What Is Machine Learning? Understanding ML

In this guide to understanding machine learning, we discuss:

• What Is Machine Learning?
• How Does Machine Learning Work?
• What Are the Four Types of Machine Learning?
• How Are Businesses Using Machine Learning?

Since the rise of mechanization in the early 1800s, humanity has continued to develop smarter machines to improve our quality of life. From the steam engine through to the self-driving car, each era of the modern age has been defined by its technological advancements—and machine learning (ML) is no different. What sets machine learning apart is that its impact isn’t limited to any single aspect of everyday life. In fact, machine learning algorithms have been behind the majority of technological innovations from the last five years.

Whether you know it or not, you likely already encounter machine learning on a daily basis. Here are four areas where machine learning has already been responsible for significant change outside the realm of science fiction:

• Generating highly-tailored user recommendations on platforms such as Spotify, Netflix, and Google
• Identifying a person or object from an image for use in facial recognition and visual search
• Powering the speech recognition and data processing behind virtual personal assistants such as Alexa and Siri
• Predicting whether financial transactions are fraudulent based on previous patterns of behavior

As early as 1959, Arthur Samuel, an early artificial intelligence (AI) pioneer, defined machine learning as “the field of study that gives computers the ability to learn without being explicitly programmed.” Now, almost 70 years later, that definition has expanded to include a wide variety of different algorithms and models. Read on to learn what machine learning is, how it differs from other forms of artificial intelligence, and why it matters to your business.

What Is Machine Learning?

Machine learning is a subfield of artificial intelligence focused on developing computers that learn similarly to humans. Through the use of algorithms that seek to imitate human intelligence, machine learning enables AI to improve outcomes through iteration. This ability to solve problems and generate predictions without explicit programming makes machine learning highly adaptable.

While machine learning enables computers to learn with a degree of independence, the computers still require human input. Data scientists feed training data into a machine learning algorithm to create a machine learning model. By adding live data to that model (once the training is complete), users are then able to generate new predictions. Finally, the outcomes of those predictions serve as further training data, increasing the model’s accuracy in what is referred to as the “flywheel effect” of accelerating progress.

Machine Learning vs. Artificial Intelligence

Artificial intelligence refers to any technology that enables machines to simulate human intelligence. AI and machine learning are closely related with similar goals. However, AI encompasses many methods beyond ML on its own, including search algorithms, rule-based systems, and genetic algorithms. Not only that but AI also has a wider cultural context, and there is significant ongoing discussion and debate regarding theoretical future AI developments and directions.

If AI refers to the broader concept, then ML is one specific application of AI. Every AI methodology shares the same broad objective: enabling a machine to complete a complex task effectively. As a subfield of AI, machine learning achieves this goal by analyzing large volumes of data. However, the focus of an ML model is narrower, since each model is typically dedicated to one specific task.

Machine Learning vs. Deep Learning

Deep learning is a type of machine learning that uses neural networks to more closely mimic the structure of the human brain. Deep learning requires much more data and computing power than machine learning, since it reduces the need for human intervention. By using multiple layers of neural network processing, deep learning models can analyze and learn from massive unstructured datasets. Just as artificial intelligence is an overarching term, machine learning is an umbrella category that deep learning sits under.

The key to understanding deep learning lies with neural networks, also known as artificial neural networks (ANNs) or simulated neural networks (SNNs). Neural networks consist of thousands (or even millions) of simple processing nodes that are connected in a layered structure. As a result, they can model complex nonlinear relationships between input and output data, and classify data more efficiently. This is particularly useful in computer vision—the process by which machines decode visual imagery like humans do.

How Does Machine Learning Work?

Machine learning models can take a wide range of forms (more on that later) but the guiding principles remain relatively consistent. According to UC Berkeley, ML algorithms typically consist of three components:

1. Pattern recognition and prediction: ML models are most often used to either classify data or predict outcomes. As such, each new decision process begins with an initial estimate. Once the input data has been provided, the algorithm will attempt to “guess” the type of pattern it’s supposed to find.
2. Error calculation: Next, the algorithm needs to compare its estimate with existing examples (if they have been provided). By quantifying how accurate it believes its initial prediction was, it can assess the scale of error.
3. Optimization: Finally, the algorithm analyzes the decision process used to reach its estimate, adjusting for future iterations. By changing the “weights”' assigned to each parameter, it reduces the discrepancies between any provided examples and its own estimates. This continuous process of iterating, evaluating, and optimizing means the final model will produce more-accurate outputs.

