What is Human in the Loop?
Wikipedia defines Human-in-the-loop (HITL) as “a model that requires human interaction” and identifies its primary benefit as “allowing the user to change the outcome of an event or process.”Simply put, HITL is a machine learning or computing level that blends together the capabilities of an AI to automate tasks or behaviors while humans make judgment calls on trickier aspects of the task.
What are examples of human in the loop computing?
Perhaps the most commonplace example of human-in-the-loop computing is Facebook’s photo tagging function. If you’re a Facebook user and have uploaded a group photo recently, you would have noticed how Facebook has the capability of automatically tagging friends in your photo. Their photo recognition algorithm has improved so much that it can actually figure out the right friend to tag about 97% of the time. However, Facebook will still ask you, the user, to confirm if the person tagged in the photo is correct. Whenever you confirm or correct Facebook, the data is fed back to its algorithm in order to further improve their photo recognition program.
Another example of HITL machine learning are self-driving cars. Engineers have been working on self-driving cars for decades. In fact, research scientist Dean Pomerleau and Ph.D. student Todd Lochem from the Carnegie Mellon University’s Robotics Institute took a trip across the U.S. in an self-driving minivan, requiring only about 3% human interaction. But even if we’ve advanced in terms of self-driving programs, 99% accuracy still means 1 in 100 trips might result in an accident.
Furthermore, the technology to follow rules of the road, that is, being able to interpret signs and cues demanded by daily driving, is still being developed. This is why Tesla, just a couple of weeks after releasing their new self-driving features for Model S cars, placed new constraints on their autopilot system, requiring that human drivers keep their hands on the wheel. While the car can mostly drive itself on highways, it still requires a human failsafe to navigate during times that need more judgment calls such as navigating through construction or roadblocks.
Another great example of the HITL model are ATM machines. Up until recently, depositing cheques in ATM machines requires you to input the amount you’re depositing. But with the advancements in optical character recognition, today’s machines are able to actually read your handwriting and recognize important details like routing numbers. However, in cases where the machine is doubtful on its accuracy such as when the handwriting is incomprehensible or the language is not recognized, the ATM will still ask you to input the amount and flag the cheque for human checking.
Where are we right now in terms of HITL?
All the examples stated above show just important human and computer interactions remain to be. Today’s available AI technology still struggles with accuracy. Technology has advanced enough that machines can easily reach 80% accuracy. But what about the remaining 20%? What business would be comfortable to make decisions or present services that could be wrong twice out of ten times?
This is where HITL seeks to solve the gap. First, programmers create the data that machines learn from. After programming, machines automate much of the task, only relying on people to handle tough judgment calls (like following rules of the road, navigating through obstructions or deciphering bad handwriting). Computers, in turn, learn from these interactions to improve their accuracy.
HITL therefore is about delivering the best of both worlds—taking the capabilities of both machines and humans and having them work together to tackle tasks more efficiently and accurately.
Human in the loop computing as a service
All this means that even if your business is working with data created entirely by a computer or software, you would most likely still need to perform manual inspection of your data at different points of the data analysis process. Take for example Google—although they already have search algorithms in place, they still have human checkers to weed out irrelevant, sensitive or inappropriate search results.
For the average business though, this type of labor-intensive data analysis and scrubbing is simply too labor-intensive a process that not every business just has the resource for. This is why partnering with the right provider like Infinit-O can help you benefit from HITL by giving you a dedicated team to check and analyze your data, helping you to improve your processes, engage your customers and follow industry best practices. Click here to learn more.