Artificial yes-men
About two months ago, I began a project to build a computer program that could learn by probability. I’m not going to write up all the details just yet—only those relevant to this particular story.
As mammals, humans are equipped with emotional reflexes. Fear, for example, allows us not only to become physically prepared for a threat, but also to recall and calculate the events caused by previous threats to aid us in making judgements regarding the supposed imminent threat. Basically, we are designed to learn on a very basic and fundamental level.
Computers do not yet have such sophisticated capabilities. The concept of learning is one of the great challenges facing the information age. However, I came up with the idea of allowing a computer to carry out simple learning processes loosely based on the human emotional reflex—i.e. assess how to handle a piece of information by assessing how previous information was handled. I refer to these processes as the Cohesive Learning Engine.
As test material, I’ve set the CLE to work reading pages from the web, and attempting to work out similarities between pages. In fact, it was the CLE that kick-started the Sri Chinmoy project. After a few weeks of reading and analysing data it picks up from the web, the CLE now has a clearer idea of what a weblog is than I do.
Now, not all the information processed by the CLE is automatically assessed. At this stage, it requires a certain level of human intervention to tell it whether or not it is making valid assumptions. To make this easier, I’ve been building a user interface so I can test the effectiveness of the automated analysis. For example, I can type in “is x similar to y?”, and the CLE will respond with either “yes” or “no”. I then will enter either “correct” or “incorrect” accordingly.
As I’ve been extending the capabilities of this UI, more and more interesting patterns have emerged in the CLE’s behaviour. Today, I’ve taught it what to do when I ask something like “do you know of anything similar to x?”. When asked this, it dutifully responds with either “yes, I know of z pages similar to x” or “no”. I then wanted to be able to get a random example, so tried to get it to answer “like what?”.
On its first attempt to handle “like what?” the CLE returned “yes, exactly like what”. It took me a moment to work out what it meant—and then a few moments more to work out why it had decided to respond in this way. Quite unintentionally, it would appear that the CLE has learned that it runs more efficiently if it agrees with me—therefore, it decided to do exactly that. I have taught a computer to kiss my back-side.