In this series, I am writing several blog articles about Artificial Intelligence. Today’s article concentrates on one of the most important aspects of the AI/ML problem….knowledge context.
I was sitting in a restaurant in Washington Regan airport the other day at a table next to a grandmother and presumably her grandson who looked to be about 15 or 16 years old. They were having a discussion about an upcoming trip he was taking to Israel and the Grandmother asked him to do some research before he went to find out how old the other children were at the small commune he was going to visit. Sitting in such close proximity to the two I couldn’t help but overhear their conversation. He immediately picked up his phone and started to search for the information. The grandmother said, “no, not that way!” and asked him to put down the phone. He looked shocked and an argument ensued with the young man claiming he had all the information at his fingertips now. The grandmother quietly argued that she wanted him to actually get the information in another way. With his phone in front of his face he proudly proceeded to tell her the ages of the children…..information that that he had found on the commune’s website during their argument. The grandmother again askied him to put down the phone and listen to what she was asking of him. The young man then tried to strengthen his position by reading the fact that there were currently x girls and Y boys at the commune. Finally, the grandmother asked the young man to please just call the commune and ask the commune’s leader the question. Reluctantly the young man dialed the number and started a conversation with a person on the other end of the phone. The young man asked this person how old the other children were and listened intently, inserting small acknowledgements and a few other questions along the way. When he hung up the phone the grandmother asked him, “so what did you find out?”. He proceeded to tell her the ages of the children, their gender, how long they had been there, the fact that one was disabled and confined to a wheelchair, and several other miscellaneous facts about their diet, and how they spent their days at the community. I could see the grandmother smiling as she asked the young man, So…do you see why I asked you not to use the Internet to get this information? The answer is obvious. The grandmother was trying to show her grandson that always relying on instant knowledge sources such as the Internet could give him some quick knowledge but not the real context that many times makes that knowledge useful.
I often tell people that semantic data elements and even knowledge itself is relatively sterile. What I mean by that is that unless there is enough context to provide you with a true understanding, the knowledge is useless and possibly even dangerous. A semantic data element is a word like “BRIDGE” or “RIVER”. Without the context of some other semantic data element these are just words. However, when we associate the data elements…for example…The bridge is over the river, we form knowledge. We now know something useful about both data elements. This is where the major problem lies with many new AI/ML applications being built today. Engineers and programmers are creating knowledgebases and attempting to use this knowledge in this form without taking the next step and turning the knowledge into deep understanding. Let’s look at how this could happen. If your user for the knowledge of the bridge over the river is a military analyst this knowledge means “River Crossing”. However, if your user is a water skier, this knowledge has a totally different meaning…i.e., “water hazard”. It is easy to see how adding other context such as how deep the river is, how wide the bridge is, where it is, etc. can make any users understanding of what this knowledge means to them even deeper and therefore more useful.
Real intelligence is built inside the brain in this way through the continuous addition of context to knowledge that makes it richer and richer. This deep understanding is how expertise is built and it is what differentiates an expert from a novice (Sweller, 1988). Cognitive psychologists call these contextual knowledge maps schemas (Miller, 1956) and like a puzzle continually being built inside the brain as new context is added the understanding of what the knowledge “means” becomes more and more complete.
This is really what machine learning is. It is not the continual addition of new facts or knowledge but the addition of new context to improve an existing understanding to make it deeper.
There is obviously much more that could be added to this discussion but if you hear someone describe machine learning as the addition of new knowledge to an application, remember that without context that knowledge may not be useful (or valid).
Miller, G. A. (1956). The magical number seven plus or minus two: Some limits on our capacity for processing information. Psychological Review, 63, 81-97.
Sweller, J. (1988). Cognitive load during problem solving: Effects on learning. Cognitive Science, 12, 257-285
Andy Bevilacqua, Ph.D.
Cognitive Optical Psychophysicist