The current AI landscape is both exciting and confusing. Phrases like “advanced analytics” and “machine learning” are often used along with AI. You need to know what the words mean before you discuss how to adopt the technology.
One of AI’s challenges is that it’s a multi-disciplinary domain where even basic definitions are tricky. Here, we will focus on three terms and the relationship among them: AI, machine learning, and data science.
Artificial intelligence (AI)
AI combines three disciplines—math, computer science, and cognitive science—to mimic human behavior through various technologies. All of the AI in place today is task-specific, or narrow AI. This is an important distinction as many think of AI as the general ability to reason, think, and perceive. This is known as Artificial General Intelligence (AGI) which, at this point, is not technically possible.
This technology is rapidly evolving, and neither the scientific community nor industry agree on a common definition.
Some common definitions of AI include:
- A branch of computer science dealing with the simulation of intelligent behavior in computers.
- Advanced statistical and analytical methods such as machine learning and artificial neural networks, especially deep learning.
- A computer system able to perform specific tasks that normally require human intelligence, such as visual perception, speech recognition, decision-making, and language translation.
AI capabilities are rapidly evolving, and neither the scientific community nor industry agree on a common definition of these technologies. In this guide, we will use the definition of AI from the National Defense Authorization Act for Fiscal Year 2019, which is also referenced in the Executive Order on Maintaining American Leadership in Artificial Intelligence.
The term “artificial intelligence” includes the following:
- Any artificial system that performs tasks under varying and unpredictable circumstances without significant human oversight, or that can learn from experience and improve performance when exposed to data sets.
- An artificial system developed in computer software, physical hardware, or other context that solves tasks requiring human-like perception, cognition, planning, learning, communication, or physical action.
- An artificial system designed to think or act like a human, including cognitive architectures and neural networks.
- A set of techniques, including machine learning, that is designed to approximate a cognitive task.
- An artificial system designed to act rationally, including an intelligent software agent or embodied robot that achieves goals using perception, planning, reasoning, learning, communicating, decision making, and acting.
It is important to keep in mind that the definition of AI is still evolving and that achievements in the field today have been in task-specific AI or “narrow AI”, as opposed to what is commonly called artificial general intelligence that can learn a wide range of tasks—like humans.
Data science is the practice and methodology of implementing analytics or machine learning techniques, with subject matter expertise, to provide insights that lead to actions.
Data science is a broad field that covers a broad range of analytics and computer science techniques. This field—and the various professions that perform data science—are a critical component to building AI solutions.
In practice, data science is a cross-functional discipline that combines elements of computer science, mathematics, statistics, and subject-matter expertise. The goal of data science is to produce data-driven insights and processes that can help solve business, operational, and strategic problems for different kinds of organizations. This is often, though not always, achieved using machine learning and artificial intelligence capabilities.
Throughout these chapters, we will frequently refer to data science and data science teams. These are the teams who support the many data and AI efforts underway in government agencies.
Read more about how data science fits into the broader government AI ecosystem, Integrated Product Teams (IPT), and Developing the AI Workforce in Chapter 2 of the AI Guide for Government, How to structure an organization to embrace AI.
Machine Learning (ML)
Machine Learning (ML) refers to the field and practice of using algorithms that are able to “learn” by extracting patterns from a large body of data. This contrasts to traditional rule-based algorithms. The process of building a machine learning model is, by nature, an iterative approach to problem solving. ML has an adaptive approach that looks over a large body of all possible outcomes and chooses the result that best satisfies its objective function.
Though different forms of ML have existed for years, recent advancements in technology provide the underlying capabilities that have enabled ML to become as promising as it is today. Increased computing capacity (especially elastic computing infrastructure in the cloud), large-scale labelled data sets, and widely distributed open-source ML software frameworks and codes propelled the development of ML models. With these advancements, the accuracy of ML prediction and the number of problems ML can address have dramatically increased in the past decade.
There are three high-level categories of machine learning: supervised learning, unsupervised learning, and reinforcement learning. Each has its own mathematical backbone, and each has its own unique areas of application. Occasionally in more complex workflows, they may be combined.
Supervised learning, also known as supervised machine learning, is defined by its use of labeled datasets to train algorithms to classify data or predict outcomes accurately.
- Input data is fed into the model.
- Weights are adjusted until the model has been appropriately fitted, i.e. generalizes and adequately represents the pattern.
- A training dataset is used to teach models to yield the desired output and includes inputs and outputs that are correctly categorized or “labeled”, which allow the model to learn over time. The algorithm measures its accuracy through the loss function, adjusting until the error has been sufficiently minimized.
Supervised learning models can be used to build and advance a number of important applications, such as:
- Image and object recognition are applied computer vision techniques that are used to detect instances of objects of a certain type of classification such as a car or pedestrian. For example, in health care an AI system can learn to recognize what are pre-cancerous cells and what are not, in order to assist medical professionals conduct an earlier diagnosis relative to what a medical professional could determine on their own.
- Predictive analytics is used to provide deep insights into various data points and allows for the anticipation of results based on given output variables. Examples of predictive analytics include credit scoring to predict likelihood of paying on time based on factors including customer data and credit history.
- Customer sentiment analysis is used to extract and classify important pieces of information from large volumes of data—including context, emotion, and intent. It can be useful for gaining an understanding of customer interactions and can be used to improve customer experience.
- Spam detection is used to train databases to recognize patterns or anomalies in new data to organize spam and non-spam-related emails effectively. As the name suggests, it can be used to detect spam and help create a better user experience and reduce cyber fraud and abuse.
