E-learning systems are quite common these days. More and more people are turning to free MOOCs (Massive Open Online Course) to expand their skills in their spare time, and most everyone under 35 has used some kind of Learning Management System (LMS) in their formal education. These platforms act as an online version of the classroom and are mainly, if not solely, focused on learning theoretical knowledge.
However, due to their static nature, these digital spaces don’t have a reputation for being very engaging, which makes it difficult to go beyond theoretical knowledge. Some companies are researching and building LMS augmented with artificial intelligence to make them adaptable to students’ learning styles and preferences, like using visual, auditive, theoretical or applied approaches. These AI-assisted systems are showing great promise and efficiency, and have gotten me thinking about how AI could function as a personal teacher outside of the classroom.
Teaching to Fly
Since 1993, I have been flying gliders, airplanes and paragliders. As the chief flight instructor of my glider association, I have trained many pilots for pleasure and aerobatics flying, as well as other flight instructors over the last 20 years. Over all these years, I have experienced my fair share of students with different learning styles.
Working in the AI industry, it was only a matter of time before I started thinking of the possibilities of an omnipresent instructor flying with me while I do a performance flight.
Working with all these different types of learners has given me loads of new perspectives with which I can reflect on my own learning style. I’ve been able to act as my own personal trainer, picking out different bits from working with others and developing my habits and methods to continuously improve. But no amount of experience will ever fully replace an outside observer who can make me aware of mistakes or bad habits, especially the ones I’m not aware of.
Working in the AI industry, it was only a matter of time before I started thinking of the possibilities of an omnipresent instructor flying with me while I do a performance flight, an instruction flight or even while I train another instructor. However, it’s a bit of a leap from what we currently have with LMS.
To understand better what our AI instructor would have to do, it is useful to take a quick look at how we, humans, learn practical skills.
From Theory to Practice
Learning practical skills, like driving a car, flying an airplane or operating machinery is a little different from learning theoretical knowledge like philosophy or calculus. The basic teaching techniques still apply, such as going from the known to the unknown, using examples and demonstrations, increasing the level of difficulty of exercises, etc. But, one aspect that is very different when learning a practical skill is that every action taken has an impact on the evolution of the situation. That means each action often requires follow-up actions to either correct the new situation or to make it evolve the way we want—to avoid getting into a worst-case situation, for example.
If you can, remember learning to drive: The force applied on the accelerator pedal influences the acceleration. Accelerating too fast may require you to brake to avoid bumping into the car in front. Brake too hard and you may surprise the guy behind and provoke an accident. Your driving instructor helped you understand these consequences, either while your were acting or after the fact. His job was to help you build a continuous “Situation – Options – Actions – Re-assess” loop, along with a good situational awareness. Following this loop is key to learning and improving practical skills.
One of the most powerful methods to learn a practical skill is called by development. It is a kind of feedback method that requires the instructor to be an expert in the matter. Instead of simply giving straight feedback on the action and resulting situation, the instructor will ask carefully chosen questions in order to stimulate thinking and create the appropriate neural relationships and reflex. Because the student develops the thinking and logic as he is guided, the information becomes much more permanent than when simply lectured at.
Development questions are used when the student has the minimal understanding to be able to construct new knowledge and when the situation is not critical, whereas direct feedback is more effective and sometimes required in emergencies or when the student is incapable of building on existing knowledge. Both techniques can be used during and after the fact, but never too long after.
An effective AI instructor would be able to do both. Airplane pilots have access to a lot of information in order to make their decisions, and modern cars have ever more too, but being aware of mistakes and bad habits is way more complicated. It requires thorough analysis of complex interactions between many factors. I analyse my glider flights when I train for contests, but that is after a long 4-7-hour flight with no feedback on my tactical decisions and flight strategy. I would much rather be told on the spot what mistake I am doing. The same goes when I do an aerobatic flight and things are moving really fast.
[Students] would have that AI to help them build the right habits of self-evaluating when I’m not there.
When I do an instruction flight, I would like to be debriefed by an AI on the clarity of the statements I used and the effect it had on the student, and if I used the right level of development questions with the student. Of course as a teacher, I especially wish that the students I let go fly solo would have that AI to help them build the right habits of self-evaluating when I’m not there.
