3 Comments
author

Thanks for your insights, Jacob.

My take, as requested:

#Relevance is close to being a synonym of #Importance, including the role of attention in each. As for Signal-to-noise ration, that's under Detectibility (your #Conspicuity) in that a stimulus with high SNR is more likely to have high detectability than the same stimulus against, say, a distracting background. So I can't think of anything #Relevance adds to the construct (although it could be a subsystem). Also, nothing in the brain's salience network exists that implies a separate role for #Relevance .

#Interestingness is synonymous with novelty, so it, too, is subsumed by the first 3.

What I *didn't* mention is that salience can be directed not only to the outside world, but also to our "inner world" (the contents of consciousness). When that happens, a different (but overlapping) network is activated, called the Default Mode Network.

Hope my take was helpful,

Bill

Expand full comment

> #Relevance is close to being a synonym of #Importance, including the role of attention in each

I don't dispute that. I just considered "importance" to be goal-oriented while "relevance" was more feature-oriented. The first deals with cognitive states while the second deals with sensory processing. It's perhaps the same but on different levels of the hierarchy.

> As for Signal-to-noise ration, that's under Detectibility (your #Conspicuity) in that a stimulus with high SNR is more likely to have high detectability than the same stimulus against, say, a distracting background.

I actually disagree with this. I consider "conspicuity" to be a function of magnitude, not signal. Because there may be no signal to be had beyond its loudness. We can call it loudness instead if you like. It's like browsing a website covered in flickering advertisements. Just blank flashing boxes will overwhelm your senses making it difficult to filter out and find the relevant and important features.

> #Interestingness is synonymous with novelty, so it, too, is subsumed by the first 3.

Again, I disagree. "Interestingness" is a desirous positive quality of evaluating received information. Whereas, "novelty" or "surprise", are agnostic about the desirousness of the information. The key aspect of novelty is the unexpectedness.

For interestingness, the desired quality is the pleasing patterns, the symmetry, and the new juxtapositions that cause rearrangements of your internal cognitive framework in ways that can create extra connections and optimizations. This is Schmidhuber's hypothesis anyway, and I have to admit that it is indeed not novelty.

I've always been interested in all forms of "intrinsic utility" functions that are *not* pre-defined by an outside user for things like reinforcement learning, reasoning, or planning. Salience is one of those. The breakdown I provided is from what I know about different forms of intrinsic motivation I've encountered over the years. There are others, but relevant to the current discussion.

Expand full comment

Like you said, the term saliency in the AI zeitgeist, is a conflation of a lot of different meanings. I broke down the meanings I've encountered. I offer two more at the end and would be interested in your view on those:

# Different Meanings of Salience

## Conspicuity

- conspicuous, dominant, intense, eye-catching, prominent

- loud, bright, overwhelming the senses

## Novelty

- surprising, novelty, outlier, unexpected

## Importance

- pertinent to the goals, values, or trained responses

- pre-cognitive pattern-matching to trigger attention

## Relevance

- data and features that are needed for the task at hand

- signal vs. noise in input

- function of, or modulated by, attention?

## Interestingness

- a drive for curiosity and play

- first derivative of subjective beauty or compressibility (Schmidhuber 2009)

- perceived patterns that create internal mental patterns that can be integrated or reduced in novel ways (I think)

Expand full comment