I’m an alien. But I’ll write this as though I were a human being.
This is a blog about the relationship between human thought and machine learning. Specifically, about identifying patterns in human thought that might then be able to be identified by a machine. The vehicle for doing this is language, but that is because it’s the best thing available to identify a lot of human thought. And human thought is cloth cut from human behavior. Many of the recombinant primitives that make up language can be seen as being derived from vision in the world, movement in the world, and sound in the world.
You may doubt this. Doubt itself is something to observe. Doubt contains the idea of incompleteness, of something missing, of something only partly visible. In fact, any of these sentences are themselves available for inspection. In supervised learning, the key element is the naming of the labels. In order to name labels, and to hand that task itself to machine learning, we need creative approaches to the task. How can we get the machine to identify complex patterns if we can’t at least point the way to the patterns that we’re after?
We live in a world of instances but we can only perceive a world of classes. This means that generalization is necessary for survival: it is the very heart of ‘recognition’ which is the most fundamental thing required to survive and thrive in the world. It is also the most fundamental basis for intelligence, whether artificial or human. The ability to recognize something as one thing, and be able to do whatever you want with it, is what makes you intelligent.
So generalization is key. We can’t keep track of everything in all its details. This means that learning must be recombinant, and as we’ll see, it is. This a portion of why thought is so varied, because it is based on a recombining not just of idea pieces (ideities) but also of dimensions. It is unclear how many dimensions there are, but it is clear that they are there.
If the key to supervised learning is proper labeling, then in order for a machine to learn higher-order structures, it mst be taught what to look for. That means that it is the job of the intelligence maker to label those more complex objects.