For example, to train an image recognition system, a data scientist could provide the algorithm with a labeled set of dog and cat pictures. The algorithm would take that input data and begin to distinguish the differences between cats and dogs. These different parameters could include the size and profile of each animal, the differing types of fur, and placement of facial features.

The algorithm would then assign each of these parameters a weight, depending on its perceived usefulness and relevance. If the algorithm correctly identified a cat, the weights wouldn’t be adjusted, but if it was incorrect, the parameters used to reach that conclusion would be given lesser weighting. That way, the model gradually reduces the likelihood of making further mistakes.

What Are the Four Types of Machine Learning?

Machine learning models are typically grouped based on how each algorithm is designed to learn. The four most common types of machine learning are supervised learning, unsupervised learning, semi-supervised learning, and reinforcement learning. However, these categories are broad, and many machine learning models will incorporate aspects of each. For example, deep-learning models can fall under any of these four categories.

Choosing the right type of machine learning for a task depends largely on the specific goal and the dataset that a data scientist is working with. In fact, algorithms will often be adapted based on the specific challenges a data science team (or their users) are facing. The best way to assess what machine learning model will work best for your needs is to understand how each algorithm works.

How Does Supervised Learning Work?

Supervised machine learning, also referred to as supervised learning, works by using labeled training data. Data scientists assign labeled data one or more tags to give the algorithm useful context, such as distinct categories or number values. For example, a set of emails may be labeled as “spam” and “not spam,” providing the ML algorithm with a structured guide to learn from.

By analyzing the relationship between the input (data) and the output (labels), the algorithm learns to map one onto the other. Once this training is complete, with the weights properly adjusted, the model can then predict the output for new data. Due to its relative simplicity, this is the most common form of machine learning today. Examples of supervised learning techniques include:

• Linear regression: Creates a linear relationship between a dependent variable (the input e.g. advertising spend) and an independent variable (the output e.g. overall revenue) to predict future outcomes. This can be used to estimate the strength of the relationship between variables, and the value of the dependent variable at a certain value of the independent variable. For example, how salary impacts employee satisfaction.
• Logistic regression: Predicts the probability of a binary outcome based on one or more independent variables. The outcomes will always be binary, such as yes/no, 1/0, or true/false. This is primarily used for prediction and classification tasks, such as identifying employee churn risk.
• Decision trees: Models future outcomes and predictions using branching-linked decisions, which form a tree-like structure. These branching-decision sequences categorize complex datasets, identifying ways to group and visualize the data. This is useful when strategizing, such as creating a budget, or assessing the impact of purchasing a new solution.

How Does Unsupervised Learning Work?

Unsupervised machine learning, also referred to as unsupervised learning, exclusively works with unlabeled datasets. Unsupervised ML algorithms analyze datasets for trends, clustering data points into different sets in the process. Working with reduced human intervention, these algorithms often find patterns in data that would otherwise remain hidden.

While unsupervised learning is primarily used for data-clustering tasks, there are many areas where it can prove useful. Common uses for unsupervised learning algorithms include:

• K-means clustering: Divides data points into sets based on their similarities, and discovers underlying patterns. K-means clustering algorithms search for a fixed target number (K) of clusters, set by the data scientist. This is one of the most popular types of clustering algorithm due to its simplicity and efficacy. Recommendation engines, such as those used by social media platforms, often use K-means clustering to suggest content based on a user’s prior behavior.
• Association rule: Identifies the strength of relationships between data items, counting the frequency of complementary occurrences. By finding associations that take place at a far higher rate than a random sample, businesses can strategize accordingly. This is particularly useful in identifying customer buying trends, such as products that are frequently bought together.
• Dimensionality reduction: Simplifies a dataset by removing redundant features and noisy data, while still retaining its meaningful dimensions. When dealing with large datasets with sparse raw data, reducing the number of variables makes analysis much easier. For example, natural language processing technology often only extracts useful vocal features for speech recognition.

How Does Semi-Supervised Learning Work?

Semi-supervised machine learning, or semi-supervised learning, bridges the gap between supervised and unsupervised learning methods—as its name suggests. Semi-supervised learning models use both labeled and unlabeled data during the training process. By feeding small amounts of labeled data into an algorithm, it can apply those learnings to the full unlabeled data set. Since labeling data can be a tedious, costly process, semi-supervised learning is often an efficient solution.

