new cluster based on the new nearest centroid.

K-means is a powerful technique that we frequently use in market segmentation, as well as anomaly detection.

Other Algorithms

Although the algorithms listed above are some of the most commonly used tools in machine learning, there are dozens (arguably hundreds) more algorithms. A large part of doing machine learning work is understanding the strengths and weaknesses of different algorithms, and knowing which algorithm to use for which problem.

Applications of Machine Learning

Machine learning is a very versatile set of tools that have transformed a variety of industries.

Healthcare

In a healthcare setting, doctors and healthcare providers can use machine learning systems to detect disease and create personalized treatment plans. More specifically, we can use classification systems to analyze and detect diseases in medical images, as well as predict diseases more generally beyond medical imaging.

In turn, these sorts of applications can improve diagnostic accuracy and lead to more tailored care. Ultimately, using machine learning in healthcare can decrease costs and lead to better patient outcomes.

Finance

In finance and banking, we can use machine learning systems for credit scoring, fraud detection, and algorithmic trading.

So for example, we can use regression models to predict creditworthiness or use various algorithms to detect fraudulent transactions (e.g., classification models). Ultimately, using machine learning systems can decrease financial risk and increase profits.

Retail and E-commerce

In retail and e-commerce, we can use machine learning to build recommendation systems that analyze customer purchase and browsing behavior to predict new products that the customer will like.  This often increases customer satisfaction, as well as revenue.

Operationally, you can also use machine learning to optimize inventory and logistics, resulting in operational efficiency.

Marketing

In marketing, you can use machine learning in a variety of ways to improve both marketing strategy and tactics.

Strategically, you can use clustering techniques for customer segmentation, which enables businesses to refine product, messaging, and channel engagement strategies. Tactically, you can use ML for things like churn prediction, by building a classification model that predicts customers who will churn.  This in turn enables businesses to intervene and therefore decrease churn.

These strategic and tactical uses of machine learning for marketing often lead to improved retention, larger cart values, higher customer satisfaction, and ultimately higher revenue and ROI. 

Autonomous Vehicles

In autonomous vehicles, we can use advanced deep learning techniques to build cars that drive themselves. We do this by collecting data from large sensory systems, and then use ML algorithms (e.g., deep learning based computer vision systems) to identify objects, predict behavior of other drivers, and plan good routes.

Using machine learning for self-driving cars in this way promises lower safety risk and decreases in driving errors.

Natural Language Processing

Perhaps the most spectacular use of machine learning in recent years has been in natural language processing, where the use of specialized architectures (i.e., the Transformer architecture), has led to ML-powered systems like ChatGPT.

These natural language ML systems can perform question answering, translation, sentiment analysis, and more. Because of their flexibility, we can use these systems in areas that promise to revolutionize business and industry.

Conclusion

Machine learning is a toolkit for building software systems that learn from data, and it has become a powerful technology for solving a variety of problems in business, industry, and beyond.

Moreover, given the current boom in machine learning and AI, we're likely still only scratching the surface. As data becomes more abundant and compute hardware becomes more powerful, we're likely to unlock even more abilities with machine learning. ML will likely play a central role in technology, but also society more broadly.

That said, you should consider learning or implementing machine learning in your own business or organization. As machine learning systems become more powerful, the opportunities will be vast.