In the world of technology and technological research, when faced with complex problems, a dichotomy in possible solutions tends to develop. Indeed, in most cases, two approaches tend to prevail at an intermediate stage, one of which will later become the more universally used one. In the case of artificial intelligence applied to industrial automation (but not only), at the moment the open issue is one in which the Human in the loop solution emerges as prevalent, while the Human out of the loop approach seems to be losing momentum.
Before understanding why the Human in the loop approach currently seems to offer better results, let’s briefly look at the difference.
Human in the loop and Human out of the loop: assisted or autonomous choices?
First of all, let us reiterate a subtle but substantial difference: much of what is now erroneously called artificial intelligence is actually machine learning. To minimize the issue, let us recall that the goal of artificial intelligence is to enable a machine to make any decision in a manner similar to humans. For its part, the goal of machine learning is to obtain the most accurate results possible in a specific (even complex) task, using past data as a basis and improving through new data and feedback.
Having said that, and having ascertained that for the purposes of automation and industrial robotics machine learning perfectly meets any need, let us see the difference between the two approaches.
When we talk about Human out of the loop we generally refer to a more radical approach in which, once the required level of accuracy is reached in testing, tasks are entirely devolved to the machines and decisions are made completely autonomously by the system. In practice, an acceptable margin of error is set at the management level, and once that is reached, the system is autonomous.
The Human in the loop approach involves constant interaction between the system and the people in charge. In practice, automation is used for the collection, management, and organization of data, and in the most common approach for the automatic management of the simplest operations to ensure the basal metabolism of the system. Human operators are guaranteed two levels of intervention: providing feedback to the system to feed it new data and making decisions based on the outputs.
Human in the loop: the ideal approach for smart manufacturing?
After initial enthusiasm for fully automated systems, experts in all major fields now point to the Human in the Loop approach as the most advantageous, for a number of reasons both purely technical and statistical.
The technical reason is actually quite simple, at least in concept. As we know, machine learning requires having one or more initial datasets with which to train the system. When it comes to manufacturing or otherwise business and industrial processes, however, finding datasets that are perfectly matched to the specific needs of the system is definitely complex. This means that the system, to reach a level where it can be fully autonomous, would require a lot of time and resources.
On the other hand, from a statistical point of view, it is easily demonstrated that fully autonomous systems, while achieving excellent performance in day-to-day operations, easily fall apart in the face of new or complex situations. This is why the mixed Human in the loop approach is gaining more and more credence. Combining the processing power of machine learning with the discretionary ability typical of humans yields the best results. In support of this view, for example, a 2018 Stanford University study shows that a Human in the Loop system applied to medical diagnoses gets better results than either artificial intelligence alone or doctors alone. In short, to use a rather common slogan, an HITL system manages to get the best of two worlds.
