Augustin Burchell

29 May 2020

I find expert systems fascinating, and I would like to learn more about them.

I work part-time as a research assistant with the lab of Denis Pelli of the NYU Psychology Department. Broadly speaking, we study vision; narrowly speaking, I study crowding. As I have come to learn, this unfinishable quest for understanding is the myriad manifestation of a simple game: consider a possible story, and then poke around at the world in ways which might help us determine if this story makes our understanding of how things are more or less elegant. The whole endeavor is like a much-too-hard Sudoku, to the point at which guesswork is required (at least at my current Sudoku/research skill level). If I put a three here, what otherwise elusive cells become inevitable? After these newly understood cells have been given their rightful number, what further identities are revealed? At the end of this process of conjecturing and following the conjecture to its logical ends, are we left with any impossibilities? Does this possible 3 here -- which could only lead us to a 5 here and a 7 there -- fly in the face of the 6, 2, and 1 we were already sure about?

It is this process of grabbing hold of this or that branch, shaking the tree, and examining whether something favorable falls out which leads me to the question (a question which seems only proper given my current recently-graduated-with-a-BA-in-Computer-Science circumstances): couldn't a computer make my life easier here?

My earliest thoughts were vague, but tantalizing. A graph seems apt; directed must be the way to go (so to speak). Concepts ought to be nodes and their relationships defined by their connections. These concepts could then be assigned a confidence level, an assessment of truth based on the current state of the literature (with said literature cited and searchable, of course). These confidence levels could then be tweaked, dialed up or down by toggling unreplicatable research out or tagging the findings of new studies in. The beauty would be that the computer would do my heavy branch-shaking for me. If we find that letter recognition efficiency is independent of uncertainty, what other questions about efficiency does that illuminate? Of what previous studies should this new finding make us wary and prompt us to try replicating? I had walked my way into what every student of computation dreams for: an opportunity to create a system which would allow me to become even lazier.

My current thoughts are still early, but with more names to which they can be affixed. Expert systems, it would appear, are where I should start my hunt. If ever I were to claim that I were born in the wrong generation (thereby clarifying that I was indeed born in a perfectly fitting generation), it would be while reading about LISP machines and ageless yearnings for the codifications of specific knowledge domains. It is not often that I have felt more reassured than while reading about truth maintenance and fuzzy logic. A possibility of novelty replaced with a more tractable sense of conceptual kinship, it felt good to know that my desires were at worst intelligible, and at best already implemented.

This is a declaration not of a solution, but of an intent to search for one. Along this search I intend to learn more about expert systems and knowledge graphs, of logic (first and second order, fuzzy or otherwise), and about what tools are available or lacking for making sense of potentially complicated or numerous dependencies between a fair number of concepts ("fair" in this case meaning too many to comfortably keep in one's head, but not enough to warrant throwing a neural net at the problem domain and hoping for the best). Whether I make something interesting, or get lucky enough to find that someone else has already made my interesting thing for me, I hope to game the system enough to get within perhaps a few degrees of becoming an expert on expert systems.