Tethering the model to human intent is, indeed, a difficult problem. It won’t be solved, at least, not directly, through the NWITS model. In order to tether it to human intent, it must be tethered to language and its corresponding intent.
How do you do that? Well, here’s a way:
We train it to recognize human intent. Intent is very likely discernible in the plurality, i.e. in more than one instance. Here are some examples in which the words are in general not shared from one example to the next:
- My heart and soul
- The apple of my eye
- My everything
- My one and only
- My precious
- The light of my life
- The center of my world
- The sunshine of my life
- Sweetheart
- Sugar plum
- Sweet one
- My angel
- Love of my life
This could easily be rethought in order to make sure that no words are in common. It could also expand outward, since someone may have a treasured object like a car that’s the apple of their eye. However, any human being can discern what is common among these concepts. This is what intent is: it requires a multiplicity (more than one) of instances in order to reveal itself. This happens to be true of all language. There is no language without repetition. There is no intelligence without repetition. There is no perception without repetition. Repetition is half of the key to generalization.
What is generalization? It’s inaccuracy. It’s imprecision. It is, if not exactly, closely related to the idea of averaging. And this may well be why deep learning works. Because whether by accident or necessity, it is imitating something that must occur in the human brain. When I see you in a slightly different light, and in a slightly different angle, from what I have seen before, I still recognize you as you. This has been recognized in psychology for many years as visual constancy, or object constancy. We live in a world of instances, but we can only perceive a world of classes. I see the class that is you, not just the one time instance, with all of the attendant details down to an arbitrary level of accuracy.
If I am an artist (or a sensitive human being) then I can apply my attention, or my automatic perception, which includes a kind of visual intelligence, to observing the particular state that you present at this moment. You may have a wistful look as you were thinking about someone you lost, or you may be irritable because someone just flipped you off in the car. Or your appearance may vary along other dimensions for any of a massive amount of factors. Here, the word “panacea“ jumped into my head. It is the wrong word, but it does have the sense of spanning across domains.
And the fact that it appeared in my thought-orbit is telling: this is how we think.