Computers have become so ubiquitous that nearly every aspect of our lives revolves around their use, yet the machines haven’t lost their ability to amaze us. The latest jaw-dropping technology is the ability of computers to teach themselves new skills by analyzing huge amounts of data. The many types of machine learning promise to make our homes and workplaces safer, our access to information easier, and our lives healthier.
Machine learning applies sophisticated algorithms to massive data sets with the goal of allowing computers “to learn without explicitly being programmed,” as artificial intelligence pioneer Arthur Samuel explained in the 1950s. The data trains a learning model that system developers choose to perform specific tasks, such as identifying patterns or predicting the future. The developers adjust the learning model to make its pattern-matching or forecasts more accurate.
If you’ve used a speech-to-text system, interacted with a chatbot, or followed a recommendation made by Amazon or Netflix, you’ve had first-hand experience with machine learning (ML). Still, these applications are just a foreshadowing of the power and promise of ML to enhance our lives and our livelihoods. Here’s a look at the different types of machine learning, how we can use them, and what the future holds for each.
Supervised Machine Learning
In supervised machine learning, the model is trained by applying labeled datasets, which are annotated beforehand to identify characteristics of the raw data, such as images, text, or video, as well as to explain the context of the data. The model adjusts its weights automatically as it receives more data to improve the accuracy of its analyses and predictions.
The datasets used to train the model supply both the input and the correct outputs, which allows the model to approximate the desired output more closely with each iteration. Accuracy is determined by the algorithm’s loss function, which indicates high prediction accuracy when the loss function is low. The two types of operations in supervised machine learning are classification and regression:
- Classification categorizes the test data by identifying and labeling the dataset’s entities. Common classification algorithms include linear classifiers, support vector machines (SVM), decision trees, k-nearest neighbors, and random forests, which apply multiple decision trees.
- Regression examines the relationship between dependent and independent variables as a way to forecast future outcomes, such as projecting a company’s sales revenues. Among the most widely used regression algorithms are linear regression, logistic regression, and polynomial regression.
In addition to predicting a business’s sales, supervised ML is used to forecast swings in stock markets, identify patients most at risk of heart failure, distinguish cancerous cells from healthy ones, forecast the weather, detect spam, and recognize faces.
Unsupervised Machine Learning
The datasets used to train models in unsupervised machine learning don’t need to be labeled beforehand. This type of ML algorithm can determine differences and similarities in data without any preprocessing by humans. Three primary functions of unsupervised machine learning are clustering, association rules, and dimensionality reduction.
- Clustering places unlabeled data in groups by identifying attributes that are similar or different in their structures or patterns. For example, exclusive clustering creates a group that contains a single type of data, while overlapping clustering allows a particular data type to exist in multiple groups at one time. Two other types of clustering are hierarchical clustering, which merges separate groups of data into a single cluster iteratively, and probabilistic clustering, which groups data points based on the likelihood that they are a member of a specific probability distribution.
- Association rules identify relationships between the variables in a dataset by applying a set of rules, such as how the products in a market basket relate to each other. This allows a firm to better understand how its different products are associated, so they can gain insight into consumer behavior. One example of association rules analysis is apriori algorithms, which identify the likelihood of a consumer choosing one product immediately after selecting another.
- Dimensionality reduction helps improve the accuracy of unsupervised machine learning algorithms by reducing the number of features in a dataset. This addresses a loss of accuracy due to the inclusion of too many data features, or dimensions, in the set. The technique attempts to preserve the integrity of the dataset while extracting unnecessary data inputs. Types of dimensionality reduction include principal component analysis (PCA), which compresses datasets by removing redundancies; singular value decomposition (SVD), which extracts noise from image files and other data; and autoencoders, which apply neural networks to create a new, smaller version of the original dataset.
Common applications for unsupervised machine learning are predicting when and where cyberattacks are likely to occur, streamlining production in manufacturing settings, accident-avoidance systems in motor vehicles, and personalizing the shopping experience for a retailer’s customers.
Semi-Supervised Machine Learning
This type of machine learning uses both labeled and unlabeled data, so it serves as an in-between method when neither supervised nor unsupervised learning is the best choice for a particular application. Semi-supervised machine learning algorithms respond to a specific data point differently based on whether it is labeled or unlabeled:
- For labeled data, the model weights are adjusted by using the annotations that are applied in the preprocessing stage, just as they would be when using the supervised approach.
- For unlabeled data, the model bases its corrections on the patterns it identifies in similar training datasets.
By using some unlabeled datasets in addition to labeled data, semi-supervised learning reduces the amount of manual annotation the system requires, which cuts costs and shortens development time without reducing the accuracy of the algorithm. This technique makes several assumptions about the relationship between objects in the model’s dataset:
- Continuity assumptions imply that objects that are near each other are more likely to share the same label or group, an assumption that supervised learning also makes by adding decision boundaries. The difference is that semi-supervised learning adds decision boundaries with the smoothness assumption in low-density boundaries.
- Cluster assumptions divide the dataset into discrete clusters and apply the same output label to all data points in the cluster.
- Manifold assumptions are based on distances and densities in the dataset. The method converts high-dimensional data distributions into a low-dimensional space called a manifold. For example, a three-dimensional space is reduced to a two-dimensional coordinate plane, which allows the model to learn without requiring extensive amounts of data or processing.
Semi-supervised learning is often the optimal approach when the algorithm is processing a great amount of data, and when identifying relevant features becomes challenging. Use cases that fall into this category include the processing of medical images, speech recognition, classification of web content, and categorization of text documents.
The reinforcement learning technique for machine learning uses trial and error to reward positive outcomes and penalize negative ones. The system works by assigning positive values to the target actions or behaviors and negative values to all other responses. The reinforcement learning agent is programmed to find the route to the maximum long-term value. The method is applicable whenever a reward can be identified, such as in gaming and when making personalized recommendations.
The application of reinforcement learning has been limited to date by the need to maintain an accurate map of changing environments. Each change to the model’s known parameters requires that it run its trial-and-error routines to determine the option with the highest value. Doing so repeatedly is both time- and compute-intensive, especially in complex real-world environments. Three types of reinforcement learning algorithms are Q-learning, deep Q-networks, and state-action-reward-state-action (SARSA):
- Q-learning (the “Q” stands for “quality”) attempts to determine how useful a specific action is in realizing the target reward, or Q-value. It is called an off-policy algorithm because it learns from operations that aren’t part of the current policy. An example is the algorithm’s ability to take random actions for which no existing policy is required.
- Deep Q-networks are neural networks trained by deep Q-learning algorithms with the goal of overcoming the high resource requirements of Q-learning techniques. The neural network approximates the Q-value for each state-action pair. The network converts the state input to Q-values for all potential actions.
- SARSA is a form of Q-learning that calculates the reward for an action by adding a second action in addition to the initial action’s reward. The second action is based on the policy the algorithm has learned, so the reward for the first state-action pair is reset according to the new result.
Among the applications for reinforcement learning are self-driving vehicles, industrial automation, finance and stock trading, natural language processing, healthcare treatment planning, news recommendations, real-time bidding for online ads, and industrial robots.
What Does the Future Hold?
Various types of machine learning and other forms of artificial intelligence are transforming how organizations leverage data technologies to achieve their strategic goals and gain a competitive advantage. These advances allow firms to automate more of their business processes and realize a greater return on their investment in business intelligence platforms. Continuing refinement of AI methods is expected to lead to new types of machine learning that will make business operations faster, more agile, and more efficient.
Image used under license from Shutterstock.com