Charity Digital – Topics – The ultimate guide to machine learning
Machine learning picked the TV show you watched last night. It likely picked the music that you’re currently playing. It almost certainly led you to the current article you’re reading. All media recommendations, in fact, are based on machine learning and most uses of artificial intelligence (AI) involve machine learning in some form.
As explained by MIT Sloan professor, Thomas W Malone: “In the last five or 10 years, machine learning has become a critical way, arguably the most important way, most parts of AI are done.” Most advances in AI, including generative AI, depend on elements of machine learning.
With the growing ubiquity of machine learning, and misinterpretations that follow confusing terminology, we thought we’d write an article that makes everything as simple as possible. So, with that in mind, let’s start from the top, with definitions of AI and machine learning.
Skip to: What is artificial intelligence?
Skip to: The different branches of AI
Skip to: The definition of machine learning
Skip to: How machine learning actually works
Skip to: Real-world examples of machine learning
What is artificial intelligence?
To define ML, you need to first define AI.
AI works by using iterative, fast processing, and intelligent algorithms, married with huge amount of data. The tech learns automatically from patterns or features of the data and uses that information to improve processing and algorithms.
AI acts as a simulation of human intelligence in machines that are programmed to think like humans. Indeed, AI refers to any machine that exhibits traits associated with a human mind, such as learning and problem-solving.
The different branches of artificial intelligence
There are myriad different branches of AI. Neural networks, for example, effectively learn through external inputs, relaying information between each unit of input. Neural networks are made up of interconnected units, which allow repeat processes to find connections and derive meaning from previously meaningless data.
Neural networks are a form of machine that takes inspiration from the workings of the human brain. Examples of a neural network include sales forecasting, industrial process control, customer research, data validation, and even targeted marketing.
Deep learning uses extensive neural networks with various layers of processing units. Deep learning utilises the vast advances of computing power and training techniques to learn complicated patterns, employing massive data sets. Face ID authentication is an often-cited example of deep learning, with biometric tech employing a deep learning framework to detect features from users’ faces and match them with previous records.
Natural learning processing is a commonly used for of AI. Natural learning processing relies on the ability of computers to analyse, understand, and generate human language – particularly around speech. The most common form is chatbots. Natural learning processing, at more evolved stages, allows humans to communicate with computers using normal language and ask them to perform certain tasks.
Expert systems use AI to mimic the behaviour of humans or organisations that possess specific knowledge or experience. Expert systems are not designed to replace particular roles but assist complex decisions. Expert systems aid decision-making processes by using data, in-depth knowledge, alongside facts and heuristics.
Expert systems are typically employed in technical vocations, such as science, mechanics, mathematics, and medicine. They are used to identify cancer in early stages, for example, or to alert dentists to unknown organic molecules.
Fuzzy logic is a rule-based system that aids decision-making. Fuzzy logic uses data, experience, and knowledge to advance decision-making and assess how true something might be on a scale of 0-1. Fuzzy logic answers a question with a number, such as 0.4 or 0.8, and aims to overcome the binary human response of true and false and give degrees of truth over vague concepts.
The application of fuzzy logic appears in low-level machines to perform simplistic tasks, such as controlling exposure in cameras and defining the timing of washing machines.
These are the main areas of AI, especially in relation to machine learning. That might seem like a lot to take in, and the definitions can be hard to absorb. That’s why we created an effective glossary, which you can access here: A glossary of artificial intelligence terms and definitions.
The definition of machine learning
All of the above branches of AI all closely relate to machine learning. Neural networks are a type of machine learning process. Deep learning is a subset of machine learning. Natural learning processing combines machine learning models with computational linguistics. Expert systems provides a different model to machine learning, with a stricter set of rules. Fuzzy logic is a method of machine learning that has been developed to extract patterns from a data set.
So machine learning is pivotal to understanding so many areas of AI. But what is machine learning? Machine learning was defined in the 1950s by Arthur Samuel: “The field of study that gives computers the ability to learn without being explicitly programmed.” The absence of explicit rules, or the expectation that the rules will evolve, is the core element of machine learning. Machine learning asks computers to program themselves.
It starts, as everything concerning AI starts, with data. Machine learning focusses on using that data and complex algorithms to allow AI systems to imitate the way humans learn, gradually improving accuracy. As ever with AI, the more data, the better the results. Machine learning is an analytic model, which allows software applications to become more accurate at predicting outcomes.
Machine learning, as highlighted by MIT Sloan, can be descriptive (explain what happened), predictive (predict what happens), and prescriptive (explain how to make something happen).
There are three subcategories of machine learning. The first is supervised machine learning, which refers to models trained on labelled data sets. These grow more accurate over time, though AI drift remains a concern. Supervised models are the most common form of machine learning.
The second is unsupervised machine learning. Unsupervised models look for patterns in data that are specially not labelled. Unsupervised models find patterns that people are not looking to find, which can provide unexpected insight. Unsupervised models are often used in sales and marketing to find opportunities that have been missed, or to provide options for engagement.
The third, and final, is reinforcement machine learning. Such models rely on trial and error, allowing the models to grow through establishing a reward system. Reinforcement learning can train autonomous vehicles, for example, by alerted them to the right, and wrong, decisions. Reinforcement models are built on the basic premise of positive/negative reinforcement.
