Human-in-the-Loop Machine Learning: Improving AI Outcomes
Human-in-the-loop machine learning, sometimes called HITL, is a process whereby humans provide input to a machine learning model to maximize accuracy and increase efficiency for applications that use artificial intelligence (AI). But how exactly does it work and how can AI applications take advantage of this approach?
Keep reading to learn the following:
What is machine learning?
What is human-in-the-loop machine learning?
Benefits of human-in-the-loop machine learning
Examples of human-in-the-loop machine learning in action
What is machine learning?
Machine learning (ML) is a subfield of artificial intelligence (AI). It describes technology that imitates the way that humans learn – from experience. Machine learning algorithms use computational methods to “learn” information directly from data, to find patterns and make predictions on what to do with the data. It does all this without relying on an explicit instruction set.
Like a human learning a new skill, the technology will usually get better and improve its accuracy over time. This is because it will gather more data to work with – more data means more accuracy.
Three subcategories of machine learning
In general, there are three broad subcategories of machine learning:
Supervised learning algorithm
Data scientists supply labeled training data. Machines use the data to predict outcomes.
Unsupervised learning algorithm
Applications look for patterns in unlabeled data.
Reinforcement machine learning
The algorithm learns by interacting with its environment and improving behavior based on positive or negative responses to actions.
What is human-in-the-loop machine learning?
Human-in-the-loop machine learning incorporates a human element that provides input as part of the machine learning model.
HITL makes machine learning even more accurate
Conventional machine learning models need a lot of data to produce accurate results. Human-in-the-loop machine learning is one way to improve accuracy and develop human-level decision making faster.
You have probably been part of a human-in-the-loop model
If you have subscribed to an entertainment streaming service, you have probably been part of a machine learning model by being a human-in-the-loop. For example, many streaming services offer a recommendation engine that offers personalized and, in many instances, pretty accurate suggestions for what you might enjoy to watch next.
The data used by this machine learning algorithm could include:
Generic information about you that it gathered when you registered. For example: where you live, gender, age etc..
Search, viewing, and rating history (your data that it collects over time)
Data from other viewers
Information about the movie or show provided by the creator
Specifically, the more you watch and provide input, the more the recommendation engine gets to know you (and other people like you) to deliver better suggestions. You are the human in that loop helping to improve its accuracy!
Human-in-the-loop approach
Commonly the HITL approach is used in supervised and unsupervised machine learning when there isn’t enough existing data available. Taking this a step further, in our recommendation engine example, if you were presented with some recommendations that were not to your liking, you could feed that information back into the model and the algorithm would learn and get better over time.
HITL can help accelerate machine learning and minimize the likelihood of incorrect decisions being made, which can be quite costly. For example, if a robot makes an incorrect decision, there could be undesirable ripple effects down the supply chain such as an inferior product passing inspection or a product getting shipped to the wrong recipient.
Confidence level or score for human-in-the-loop machine learning
In those cases where the consequences of an incorrect decision could be substantial, using a confidence score is a good HITL approach. Say a machine is asked to solve a problem, and it isn’t able to solve the problem with a high degree of confidence, humans can step in and provide additional information. This feedback to the ML algorithm helps it learn, improve its predictions, and increase its accuracy.
This concept of “high degree of confidence” mentioned above can come in the form of a score that machine learning models deliver alongside their proposed decisions. A commonly used Confidence Level method is a number between 0 and 1, or between 0 and 100 that represents the likelihood that the output of a Machine Learning model is correct. The higher the score, the more confident the ML is that its response will satisfy the user.
Example of a robotic human-in-the-loop machine learning solution
Plus One Robotics makes vision software for robots that are used in warehouses and distribution centers. They created the PickOne Mixed Depalletizing Solution to address common problems of a mixed pallet. A mixed pallet is a pallet stacked in a non-uniform way with a variety of products. Check out our customer story where we explain the challenges of a mixed pallet including when familiar objects have a packaging change or come wrapped in highly reflective outer plastic.
How PickOne Mixed Depalletizer works with a human-in-the-loop
Computer vision images the pallet.
Computer vision images each pickable item and assigns each one a confidence level as defined by the customer.
If the item has a high confidence level, the robot is instructed to pick and place the items.
If there is an item with a low confidence level, the solution generates a request for a review. This human-in-the-loop, usually a crew chief, handles the exception.
The human response is stored with the other machine-learning data to improve the algorithms. The next time the robot encounters the same or similar object, it will know what to do.
With this human-in-the-loop machine learning solution, the crew chief can be located anywhere – they don’t need to be near or even in the same building. And, a single supervisor can monitor an entire fleet of robots simultaneously.
Benefits of a human-in-the-loop
There are many benefits to using a HITL machine learning approach.
Creates a better dataset
Human interaction ensures a more accurate dataset. This reduces the chance of errors thus improving accuracy and productivity.
Promotes a quick start
Businesses don’t have to spend weeks, months or even years creating a dataset before they get started with automation. Rather, they can build the dataset as they go. Initially, the confidence level can be set high. That means that the robot will only act if they are very confident that their response is correct. At first, the robot may require a lot of human input. Over time, it will require less.
Ensures agility
Models can be adjusted based on changing requirements or conditions using the human’s input.
Allows for in-situ error detection
A human-in-the-loop can often identify an error faster and therefore resolve more quickly before it gets too far down the line.
Enables values-based decisions
Machine learning lacks the context and social cues that humans naturally provide. HITL enables operators to adjust outputs where societal norms may be key to informing an output.
Hardware platform for human-in-the-loop machine learning solution
For their solution, PlusOne selected the Karbon 804. This powerful rugged computer supports their GPU and offers enough USB data lanes for up to 20 Intel RealSense cameras. In addition, the system helps them to meet some of their customers requirements of having a UL listing.
Our team has years of experience helping customers with their platforms for their AI, machine learning and computer vision solutions. Our ML confidence level is very high! Contact us to get us in the loop!
