Human-in-the-loop (HITL)

Human-in-the-loop (HITL) is a type of artificial intelligence (AI) in which human intelligence is utilized alongside machine intelligence to make machine learning (ML) as efficient and effective as possible. Unlike fully automated AI systems that analyze data and make decisions on their own, HITL systems involve a certain degree of human input. This human input contributes significantly to ML and improves a system’s ability to learn. HITL is often particularly useful in situations where an AI system is required to handle edge cases, which are problems that occur at an extreme operating parameter. Because it is often more difficult for AI systems to accurately assess and respond to edge cases, human intervention through HITL may be critical in such scenarios. Overall, HITL effectively incorporates human judgment, improves transparency, enables more powerful systems, and reduces the pressure on developers to create perfect algorithms. HITL already has a number of practical applications, such as in self-driving cars and modern automated telling machines (ATMs).

Background

AI is defined as the simulation of human intelligence in machines that are specifically designed to think like humans and mimic human actions. The term can also be used to describe any machine that exhibits learning, problem-solving, or any other traits typically associated with the human mind. In any event, the defining element of AI is its ability to use reason to take actions aimed at accomplishing a particular goal.

There are two main types of AI: weak AI and strong AI. Weak AI systems are those designed to do one specific job. A personal assistant like Amazon’s Alexa is an example of a weak AI system. Such systems are simply designed to answer questions posed by human users. Strong AI systems are those designed to actually think and learn so they can carry out tasks in a human-like fashion. These systems are far more advanced than weak AI systems and require a much greater degree of machine intelligence.

One of the most important aspects of AI is machine learning (ML). ML refers to the process through which machines learn how to make accurate predictions. In other words, ML makes it possible for a machine to wade through and analyze mountains of data to draw conclusions and make decisions. To understand what makes ML unique, it is helpful to compare an ML system with a system that runs on traditional computer software. In traditional computer software, a human developer writes code that explicitly tells the system how to solve a problem. With ML, a machine is taught how to reliably solve a problem on its own through the input of a large amount of data. There are two main types of ML: supervised and unsupervised learning. In supervised learning, machines essentially learn by example. When undergoing supervised learning, a system is fed a large amount of labelled data to learn from so that it can eventually master some sort of task. In unsupervised learning, a system relies on algorithms to seek out and identify patterns in data with the aim of dividing the data in question into distinct categories. By doing this, the system will ultimately learn simply by grouping similar forms of data and identifying anomalies.

Overview

At its core, HITL is a way of creating ML models by combining both machine and human intelligence. In most variations of HITL, one or more people are part of a virtuous circle with a machine system. The human’s job within that circle is to oversee the ML model controlling the system. There are three distinct aspects of this job. First, the human user must label the training data that the ML algorithm used to learn to make decisions. Next, the human user must tune the ML model by scoring data to introduce new data categories, help the model to learn about edge cases, or account for overfitting, which is an error that can happen when a function is too closely fit to a limited set of data points. Finally, the human user must test and validate the model by reviewing its outputs. This last step helps to ensure that the model is capable of making sound decisions and reasonably confident about its judgments.

HITL has a number of advantages over other approaches to AI. Most importantly, it serves to make an ML model as transparent as possible. Because the HITL process involves at least some degree of human interaction, the model has to be designed so that it can be understood by humans and so that humans play a direct role in making critical decisions. This inherently makes the whole process more visible. HITL is also advantageous because it incorporates human judgment in a particularly effective manner. In other words, HITL is a more effective approach to ML simply because it allows for the inclusion of people in the decision-making process. Another advantage of the HITL is that it saves developers from the difficult task of having to create perfect algorithms from the get-go. Because HITL systems are supplemented with human intelligence and judgment, the automated part of the system does not have to be able to get everything perfectly right at every stage of the process. Rather, the system simply has to be able to make enough progress to reach the next point of interaction as the process unfolds. One final advantage of HITL design strategies is that they frequently lead to the creation of systems that are more powerful than either fully automated or fully manual systems.

AI systems utilizing HITL computing already have a variety of real-world applications. One of the most notable is in certain types of self-driving cars. While self-driving cars are often able to navigate roadways with near-perfect accuracy, fully automated self-driving cars always have at least a small margin of error. By incorporating a HITL approach in which human drivers are asked to take control in some situations, car manufacturers can produce self-driving cars that are safer and more reliable than when they operate entirely on their own. HITL also plays a role in some modern ATMs. While most ATMs now use visual algorithms to decipher information like dollar amounts and account numbers on deposited checks, these systems sometime have difficulty interpreting visual data correctly. In such cases, ATMs with HITL programming ask the user to manually enter the required information and then flag the check in question to be reviewed by a human bank employee.

Bibliography

Anadiotis, George. “Human in the Loop: Machine Learning and AI for the People.” ZDNet, 23 May 2018, www.zdnet.com/article/human-in-the-loop-machine-learning-and-ai-for-the-people. Accessed 9 Sept. 2020.

Biewald, Lukas. “Why Human-in-the-Loop Computing Is the Future of Machine Learning.” Computerworld, 13 Nov. 2015, www.computerworld.com/article/3004013/why-human-in-the-loop-computing-is-the-future-of-machine-learning.html. Accessed 9 Sept. 2020.

Copeland, B.J. "Artificial Intelligence." Britannica, 16 Jan. 2025, www.britannica.com/technology/artificial-intelligence. Accessed 17 Jan. 2025.

Fages, Bernat. “Introduction to Human in the Loop.” Human Lambdas, 30 Apr. 2020, www.humanlambdas.com/post/an-introduction-to-human-in-the-loop. Accessed 9 Sept. 2020.

Frankenfield, Jake. “Artificial Intelligence (AI).” Investopedia, 13 Mar. 2020, www.investopedia.com/terms/a/artificial-intelligence-ai.asp. Accessed 9 Sept. 2020.

Heath, Nick. “What Is Machine Learning? Everything You Need to Know.” ZDNet, 14 Sept. 2018, www.zdnet.com/article/what-is-machine-learning-everything-you-need-to-know. Accessed 9 Sept. 2020.

Johnson, Jonathan. “What Is Human in the Loop (HITL) Machine Learning?” BMC, 12 Aug. 2020, www.bmc.com/blogs/hitl-human-in-the-loop. Accessed 9 Sept. 2020.

Wang, GE. “Humans in the Loop: The Design of Interactive AI Systems.” Stanford University— HAI, 20 Oct. 2019, hai.stanford.edu/blog/humans-loop-design-interactive-ai-systems. Accessed 9 Sept. 2020.

“What Is Human-in-the-Loop Machine Learning?” Appen, 15 Jan. 2019, appen.com/blog/human-in-the-loop. Accessed 9 Sept. 2020.