Creating the right receptive state
An instructor must also be able to recognize the current receptive state of a student and will generate the needed emotions to put the student in the right mindset for learning new ideas. According to the Comfort Zone Principle, the best learning is done when in the Stretch zone.
In the comfort zone, the brain is on cruise control and wanders about rather than stay focused. If you’re an experienced driver, remember the times when you “woke up” because of sudden braking and realized that you just don’t remember the last 10-15 km. You were in the comfort zone. In the stretch zone, full attention is available, the brain absorbs the information and new neural connections are made. In the risk zone, there’s an information overload; a tunnel vision forms and situational awareness is lost. We react with our reflexes and can’t think. This is when the basic training is most important as we’ll simply repeat what we’ve learned with practice. In flight instruction we like to say that in these cases, “we sink to the level of our training.” In the danger zone, we freeze, brace, stare with empty eyes; we simply cease to function and think. Hopefully that never happened to you, but I have seen this happen to student-pilots: no fun.
Creating the right receptive state is possible with some smart use of direct feedback and development questions, gradually forcing focus and increasing the workload. A flight instructor, or any instructor, needs to adapt to the level of the student, which depends on many environmental and personal factors: fatigue, stress at work, turbulent air, prior experience, natural talent, etc. For example, I will tell a beginner or distracted student to “watch the yaw string, we’re flying sideways” and a more advanced student “how does the glider feel now?” In the former case, the student simply has to detect which side we’re skidding, in the later one he has to process a lot more signals.
Learning a practical skill until we become safe and competent enough is one thing. Continuing to improve is another. Building experience is often done alone or with little supervision, which means that we are prone to build bad habits because the outcome of our actions on the situation was not perceptible with our current competence level.
People with more talent, but especially those with better situational awareness, are more effective at improving practical skills by themselves. In any case, everyone could use continuous training to improve quicker and better. Once we get okay at a practical skill, we get into the comfort zone, which means that we are on cruise control and not really learning anything. Continuous training is a process for keeping ourselves in the stretch zone, which is best done with an instructor. The problem is that continuous training can be expensive, disruptive and often not well accepted because of human ego. Ask yourself how you would react if someone commented on your driving, after all those years without accidents.
Using direct feedback and development questions, the AI would ensure that we don’t build bad habits and would bring us into higher levels of competency.
I think that continuous training from an AI instructor would be much better received. The AI would be trained with all the best practices and experience from all the people that it is used by, globally. That means that there would be very little bias (there might be some, but that is a subject for another, more technical article). Using direct feedback and development questions, the AI would ensure that we don’t build bad habits and would bring us into higher levels of competency.
That works until we doubt what the AI would suggest. Unless we can understand why it is suggesting something, we might reject it. This is where a human instructor will shine. That brings us to a current field of research in deep learning and other forms of machine learning: explaining the output of an AI and how it comes to some conclusion. Until the research finds a solution to this challenge, can we say that the AI could have a “gut feeling” like we humans often do?
So, what if we can replace instructors in cars, machinery, flight training simulators and real airplanes for continuous training, and other machinery training? It brings back the important and oft repeated societal question: how many jobs will be lost to AI? One thing that AI will not replace anytime soon, if ever, is the empathy of a human. In every practical activity we learn, we will reach a plateau after some good progress. This plateau can be depressing and is when one might quit. At that stage a human instructor can have a conversation about this phenomena and find ways to stimulate motivation, which helps to create the appropriate state of mind to start learning again. We are many years from having an AI that will have the creativity to stimulate the student’s motivation in such cases.
The real power of AI is not in replacing humans at doing things, it is to make humans more efficient in everything they do.
In the end, an AI that can train us to become ever more efficient and performant rests on LMS being able to put to use these simple principles of teaching practical skills. However, as challenging as it is to train an AI with the knowledge of flight and instruction, the true hurdle is in developing a trusting and productive human-machine relationship. The real power of AI is not in replacing humans at doing things, it is to make humans more efficient in everything they do. After all, we are the ones who want to be flying.
Also published on Medium.