As semi-supervised learning is a happy medium between the two previously mentioned methods, its applications remain similar. Here are three situations where semi-supervised learning can prove valuable:

• Fraud detection: In the situation where a financial team only has a handful of confirmed examples of fraudulent activity, semi-supervised learning systems can learn from the smaller data set. Since fraud is both abnormal and hard to detect, this method saves accountants from having to sort through thousands of transactions.
• Content classification: Reading through and annotating large volumes of content can take humans an incredibly long time. With semi-supervised learning, human annotators only have to assemble a small selection of hand-labeled examples. This can apply to everything from classifying web pages for search engines to classifying incoming emails for email clients.
• Speech recognition: Capturing the breadth and range of human speech, including accents and vocal variance, is a major task. Semi-supervised learning works from a small training set of human-annotated audio before performing its own self-learning. In self-training trials by Meta, the word error rate decreased by 33.9%.

How Does Reinforcement Learning Work?

Reinforcement machine learning, also called reinforcement learning, works using trial and error. Unlike other methods, a reinforcement learning algorithm is programmed with a specific goal and a clear set of rules to follow. Additionally, the data scientist further includes a points-based goal—positive results add points, while negative results subtract them. This feedback loop serves to improve results over time.

Reinforcement machine learning algorithms are most useful for sequential decision-based scenarios, such as games, robotics, or project management. By finding the most efficient route to the maximum reward, reinforcement learning can be a powerful decision-making tool.

How Are Businesses Using Machine Learning?

In the digital world of work, data has become a company’s most valuable asset. Machine learning represents an opportunity for companies to leverage historical data to better strategize for the future. As augmented workforces become the norm, companies that continue to rely on manual processes and fail to fully utilize their data will fall behind.

According to Workday research, 80% of decision-makers believe that AI is necessary to keep their business competitive. Despite this, 76% say their knowledge of AI and ML applications needs improvement. To succeed, business leaders must understand where machine learning can bring the most value to their business.z

Below are a few examples of how Workday customers are already using our embedded machine learning:

• Recruiting the best candidate: Manually evaluating high volumes of job applications can be a mammoth task. With machine learning, recruiters can quickly match job requisitions with potential candidates, clustering them based on the strength of their match. A large multinational automotive manufacturer experienced a 70% increase in candidate screening efficiency by using HiredScore AI for Recruiting.* 
• Identifying and tracking skills: Understanding the full breadth and depth of talent in your workforce is no easy feat. Rather than relying on a basic catalog of skills, machine learning enables a multidimensional overview. Whether surfacing insights on skills gaps or clustering skills based on industry, region, and proficiency, ML is critical for developing a skills-based talent strategy.
• Enhancing internal mobility: If your talent doesn’t have regular opportunities to develop and grow, they’re at risk of attrition. Machine learning can surface tailor-made learning recommendations and job openings based on an employee’s skills, role, and tenure. By using our ML-generated role recommendations, a major global real estate company saw a 10% increase in internal mobility engagement. 
• Improving process efficiency for managers: People leaders spend a lot of valuable time on manual processes. With ML, it’s possible to streamline scheduling, surface insights from employee feedback, and address time anomalies. In fact, a corporate ventures organization was able to achieve a 50% manager self-service rate for HR processes, enabling far greater oversight and accountability. 
• Automating finance intelligently: While automation has touched many parts of the finance function, too many processes remain manual. The intelligent automation enabled by machine learning includes supplier invoice scanning, receipt scanning for expenses, and customer payment matching.
• Detecting data anomalies: A business is only as good as the quality of its financial data. Machine learning flags any anomalies in the general ledger early in the cycle, improving forecast accuracy. That way, your financial professionals can focus on more strategic and valuable work.

Taking the Next Step with Workday

Workday C-suite global AI research shows that 98% of CEOs believe there would be some immediate business benefit from implementing AI and ML. However, only 1% would classify their adoption as mature. Accordingly, the potential competitive advantage for businesses that integrate AI across their organization is huge.

One of the major barriers business leaders face is trust. At Workday, we believe in responsible AI that is transparent, explainable, private, and secure. That means providing our customers with AI technologies that have been designed with responsible AI at the core. We document our alignment to existing and developing regulations and best-practice frameworks, in addition to providing customers with fact sheets describing how our solutions are developed, assessed, and maintained.

Furthermore, Workday AI is built into the core of our platform so that we can rapidly deliver new AI capabilities that offer meaningful business value. Our AI is trained on the industry’s largest and cleanest set of finance and HR data, so it’s always grounded in reality. With over 65 million users on the same version of Workday, only our customers have the trusted data necessary to move forever forward with machine learning.

To learn more about how Workday can support you in your AI and ML journey, explore our technology page.

*HiredScore is a Workday company.