From fraud detection to image recognition to self-driving cars, machine learning (ML) and artificial intelligence (AI) will revolutionize entire industries. Together, ML and AI change the way we interact with data and use it to enable digital growth.
ML is a subset of AI that enables machines to develop problem-solving models by identifying patterns in data instead of leveraging explicit programming. The learning refers to the training process — the algorithms identify patterns in data and then use those patterns to tweak the model, aiming to provide a more accurate output each time. ML can be supervised, unsupervised or reinforced.
Through 2022, supervised learning will remain the type of ML utilized most by enterprise IT leaders
“Most of the current economic value gained from ML is based on supervised learning use cases,” says Saniye Alaybeyi, Senior Director Analyst, Gartner. “Yet unsupervised learning may be a better fit for certain problems — for example, when the goal is clustering entities and labeled data isn’t available. Reinforcement learning is still limited in its enterprise deployments, but its superior precision and targeting is promising for the future.”
Alaybeyi examines the three types of ML used in enterprise AI programs today and the business problems that each can solve.
Through 2022, supervised learning will remain the type of ML utilized most by enterprise IT leaders. Supervised learning is effective in many business scenarios, such as fraud detection, sales forecasting and inventory optimization.
Supervised learning works by feeding known historical input and output data into ML algorithms. In each step, after processing each input-output pair, the algorithm alters the model to create an output that is as close as possible to the desired result.
Supervised learning can be used to make predictions, recognize data or classify it
For example, a model could be fed data from thousands of bank transactions, with each transaction labeled as fraudulent or not. The model will identify patterns that led to a “fraudulent” or “not fraudulent” output, and over time, learn to more accurately predict whether a given transaction is fraudulent.
Input and output data can be derived from historical data, through simulations or through human data labeling. In cases involving unstructured data, like images, video, audio or text, certain properties or categorizations can serve as output data. Supervised learning can be used to make predictions, recognize data or classify it.
Example use cases for supervised learning include:
- Identifying risk factors for diseases and planning preventive measures
- Classifying whether or not an email is spam
- Predicting housing prices
- Predicting whether or not people will vote for a given candidate.
- Finding out whether a loan applicant is low- or high-risk
- Predicting the failure of mechanical parts in industrial equipment
Read more: The CIO’s Guide to Artificial Intelligence
Unsupervised learning is used to develop predictive models from data that consists of input data without historical labeled responses. For example, a list of customers or a set of unlabeled photos could serve as input data in an unsupervised learning use case.
The most common applications of unsupervised learning are clustering and association problems. Clustering produces a model that groups objects based on certain properties, such as color. Association takes those clusters and identifies rules that exist between them.
Example use cases for unsupervised learning include:
- Grouping customers by purchase behavior
- Identifying associations in customer data; for instance, people who buy a certain style or shoe may also be interested in a certain style of bag.
- Classifying people based on different interests
- Grouping inventories by manufacturing and sales metrics
Unsupervised learning can also be used to prepare data for subsequent supervised learning. This is done by identifying patterns or features that can be used to categorize, compress and reduce the dimensionality of data.
Reinforcement learning (RL) is based on rewarding desired behaviors or punishing undesired ones. Instead of one input producing one output, the algorithm produces a variety of outputs and is trained to select the right one based on certain variables. So, for example, a computer program could be trained to win a video game by identifying patterns in the actions that lead to it scoring more points than the other players.
Although it has been around for decades, RL has recently seen a renewed interest. RL requires less management than supervised learning, making it easier to work with unlabeled datasets. There have been some recent successes in RL implementations in the gaming world. However, practical RL applications are still emerging.
Recognize the potential opportunities for RL, but only employ it in limited scenarios
Most current data science and ML platforms don’t have native RL capabilities, and it requires much higher computing power than most enterprises have available. Right now, RL is only applicable in areas that can be fully simulated, that are quite stationary or where massive amounts of relevant data are available.
Example use cases for reinforcement learning include:
- Load balancing, for example, by teaching the algorithm to minimize the number of jobs waiting based on available compute resources
- Solving traffic jam problems by dynamically controlling traffic lights
- Training a robot to learn policies by mapping input from raw video images and replicating the actions it has “seen”
- Teaching a car to park itself, thus reducing time-consuming and trial-and-error work
AI leaders must develop significantly better simulation capabilities before RL can enter mass adoption. Recognize the potential opportunities for RL, but only employ it in limited scenarios.