3 April, 2019 | Written by: Chris Williams
Categorized: Artificial Intelligence
Share this post:
When it comes to teaching Artificial Intelligence (machine learning), who is best placed to be the teacher?
The ability to pass on your knowledge to others is a great gift and the best teachers are those who not only can “do” (unlike the old adage) but also excel in doing.
But what about teaching a machine? Who are the best people to take on this task? And what skills do they need?
In my five years working with IBM Watson I’ve had the privilege of watching colleagues and clients training and teaching many different machine learning systems. I’ve also done a fair bit of this myself and have learned what works and what doesn’t. It became clear very quickly that it is really hard to predict is who is going to be good at teaching a machine and who isn’t.
Teaching artificial intelligence isn’t the same as programming artificial intelligence. Not only are new techniques necessary, but a substantially different philosophy and mindset is required. It may come as a surprise to learn that just as with humans, it might take several attempts for the learning to “stick and a good “machine trainer” will expect to get surprising results at first, relishing the challenge of trying to reduce the number of times the machine reaches the wrong conclusion.
Machine learning takes us into a world of uncertainty where seldom is the outcome expressed as 100% certainty. The mathematics underpinning AI is that of probability and statistics. We strive to teach the machine to be more certain, but there is no guarantee that the outcome will always be correct or indeed the same every time. Some people can handle this better than others.
Matters become more complicated when the meaning of the written word using natural language processing is the subject for machine learning. This might be a chatbot or virtual assistant, or a system that reads and understands documents. To be successful it is important to be adept at understanding how language is used in the domain and the context in which the AI will operate. Often there are many ways to express the same thing and the machine must be able to interpret these variances as the same intent. For example, how might an expert use words compared to a lay person? How might a young person use words compared to an older person? How might a stressed person use words compared to a calm person?
There is also a diagnostic or analytical element to teaching a machine. When the outcome appears to be wrong, why did it reach that conclusion? what might have misled it? And what is the best strategy to correct the error without impacting anything else?
So, what kind of person will be good as a teacher of machines? It is tempting to try and define this by role, (a business person or data scientist or developer or a natural language processing expert), but my experience is that all of these can be either good teachers or poor teachers.
My conclusion is that rather than relying on one person, a multi-disciplined team that has the required set of skills and experience will be most successful at the task of teaching artificial intelligence. These include:
- Deep real-world domain knowledge
- Great language skills (assuming you are using Natural Language Processing)
- Attention to detail and the ability to stick at what can be a tedious task at times
- Willing to keep trying different strategies
- Creative Thinkers
- Logical but flexible thinkers
- Good communicators
Teaching artificial intelligence well is actually very satisfying but it isn’t always easy. The business benefits of getting it right are huge so I would encourage you to have a go.
Since AI will quickly become a factor in every role, we need to educate the next generation on how to be good at teaching machine learning systems.