(Here’s an observation about adjectives, verbs, and language in general. It might be important even if I’m misinterpreting the definition of natural latents.)
For many adjectives, we can define the concept “salience of <insert an adjective>”. Salience of color / texture / shape / size / etc.
For example, what’s “salience of a texture”? It’s a function of how much of the texture is present (in your field of view) and how strongly the texture contrasts with other present textures.
We can learn an empirical rule: “if a texture is salient enough, then it’s probably caused by a single object or a single kind of objects”.[1] Yet this object or kind is random. Would this make “being caused by an object/kind X” a natural latent over “pixels with a salient texture Y” and vice-versa? An object’s texture tends to be similarly salient in many different situations, so a particular value of “salience of a texture” can itself be a natural latent. “Salience of a texture” is not the same thing as “a texture”, but it’s one of the reasons why textures are important.
landscape example
Similarly, we can consider “salience of an action”. With an empirical rule like “salient actions (e.g. salient movements) are usually caused by a single object / a single kind of objects / a single causal process”.[2] Such rule makes fine-grained classification of actions important, making action-related words important.
Beyond “salience of a texture”, we can consider concepts like “danger of an object” (is it a good idea to touch or step on?) and “traversability of an object” (can you walk/climb on it?). For many types of objects, those concepts will be independent from most facts about the objects but their texture. Let’s call concepts like this (salience, danger, traversability, etc.) “auxiliary”.
Maybe this sort of reasoning can explain analogies, connotation. Because auxiliary concepts are a type of connotation.
Imagine looking at a nature landscape. You notice a bunch of angular texture (mountains), a bunch of puffy texture (clouds), a bunch fluffy texture (trees), a bunch of smooth texture (fields). The rule says each type of texture most likely belongs to a single object or kind. Note that it’s not trivial—we could live in a world where we often see radically different things with similar, equally salient textures at the same time.
(Here’s an observation about adjectives, verbs, and language in general. It might be important even if I’m misinterpreting the definition of natural latents.)
For many adjectives, we can define the concept “salience of <insert an adjective>”. Salience of color / texture / shape / size / etc.
For example, what’s “salience of a texture”? It’s a function of how much of the texture is present (in your field of view) and how strongly the texture contrasts with other present textures.
We can learn an empirical rule: “if a texture is salient enough, then it’s probably caused by a single object or a single kind of objects”.[1] Yet this object or kind is random. Would this make “being caused by an object/kind X” a natural latent over “pixels with a salient texture Y” and vice-versa? An object’s texture tends to be similarly salient in many different situations, so a particular value of “salience of a texture” can itself be a natural latent. “Salience of a texture” is not the same thing as “a texture”, but it’s one of the reasons why textures are important.
landscape example
Similarly, we can consider “salience of an action”. With an empirical rule like “salient actions (e.g. salient movements) are usually caused by a single object / a single kind of objects / a single causal process”.[2] Such rule makes fine-grained classification of actions important, making action-related words important.
Beyond “salience of a texture”, we can consider concepts like “danger of an object” (is it a good idea to touch or step on?) and “traversability of an object” (can you walk/climb on it?). For many types of objects, those concepts will be independent from most facts about the objects but their texture. Let’s call concepts like this (salience, danger, traversability, etc.) “auxiliary”.
Maybe this sort of reasoning can explain analogies, connotation. Because auxiliary concepts are a type of connotation.
Imagine looking at a nature landscape. You notice a bunch of angular texture (mountains), a bunch of puffy texture (clouds), a bunch fluffy texture (trees), a bunch of smooth texture (fields). The rule says each type of texture most likely belongs to a single object or kind. Note that it’s not trivial—we could live in a world where we often see radically different things with similar, equally salient textures at the same time.
For example, imagine you see a burst of flame and rubble flying around—most likely it’s a single causal process, an explosion.