For instance, suppose Google collects a bunch of photos from the streets of New York City, then produces a streetmap from it. The vast majority of the information in the photos is thrown away in the process—how do we model that mathematically? How do we say that the map is “accurate”, despite throwing away all that information? More generally, maps/beliefs tend to involve some abstraction—my beliefs are mostly about macroscopic objects (trees, chairs, etc) rather than atoms. What does it mean for a map to be “accurate” at an abstract level, and what properties should my map-making process have in order to produce accurate abstracted maps/beliefs?
Representation learning might be worth looking into. The quality of a representation is typically measured using its reconstruction error, I think. However, there is some complexity here, I’d argue, because in real-world applications, some aspects of the reconstruction usually matter much more than others. I care about reconstructing the navigability of the streets, but not the advertisements on roadside billboards.
This actually presents a challenge to standard rationalist tropes about the universality of truth in a certain way, because my friend the advertising executive might care more about minimizing reconstruction error on roadside billboards, and select a representation scheme which is optimized according to that metric. As Stuart Russell puts it in Artificial Intelligence: A Modern Approach:
We should say up front that the enterprise of general ontological engineering has so far had only limited success. None of the top AI applications (as listed in Chapter 1) make use of a shared ontology—they all use special-purpose knowledge engineering. Social/political considerations can make it difficult for competing parties to agree on an ontology. As Tom Gruber (2004) says, “Every ontology is a treaty—a social agreement—among people with some common motive in sharing.” When competing concerns outweigh the motivation for sharing, there can be no common ontology.
(Emphasis mine. As far as I can tell, “ontology” is basically GOFAI talk for “representation”.)
Representation learning might be worth looking into. The quality of a representation is typically measured using its reconstruction error, I think. However, there is some complexity here, I’d argue, because in real-world applications, some aspects of the reconstruction usually matter much more than others. I care about reconstructing the navigability of the streets, but not the advertisements on roadside billboards.
This actually presents a challenge to standard rationalist tropes about the universality of truth in a certain way, because my friend the advertising executive might care more about minimizing reconstruction error on roadside billboards, and select a representation scheme which is optimized according to that metric. As Stuart Russell puts it in Artificial Intelligence: A Modern Approach:
(Emphasis mine. As far as I can tell, “ontology” is basically GOFAI talk for “representation”.)