Unsupervised learning is often used in data exploration before a learning goal is established. Unsupervised machine learning uses unlabeled data. From that data, it discovers patterns that help solve clustering or association problems. It’s useful when subject matter experts are unsure of common properties of a data set. Unsupervised learning models are utilized for three main tasks—clustering, association, and dimensionality reduction. Clustering is a data mining technique which groups unlabeled data based on their similarities or differences. Association is used to discover interesting relationships between variables in a dataset. Dimensionality reduction is used to reduce the number of dimensions while still maintaining meaningful properties close to the original data.
Machine learning techniques have become a common method to improve a user experience. Unsupervised learning provides an exploratory path to analyze data to identify patterns in large volumes more quickly when compared to manual observation to determine clusters or associations.
Some of the most common real-world applications of unsupervised learning are:
- News feeds: used to categorize or “cluster” articles on the same story from various online news outlets.
- Computer vision: used for visual perception tasks such as object recognition.
- Medical imaging: used in radiology and pathology to diagnose patients quickly and accurately.
- Anomaly detection: used for going through large amounts of data and discovering atypical data points within a dataset.
- Customer personas: used to understand common traits and to build better buyer persona profiles.
- Recommendation engines: uses past behavior data to discover data trends that can be used to develop tailor such recommendations.
Reinforcement learning is a behavioral machine learning model that is similar to supervised learning, but the algorithm isn’t trained using sample data. This model learns as it goes by using trial and error. A sequence of successful outcomes will be reinforced to develop the best recommendation for a given problem.
Applications using reinforcement learning:
- Autonomous vehicles: used for self-driving cars, reinforcement learning improves safety and performance
- Industry Automations: used to control HVAC systems in buildings, data centers and various industrial centers, which leads to increased energy savings.
- Trading and Finance: time series models can be used for predicting future sales as well as predicting stock prices
- Language and text: used for text summarization, question and answering, and language translation using natural language processing
- Healthcare: used to find optimal policies and procedures using previous experiences of patient care without the need for previous information.
- Supervised learning uses labeled datasets to train algorithms to classify data or predict outcomes.
- Unsupervised learning uses unlabeled data. From that data, it discovers patterns that help solve clustering or association problems.
- Reinforcement learning sequence of successful outcomes will be reinforced to develop the best recommendation for a given problem.
- AI solutions use one, or in some cases several, of these ML techniques.
Myths about Artificial Intelligence
Though AI is a powerful technology already providing deep insight and business value, it is not magic. Understanding AI’s limitations will help you choose realistic and attainable AI projects. Below are some common myths about AI and pitfalls to avoid when evaluating it as a potential tool.
Myth about AI:
AI will replace humans in the workplace.
AI is more likely to replace tasks within a job, not the entire job itself. Almost all present-day AI systems perform specific tasks rather than entire jobs. The purpose of AI and automation is to make low-value tasks faster and easier, thus freeing up people to focus on high-value work that requires human creativity and critical thinking.
Historically, automation has created more jobs than it replaces. AI will mostly replace tasks, not jobs. It is more appropriate to think in terms of human-machine teams where each does the tasks for which it is best-suited. Many forecasts predict that new jobs will be created, i.e. people are and will continue to be needed for certain tasks and jobs.
Myth about AI:
AI can think like a human and learn on its own.
AI uses mathematical models and finite computing power to process information. Though some AI techniques might use ”neural nets,” these algorithms only remotely resemble human biology. Their outputs are still entirely based on data and rules prepared by humans.
Myth about AI:
AI is always more objective than humans.
AI applications are a product of data and algorithms combined into models. Data is collected, prepared, and managed by humans. Combining it with algorithms may still produce unfair and biased results. Machines and humans have different strengths and limitations. Humans are good at general tasks and big-picture thinking. Machines are good at doing specific tasks precisely. Human plus machine combinations are almost always superior in performance to a human alone or a machine alone.
Myth about AI:
You can just buy AI solutions that will work across the board.
Identifying AI use cases and the data required for them can be specific and localized. Further, the nature of algorithms and model training can require varying degrees of customization as the data is aggregated, cleansed, assimilated, and the outcomes are generated. Barriers to consider beyond technology include organizational culture, appetite for risk, the acquisition process, and agency willingness to experiment. Buy vs. build decisions require careful assessment.
Myth about AI:
Artificial General Intelligence (AGI) is just around the corner.
Artificial General Intelligence refers to AI that achieves general human-level intelligence. For most systems, there is a trade-off between performance and generality. An algorithm can be trained to perform one specific task really well, but not every possible task. Whether AGI takes decades or centuries to achieve, it’s more complex than most imagine. The more tasks we want a single machine to perform, the weaker its general performance becomes.
Myth about AI:
A large team of data scientists is required to implement an AI project.
Developing AI solutions might require only a couple of people a few weeks, or it could take years with a large team. It all depends on the nature of the objective, data, required technical infrastructure, and integration into the existing environment. Depending on the maturity of the AI applications related to the specific problem of interest to your agency, the level of data science involvement can vary significantly. Examples of how this may depend based on agency need are:
- Some applications, such as voice recognition, can be deployed from commercial-of-the-shelf (COTS) products.
- Some AI applications require training of an existing algorithm using agency-specific data, needing a small data science team.
- Some AI applications are still in the research and development stage. A relatively large data science team is needed to explore the data characteristics and identify the suited AI method to solve the problem.