For a while I have been wondering about the concept of salience. This page stores notes on this topic. I would be interested to advise an undergraduate project on this topic.
Salience in Machine Learning
[@adebayo_sanity_2020]
Salience in Expert Systems
Production Systems are a type of expert system that uses a set of rules to make decisions. The rules are applied to a set of facts, and the system takes actions based on the rules that match the facts.
Rule-based expert systems like Drools allow users to define the salience of rules by giving each rule an integer value that determines its priority in the activation queue.
Salience in Naga
The Naga library (Datalog-based) provides a salience implementation.
Defines a Queue structure that can be added to the tail, and removed from the head. Anything already in the queue (compared by ID) will not be added again, but a function can be provided that will update the element when it is already present. Includes a ‘salience’ which allows elements to be promoted through the queue ahead of less salient elements.
Salience in Red Hat Decision Manager
Red Hat Decision Manager also has a rule salience implementation.
Each rule has an integer salience attribute that determines the order of execution. Rules with a higher salience value are given higher priority when ordered in the activation queue. The default salience value for rules is zero, but the salience can be negative or positive.
Salience in Political Science and Social Choice Theory
For now this section is a list of papers that mention salience in the domain of political science and social choice theory.
Salience in Linguistics
For now this section is a list of papers that mention salience in the domain of linguistics.
Salience in Game Theory
“Salience” appears in game theory (at least) when noncooperative games with multiple equilibria are considered. (Nash) Equilibria in competitive games are fixed points where neither player is incentivized to change their strategy. Some games have multiple equilibria, and in some cases the empirical frequency at which certain equilibria are selected differs from what would be expected from rational agents.
As first hypothesised by Thomas Schelling (1960), players recognise (and expect their co-players to recognise) that, by virtue of differences in labelling, some equilibria are more ‘prominent’ or ‘salient’ than others; there is a systematic tendency for the most salient equilibrium (the ‘focal point’) to be selected. (Alberti et al., 2012)
My Interpretation: The “by virtue of differences in labeling” above refers to the labels of a payoff matrix, a common object of analysis in game theory. The payoff matrix represents the utility of combinations of strategy choices for each player. “Labels” are names for strategy or combinations thereof. These labels are usually not considered as part of the analysis and are chosen arbitrarily. But in the real world the semantic content of the labels influence to what extent a particular strategy is “salient”.
(Alberti et al., 2012) cites a few examples of how salience had been treated in the game theory literature:
- Salience as related to uniqueness of a particular label (Casajus 2001; Janssen, 2001; Bacharach, 2006)
- Salience as related to frequent public reference to a label (Sugden, 1995)
The authors argue that salience is an emergent property of human interation that arises in recurrent coordination problems.
Evolutionary game theory is a niche within game theory. From (Alberti et al., 2012)
evolutionary game theory studies games that are played recurrently by individuals drawn from large populations. Individuals … gravitate towards utility-maximizing choices in dynamic processes of experiential learning.
- Became popular in the late 80s-early 90s
(Alberti et al., 2012)‘s stated contributions:
- model of salience as an emergent property of recurrent interaction
- individuals make similarity comparisions between current and old problems, and current and old strategies
- individuals tend to choose strategies similar to those that performed well on similar problems
- This is due to Gilboa and Scheidler (case-based decision theory; 1995)
- Intuition goes back to Hume’s theory of induction
- This is due to Gilboa and Scheidler (case-based decision theory; 1995)
- similarity estimates can be based on labeling
- provides a formal model of similarity-based learning
- experiments with human subjects show that co-players learn to use similarity-based rules which increase coordination success
(Alberti et al., 2012) cites a few examples of emergence (spontaneous order):
- scree slopes
- checkerboard model of racial segregation (Schelling, 1978)
(Alberti et al., 2012) argue that:
- analysis of repeated games usually treat “convention” (seemingly spontaneous choice of one equilibrium over another) as originating in random perturbations (symmetry breaking)
- in other words, the benefits of convention are recognized, but the explanation for how one convention is chosen over another is not very satisfying
- effective coordination in real-world scenarios like the Right Turn Problem (how do drivers know who has the right of way at a crossroads) depends on multiple players having coincident salience values for strategies. No two instances of the Right Turn Problem are exactly the same, though they are all mutually similar. Failure to coordinate has high stakes (traffic accidents).
Salience in Cognitive Science
The cocktail party effect is a phenomenon where human listeners can partition auditory stimulus into distinct streams. Think of someone keeping track of a conversation at a noisy cocktail party. https://en.wikipedia.org/wiki/Cocktail_party_effect
The availability heuristic is a cognitive bias by which humans make decisions based on information that is most readily available. I feel like I have encountered some instances where “available” and “salient” were used interchangeably in describing this heuristic.
Salience in Human-Computer Interaction
Salience appears in Human-Computer Interaction (HCI) [@ruksenas_formal_2008]. Watching humans operate a device (like an ATM), slip errors are frequently observed, where users forget to take an appropriate action at an appropriate time. This is attributed to (cues for; availability of) the appropriate action not being sufficiently salient. [@ruksenas_formal_2008] models salience in HCI using higher-order logic implemented in SAL.
Salience in Neuroscience
The brain component named the hippocampus helps with the assessment of salience and context by using past memories to filter new incoming stimuli, and placing those that are most important into long term memory. Wikipedia (https://en.wikipedia.org/wiki/Salience_(neuroscience))
One interpretation of “salience” is the extent to which attention is directed towards a particular thing. Attentional control is a related research area. Models of attentional control can be broadly divided into bottom-up and top-down. I personally prefer to think of these as direct-bottom-up and indirect-bottom-up. I am not sure what the “top” is. It seems that for the control mechanisms referred to as “top-down”, the “top” is the latent state of the organism when experiencing some percept, and the “bottom” is the percept itself. The “top” is just the accumulated state from previous “bottoms”.
It seems that a similar argument was made in (Rumbaugh 1997) as cited in [@rumbaugh_salience_2007]. Rumbaugh in general seems to be arguing that some phenomena in behavior science cannot be easily explained by “stimulus-response” psychology. I take this label to refer to theories of psychology that are driven by reward/reinforcement. Rumbaugh advocates for “rational behaviorism” which “embraces instinctive, conditioned, and emergent behaviors” (whereas traditional behaviorism placed too much emphasis on conditioning; my interpretation).
[@rumbaugh_salience_2007] cites Jacob von Uexkull’s Umwelt theory. [@rumbaugh_salience_2007] calls salience “a property of a stimulus, or of previous experiences associated with a stimulus, that causes an organism to focus its attention toward that stimulus”.
It seems that the emphasis in neuroscience has been on bottom-up models. Some models derive “salience” values from signals (spike trains) coming from visual procesing neurons and their accumulation.
