What Is Human in the Loop (HITL) in Machine Learning?

Human in the loop in machine learning means pairing humans and machines to speed up processes, efficiently sort through masses of data, prevent bias and fine-tune training models.

HITL in ML, as it’s sometimes abbreviated, is used in four stages as a machine model is being built and perfected. First in training the model, that is, showing it how to sort through the data that’s creating the model; second, testing the model to ensure it uses the data correctly; deployment, or actually commissioning the model to do its job; and monitoring, or making sure the model continues to do its job correctly, and stepping in to make fixes if it is not.

What Is Human in the Loop?

“Human in the loop incorporates human knowledge in training and tuning machine learning models,” said Chen Zhang, CTO at RAIN, a voice technology company based in New York. HITL improves model accuracy, handles scenarios when the AI model isn’t confident enough to do so and overrides erroneous AI decisions.

It also helps overrule biased AI decisions, Zhang said. “For example, training data could contain mostly white people’s faces for a face recognition algorithm, or contain mostly Australian accents for a speech recognition algorithm,” he said. “Humans in the loop help root out those biases in favor of generalizing a model,” he said.

How Does Human in the Loop Work?

Because the singularity hasn’t quite arrived, humans play a role in every part of the process, from labeling the data to tuning the model to fix anomalies such as overfitting, the term for inaccurate predictions, to edge cases, which are scenarios the machine has not previously encountered. 

Julia Valentine, managing partner of AlphaMille, a New York-based technology consultancy, details four main ways humans step into the machine-learning process and add the human in the loop factor. 

Training

In many cases, humans train the model by demonstrating how tasks should be accomplished, Valentine said. Humans also evaluate and validate the results when they accept or correct the model. “At the basic level, that’s reinforcement learning,” she said. Just as a supervisor might take over a trainee’s task if time is of the essence, this approach works faster than traditional supervised learning algorithms, Valentine said. 

Testing 

Humans test the model results and determine if the model performs as expected, Valentine said. This is not without risks, though. One scenario: If the human training the model has unconscious biases, the model will be trained to have a similar bias, she said. Human brains and knowledge are required to test the models and make sure they are ethically sound.

Deployment 

It is difficult to foresee every potential use case, especially for edge cases, Valentine said. Issues can also arise when a model encounters incomprehensible or imbalanced data. “Humans use their imagination to think through extreme scenarios and decide whether the model is ready for deployment, and also when to instruct it to raise a red flag,” Valentine said.

Monitoring

When models successfully pass through training, testing and deployment stages, constant or intermittent human monitoring might be required to prevent costly mistakes or evaluate model drift and harmful biases.

Benefits of Human in the Loop

Teams are choosing to combine automation and human guidance because of the many advantages that human in the loop offers: 

• Increased accuracy. Human in the loop allows workers to provide immediate and constant feedback, correcting mistakes algorithms make.   
• Improved transparency. Human oversight provides a clearer view of how and why algorithms make decisions, creating more transparency around automated tools. 
• Greater efficiency. Machine learning algorithms can process large volumes of data, speeding up tedious tasks while people focus on more complex challenges. 
• Enhanced problem-solving. Humans can provide guidance when algorithms come across edge cases and other anomalies, resulting in more advanced problem-solving. 
• Reduced bias. Algorithms may rely too much on historical data, so human in the loop enables people to catch and root out any biases early on. 

Drawbacks of Human in the Loop

While human in the loop presents many upsides, teams must also be aware of a few drawbacks:  

• More human errors. Humans may make mistakes when training algorithms, such as missing an edge case or biased decision.   
• Higher expenses. Embracing human in the loop requires more personnel and software, leading to more expenses for companies. 
• Slower processes. Humans can’t work as quickly as machines, so human in the loop may take longer compared to a fully automated approach.  

Human in the Loop Examples

Here’s how the experts Built In spoke to for this story use HITL in ML at their own organizations. 

Sifting Through Mountains of Data

AlphaMille helps investment firms, including VC and private equity investors, incubators, family offices and investment banks, automate private-market deal discovery; that is, discern what’s out there and what’s available to buy. At any given time, thousands of companies can be available, and more come on the market every day, Valentine said. “It is humanly impossible to sift through that much information even with dozens of analysts,” she said. “Machine learning plus human in the loop is a great use case for deal discovery.”

Refining Machine Learning Models

Fintech startup DFD Partners matches asset and wealth managers with their next clients. The company’s mission is to help asset managers who lack the infrastructure of a giant company scale quickly. DFD’s ranking algorithm, which ranks advisors on how good of a match they are with the company’s users (asset managers), is central to the company’s mission. “It automates the time-intensive process of vetting potential clients for the asset managers,” said Devon Drew, founder and CEO. 

By holding focused feedback sessions, DFD gets input from wealth and asset managers, accredited investors and other industry experts to refine the model and thus help users get more accurate results. 

This approach builds on DFD’s original sorting algorithm. “We knew how important it would be to have something that used the latest tech available to get more accurate results, which is when we switched over to ML,” Drew said. “Human in the loop helps create a more amazing client experience.” 

Generating More and Better Labeled Data 

RAIN, the voice technology company, uses machine learning models to recognize intents (figuring out what customers mean from what they say to a voice assistant) and entities (data points related to that intent, such as a menu item or car part). 

RAIN is using ML to develop an automotive technician voice assistant and uses models to determine the correct intent — is the customer looking for info related to a specific car repair? — and to disambiguate between synonyms that map to the same entity, like the various ways to ask for ‘wheel nut torque’ on a vehicle, Zhang said. “Humans help to provide more and better labeled data continuously in order to improve these models,” he said. 

Executing Text Clustering Tasks

RAIN is also using unsupervised learning models for text clustering tasks, he added. To do so, it’s analyzing what users actually say to voice assistance and grouping them into linguistically similar utterances. That saves time because the entire log doesn’t need to be reviewed one query at a time. It turns lengthy utterances into feature requests, discovers linguistic diversity (such as accents) and also spots user frustration.  

“Humans help to evaluate the results from these models, and then transform the results into data that is consumable for other AI models,” Zhang said. 

Frequently Asked Questions