Press the play button to watch the video
Remember! if you experience accessibility issues with any of our rich media, you can download the accompanying textbook at the start of each module
The central focus of machine learning (ML) is on the development and implementation of computer systems engineered not just to follow instructions, but to actively learn from data. This capability for learning allows software applications integrated with ML techniques to adapt and progressively enhance their operational performance over time as they encounter and process more information, moving beyond the limitations of static, pre-programmed logic.
The technology underpinning machine learning consists of specialised algorithms. These algorithms are meticulously trained to discern and understand intricate relationships, trends, and patterns hidden within datasets. By using historical data as their primary input, these sophisticated algorithms can execute a wide array of valuable tasks. They possess the ability to generate predictions about future events or outcomes, classify diverse pieces of information into relevant categories, group similar data points together through clustering techniques, simplify complex datasets by reducing their dimensionality (the number of variables considered), and even engage in generative tasks, creating entirely new content such as text, images, or music that mimics the patterns learned from the training data.
The practical uses of machine learning are wide and varied, and it has been deployed in numerous sectors of modern life. In the financial industry, for example, ML algorithms play a pivotal role in safeguarding against fraudulent activities. Banks and financial institutions leverage these algorithms to scrutinise enormous streams of transaction data in real-time, identifying suspicious patterns indicative of fraud with a speed and level of accuracy that is simply unattainable through manual human review. Within the healthcare domain, machine learning serves as a powerful assistive tool for medical professionals. It helps clinicians in diagnosing diseases by analysing complex medical imagery like X-rays, CT scans, or MRIs, often detecting subtle anomalies missed by the human eye. Furthermore, ML models can predict patient outcomes based on historical data and individual patient characteristics, thereby informing more personalised and potentially effective treatment strategies. The retail sector also makes significant use of ML, using it to create highly personalised shopping experiences tailored to individual consumer preferences, predict inventory needs with greater accuracy to avoid stockouts or overstocking, and to optimise complex global supply chains for enhanced efficiency and cost reduction.
Beyond these specific industry applications, machine learning tackles data-intensive tasks that are beyond the scope of human execution due to their sheer scale. Consider the vast quantities of data generated daily by smartphones, sensors, websites, and countless other digital sources. ML provides the necessary tools to process, analyse, and extract meaningful insights from these vast data repositories. This capacity to unlock hidden patterns and actionable knowledge from large-scale data has evolved into a significant competitive advantage in diverse fields, ranging from quantitative finance and high-frequency trading to cutting-edge scientific research and discovery.
Many of today’s most influential technology companies, including giants like Meta (formerly Facebook), Google, and Uber, have deeply embedded machine learning into their core operations. They use it extensively to inform strategic decision-making, automate processes, improve operational efficiency, and ultimately deliver better products and services.
The relentless growth in data generation over recent years underscores the increasing necessity of machine learning. Making sense of this ever-expanding quantity of information requires the analytical power that ML provides. Conversely, this very abundance of human-generated data serves as invaluable fuel for training and continually refining machine-learning models. This creates a powerful, self-reinforcing cycle: more data leads to better-trained models, which in turn lead to more advanced AI systems capable of generating further insights or facilitating new data collection methods. This continuous learning loop is a fundamental principle underpinning many of today’s most sophisticated AI systems, driving progress with profound implications across virtually all aspects of life and work.
ML may be categorised based on how an algorithm learns to improve its predictions. There are four fundamental types of machine learning:
| Supervised Learning | This method involves providing the algorithm with a training dataset that is fully labelled. This means each data point in the training set has a known, correct output or category associated with it. The algorithm’s task is to learn the mapping function between the input variables and the specified output variables. Historically, supervised learning was the most common approach, though other methods are rapidly gaining prominence. Its applications include binary classification (dividing data into two groups), multiclass classification (choosing among several categories), regression modelling (predicting continuous numerical values), and ensemble modelling (combining multiple models for improved accuracy). |
| Unsupervised Learning | In contrast to supervised learning, unsupervised learning algorithms are presented with unlabelled data. Their objective is to independently discover hidden structures, patterns, or relationships within the data without prior guidance on what constitutes a “correct” output. Techniques like gradient descent are often employed to help group data points or identify underlying distributions. Many forms of deep learning, particularly those involving complex neural networks, fall under the umbrella of unsupervised learning. Common tasks include clustering (grouping similar data points), anomaly detection (identifying outliers or unusual data), association rule mining (finding frequently co-occurring items in a dataset), and dimensionality reduction (simplifying data by reducing the number of features). |
| Semi-supervised Learning | This approach represents a middle ground. The algorithm is trained on a dataset containing a mixture of labelled and unlabelled data, typically with only a small portion being labelled. The algorithm uses the limited labelled data to infer underlying structures and then applies this learned knowledge to the larger pool of unlabelled data. While training on fully labelled data generally yields higher accuracy, labelling can be extremely time-consuming and expensive. Semi-supervised learning aims to strike a balance, leveraging the performance benefits of labelled data while maintaining the efficiency associated with using primarily unlabelled data. However, insufficient labelled data can lead to overfitting, where the model performs well on the training data but poorly on new, unseen data. Applications include machine translation (learning language patterns from partial dictionaries), fraud detection (identifying fraud with few confirmed examples), and automated data labelling. |
| Reinforcement Learning | This paradigm operates differently, focusing on learning through trial and error within an environment. An algorithm (or ‘agent’) is programmed with a specific goal and a set of rules governing its actions. As the agent interacts with its environment, it receives positive rewards for actions that bring it closer to its goal and penalties or negative rewards for actions that move it further away. The agent’s objective is to learn a strategy or ‘policy’ that maximises its cumulative reward over time. Reinforcement learning is frequently employed in robotics (teaching robots physical tasks), game playing (training bots to master complex video games), and resource allocation optimisation within enterprises. |
The choice among these learning paradigms and specific algorithms is highly dependent on the nature of the problem being solved and the characteristics of the available data. Furthermore, many algorithms exhibit flexibility and can be adapted across different learning types, highlighting the dynamic and interconnected nature of the machine learning field.