Human-in-the-loop (HITL)

Human-in-the-loop (HITL, also translated as "human-computer collaboration", "human in the loop", "human in the cycle" or "human-computer interaction", etc.) is a branch of artificial intelligence that uses human and machine intelligence to create machine learning models.

HITL is an iterative feedback process by which a person (or team) interacts with an algorithmically generated system (e.g., computer vision, machine learning, or artificial intelligence).

Every time humans provide feedback, the computer vision model updates and adjusts its view of the world. The more collaborative and effective the feedback, the faster the model updates and produces more accurate results from the datasets provided during the training process. Just like a parent will guide a child's growth, explaining that cats "meow" and dogs "woof", until the child understands the difference between cats and dogs. 

HITL How does it work?

HITL aims to achieve goals that algorithms and humans cannot manage on their own. Especially when training algorithms (such as computer vision models), it is often helpful for human annotators or data scientists to provide feedback so that the model can better understand what it is showing. 

In most cases, the HITL process can be deployed in either supervised or unsupervised learning.

In supervised learning HITL model development, annotators or data scientists provide computer vision models with labeled and annotated datasets. The HITL input then allows the model to map new classifications for unlabeled data, filling in the gaps with greater accuracy than a human team could. HITL improves the accuracy and output of this process, ensuring that computer vision models learn faster and more successfully than without human intervention. 

In unsupervised learning, computer vision models are given large unlabeled datasets, forcing them to learn how to structure and label images or videos accordingly. HITL inputs are usually broader and fall into the category of deep learning exercises. 

How does this strategy improve machine learning results?

The overall goal of HITL input and feedback is to improve machine learning results. Through continuous human feedback and input, a machine learning or computer vision model becomes smarter. With constant human help, the model can produce better results, increase accuracy, and more confidently identify objects in images or videos. 

Over time, models are trained more effectively and produce the results project leaders need through human-machine feedback. In this way, machine learning algorithms can be trained, tested, tuned, and validated more effectively. 

Disadvantages of the HITL Workflow

Although human-computer interaction systems have many advantages, they also have disadvantages.  

Using a HITL process can be slow and cumbersome, and while AI-based systems can make mistakes, humans can make mistakes too. In the process, human errors can go unnoticed and then inadvertently negatively impact the performance and output of the model.

Therefore, machines are needed to annotate datasets. However, once humans are more deeply involved in the process of training machine learning models, it may take more time than when humans are not involved. 

HITL AI training example

An example in the medical field is with healthcare-based image and video datasets. A 2018 Stanford study found that AI models performed better with HITL input and feedback than when the AI models worked without human supervision or when human data scientists processed the same datasets without AI-based automation support. 

Humans and machines can work better and produce better outcomes together. The medical field is just one of many examples of using machine learning models for human-machine interaction. 

AI-based automated systems are extremely useful when performing quality control and assurance checks on critical vehicle or aircraft components; however, for peace of mind, human oversight is essential. 

HITL input is valuable when datasets are rare and are being fed into models. For example, datasets containing rare languages or artifacts where machine learning models may not have enough data to learn from and human input is invaluable in training the model generated by the algorithm.  

References

【1】https://encord.com/glossary/what-is-human-in-the-loop-ai/