Human-in-the-loop (HITL) machine learning algorithms are an exciting new branch of artificial intelligence that combines the best of human and machine learning models. With HITL systems, human beings themselves play an integral role in the machine learning loop.
HITL systems and algorithms have a lot of applications for contact tracing. Unfortunately, purely automatic contact tracing apps are not very useful. Purely automate contact tracing algorithms are fairly susceptible to triggering false positives and false negatives. Thus, we cannot rely on automated systems alone to protect us from COVID-19.
That is where Human-in-loop machine learning and contact tracing intersect. Using humans’ input, developers can fine-tune contact tracing systems to be more accurate and more responsive to data. The result is more accurate contact tracing devices that can more efficiently respond to signals.
What Is Human-in-the-Loop Learning?
Human-in-the-loop learning is a revolutionary new method of machine learning that combines both human learning and machine learning. The idea of HITL learning is that humans are directly involved in the training, calibration, and testing of machine learning algorithms. HITL systems use human feedback to make changes to machine algorithms so they can work more efficiently. AI can take the input of large crowds of humans to modify their behavior and become more effective.
Why Is HITL Learning Necessary?
HITL learning is necessary because, despite what you might think, machines are actually pretty bad at learning. More specifically, machines are really bad at looking at a bunch of data and accurately categorizing objects, properties, and features into clearly defined groups. Machines may also be bad at picking up intentions or patterns in data that humans can pick out quite easily.
Here is an example of how this HITL learning works. Say you have an AI traffic cam that is supposed to take pictures of cars and determine whether they are staying within traffic lanes. However, these kinds of systems are prone to making false alarms. For example, the traffic light cam might not be able to accurately make a call if the car is very close to the lane lines.
HITL systems incorporate human beings into the learning process to prune false negatives and provide important guidance on edge cases. So in our AI-traffic light example, human beings can be used to go over the data, pick our false negatives, and help to program pick out edge cases. This allows humans to prevent overfitting of data, better delineate edge cases, or get the algorithm to detect new categories.
Humans can also play the role of validating output data. Humans can score outputs to identify problem areas and places where the algorithm can’t make a clear judgment.
The thing about HITL learning is that all of this is happening in a continuous loop. At each cycle of the loop, human-modified data is put back into the AI and it gets smarter the next round. This continuously updating feedback loops is superior to basic machine learning as it allows for the unique intuition of a human in the process.
HITL systems have been implemented in all kinds of fields, from text classifiers, computer vision algorithms, search/retrieval models, and more. Because humans can be the arbiter in edge cases, the algorithm has more accurate and more precise feedback.
How Can HITL Learning Be Used for Digital Contact Tracing?
HITL learning can be used to make contact tracing notifications and alerts more accurate. Combining digital and human tracers in a HITL system is the most efficient way to train algorithms.
Automated algorithms will always sometimes generate false positives or false negatives. This is just an inevitable fact about machine learning. Humans are also prone to making mistakes. The main difference is that humans are more-or-less able to account for systemic bias in their reasoning process. Machines are generally unable to do this, or not as able as humans.
Epidemiologists divide contacts into three kinds: close, causal, and transient. Whether or not an encounter falls into one of these categories depends on the context. For instance, contact tracing algorithms might not classify a short-term encounter in a cramped, poorly ventilated place as a contact because it does not trigger the algorithm’s benchmarks.
The problem is that contact tracing services do not generally record location and environmental data. Thus, they would be unable to accurately sort through and identify the distinct kinds of encounters that are relevant to epidemiologists.
HITL systems can be used to get around this issue. HITL systems can be used to classify these kinds of edge cases then feed that updated data back into the machine. This allows the machine to more accurately classify different types of contacts based on contextual clues.
Humans can also be used to make specific judgments that the machine system would not be able to. For example, a human can make the judgment to follow up with relatively distant contacts to account for the possibility of asymptomatic transfer. There might be epidemiological benefits to monitoring contacts that are twice removed, even if the closer contact has not been diagnosed.
Moreover, humans need to be involved in the contact tracing process on the front lines. Contact tracing efforts require humans to perform follow-ups and perform questionnaires with patients. Patients also need a human to guide them through the process and provide guidance on how the tech works and what steps to take next. Digital solutions can provide them with a continuous network of data to make decisions, while they work to refine the systems that generate that data.
Many people are under the impression that contact tracing solutions are purely automatic and require no effort on our part. According to Microsoft employee and contact tracing developer Sham Kakade: “There is automation, but there are humans in the loop in all of them. The humans are playing detective.”
Digital contact tracing is a necessary new approach to contact tracing, and can only be a great benefit for human contact tracers. Digital contact tracing can handle the brunt of the leg work, while the system can be managed by humans that make important top-level decisions that are fed back into the algorithm. We are still not at a place where technology can handle the full brunt of the work, but digital solutions can play an important complementary role to human tracers.