The above offers an overview of machine learning, providing you with an overarching glimpse of how it works. But how does it actually work? What are the technical processes that inform machine learning? We cover that next, so it’s about to get a bit technical.
How machine learning actually works
Machine learning, as shown above, has so many applications. And while models are often trained for various purposes, and require different forms of training, some elements inform most machine learning models. Below is a step-by-step guide to how machine learning actually works.
Data collection: The first step is always data collection. Data is typically gathered from various sources, which could be structured or unstructured. The best outcomes typically rely on reliable data, which means data that is clean, concise, accurate, and enriched.
Data processing and standardisation: The data is processed to ensure it’s suitable for analysis. That process depends on effectively cleaning the data, as above, then transforming it through data normalisation, rendering data standard and scaled, and splitting into relevant training sets.
Training: The model, selected from one of the three mentioned above, is then trained on the clean and standardised data. The algorithm, during training, will likely adjust its parameters to minimise the difference between predictions and actual outcomes in the training data. That process often involves optimisation techniques, such as gradient descent and stochastic gradient descent.
Evaluation: The model’s performance is evaluated using the testing data. The testing phase assesses how the model generalises to unseen data, whether it can make predictions, decisions, or suggestions, depending on the desired outcome of the model, as explained above.
Fine-tuning: The results may lead to changes. That might mean adjusting hyperparameters or suggesting different features to improve performance. Changing hyperparameters can be done manually, or through automated means, and is usually conducted iteratively. These typically follow common techniques, such as Bayesian optimisation or grid search. Once the fine-tuning is complete, the model will likely go through the evaluation phase again to check success.
Deployment: Once the machine learning model is trained and evaluated satisfactorily, it can be deployed to make predictions or decisions on new, unseen data. That might involve integrating the model into software applications or systems where it can automate tasks or provide insights.
Re-evaluation: Machine learning is always an iterative process. Models need to be continually refined and improved, based on feedback and new data. Models can also be subject to AI drift, which can reduce the accuracy and validity of the model. Taking an iterative approach to evaluation helps the machine learning model remain effective and relevant over time.
Real-world examples of machine learning
So, we know the definition of machine learning, the various subcategories, and how machine learning models work in practice. Now let’s look at some real-world examples of machine learning, with reference to the application and some familiar examples that you likely know.
Image analysis and detection
Machine learning can recognise patterns, objects, and images. One common approach is through the use of convolutional neural networks, which are perfect for recognition, classification, and segmentation. The tech works by utilising neurons to automatically learn features from the images, enabling them to identify objects with high accuracy.
Real-world applications include facial recognition, a controversial piece of tech prone to practicing (and furthering) bias. Medical imaging relies on machine learning image analysis. And, on many smart phones, deciphering people by their faces also relies on image detection.
Recommendations
Machine learning can predict and suggest items based on preferences and user behaviour. The systems. One common approach is through content-based filtering, in which recommendations are made based on the characteristic of items and user’s historical preferences. In the simplest terms, the machine takes all your data, as well as the data of all potential suggestions, and simply aligns recommendations based on successful patterns in the past.
There are some obvious real-world applications, such as Netflix and YouTube suggestions, the information that appears on your social media feeds, product recommendations on almost any shopping website, the next song that plays on Spotify, and so on.
Chatbots and customer service
Machine learning enhances the capabilities of chatbots, making them more intelligent, more reactive, and more capable of providing an adequate response to user queries. Chatbots use natural language processing, and natural language understanding, along with intent recognition and sentiment analysis to allow the chatbot to respond in the best way. Machine learning enhances through personalisation, allowing chatbots to learn through each interaction.
Real-world applications include WaterAid’s Sellu, which provides an immersive experience and gives insight into the work of the charity. Another is ‘Is This OK?’, result of a partnership between Runaway Helpline and Childline, with funding provided by Children in Need, which provides support and useful information for teens that are feeling pressured and confused.
Sentiment analysis
Machine learning algorithms analyse text data from social media, customer reviews, or surveys to determine the sentiment (positive, negative, or neutral) associated with a particular topic, product, or service. That informs recommendations, as above, but can also provide valuable information to organisations, such as customer opinions, market trends, and brand reputation. Sentiment analysis is often used for inform decision-making.
Fraud detection
Machine learning algorithms detect fraudulent activities in financial transactions, insurance claims, or online transactions by analysing patterns and anomalies in data. These models can identify unusual behavior, suspicious patterns, or fraudulent transactions, helping businesses mitigate risks and protect against financial losses.
Machine learning is already playing a significant role in detecting, and reacting to, cyber crime across the wide economy. And governments are using it to detect fraud in programs such as tax evasion, social welfare benefits, and healthcare claims.
Other uses
These are just some of the many uses of machine learning. Other perhaps less ‘everyday’ uses include driverless cars, filtering through spam, traffic prediction, medical diagnosis, autonomous drones, speech-to-text transcription, agricultural yield prediction, and so on.
Machine learning, whether we are cognizant of it or not, features in so many parts of everyday life, often playing an essential role in AI elements of tech. Machine learning is iterative, which means it’ll improve with every usage, which means it’ll vastly improve in the years to come.
