Here’s an interesting paradox. Expertise is mostly about better mental representation, say Ericsson and Pool in the book I was discussing earlier.
What sets expert performers apart from everyone else is the quality and quantity of their mental representations. Through years of practice, they develop highly complex and sophisticated representations of the various situations they are likely to encounter in their fields – such as the vast number of arrangements of chess pieces that can appear during games. These representations allow them to make faster, more accurate decisions and respond more quickly and effectively in a given situation. This, more than anything else, explains the difference in performance between novices and experts.
This is an extension of Newell and Simon’s concept of chunks, which I talked about before here. (Ericsson worked with Simon at one point.)
Importantly, such representations are not at all the same as the parsimonious mathematical models beloved by economists, either. An expert may know tens of thousands of specific chunks in their field. But we still often know very little about exactly what that means, or how it works.
Now consider what this means for artificial intelligence. AI mostly abandoned attempts to develop better mental representations decades ago. Marvin Minsky of MIT advocated “frames” in 1974, but didn’t follow up with working systems. To be sure, the 1980s era expert systems had many simpler “if-then” type encoded rules, but faltered in practice. And today there is a great deal of attention to semantic networks to represent knowledge about the world. One example is DBpedia, which tries to present wikipedia in machine-readable form. A company called Cyc has been trying to hard-code information about the world into a semantic network for twenty years.
But semantic networks are flat and relatively uncomplicated. They are simple webs: graphs and links without much emergent structure. As for vector models for things like machine language translation, there’s almost no representation at all. It’s a brute force correlation, reliant on massive amounts of data. Recent advances in machine learning owe nothing to better representations. Chess programs effectively prune search trees efficiently, rather than use representations.
Meanwhile, the main data structures in machine learning, like ndarrays in the Python Numpy scientific module, or Dataframes in the Pandas module which is commonly used by data scientists, are essentially glorified tables or spreadsheets with better indices. They are not sophisticated at all.
So the most important thing in expertise is something that researchers struggle to reproduce or understand, despite all the hype about machine learning. Correlation is not the same thing as representation or framing or chunks. That’s not to say AI can’t make much more progress in this area. But there’s an enormous gap here.
Future advances are likely to come from better kinds of data structures. There’s little sign of that so far.