Humans parsimoniously represent auditory sequences by pruning and completing the underlying network structure

Abstract

Successive auditory inputs are rarely independent, their relationships ranging from local transitions between elements to hierarchical and nested representations. In many situations, humans retrieve these dependencies even from limited datasets. However, this learning at multiple scale levels is poorly understood. Here We used the formalism proposed by network science to study the representation of local and higher order structures, and their interaction, in auditory sequences. We show that human adults exhibited biases in their perception of local transitions between elements, which made them sensitive to high-order structures such as network communities. This behavior is consistent with the creation of a parsimonious simplified model from the evidence they receive, achieved by pruning and completing relationships between network elements. This observation suggests that the brain does not rely on exact memories but compressed representations of the world. Moreover, this bias can be analytically modeled by a memory/efficiency trade-off. This model correctly account for previous findings, including local transition probabilities as well as high order network structures, unifying statistical learning across scales. We finally propose putative brain implementations of such bias.

Probing sensitivity to statistical structure in rapid sound sequences using deviant detection tasks

“…Experiment 2 used the same deviance detection task, but the structured sequences were generated using a community structure (Benjamin et al, 2023;Karuza et al, 2017Karuza et al, , 2019Lynn et al, 2020;Schapiro et al, 2013). This structure, depicted in Figure 3 (top panel), contains 15 nodes; each node is connected to four other nodes and these interconnections are structured so as to group the nodes into three "communities".…”

“…While it is known that humans and other organisms are sensitive to environmental statistical patterns, it is still unclear what type of statistics are learned under different situations, how these patterns are represented, and what impact the resulting representations have on ongoing sensory processing (Conway, 2020;Maheu et al, 2022;Planton et al, 2021). Research focused on auditory processing has historically emphasized deterministic sequences (i.e., precisely repeating frequency patterns; Barascud et al, 2016;Heilbron & Chait, 2018;Southwell et al, 2017;Southwell & Chait, 2018;Winkler & Denham, 2024), yet it is becoming increasingly clear that listeners track sophisticated statistics over multiple time scales (Benjamin et al, 2023(Benjamin et al, , 2024Maheu et al, 2019;Planton et al, 2021;Saffran et al, 1999;Skerritt-Davis & Elhilali, 2019). Thus, further research is needed to probe the impact of more complex temporal statistical relationships when processing rapid auditory sequences.…”

“…Neural evidence shows that the brain rapidly segments the stream using these triplets, even under passive conditions (Batterink & Paller, 2017). The second structure, a community structure, is more intricate (Benjamin et al, 2023;Karuza et al, 2017Karuza et al, , 2019Lynn et al, 2020;Schapiro et al, 2013). Community structures have recently attracted attention due to their closer resemblance to the statistics of natural environments (Girvan & Newman, 2002).…”

“…In this community structure, each node is connected to an equal number of other nodes, but certain nodes are more interconnected, resulting in groups of nodes sharing contextual associations and being presented in close temporal proximity (Karuza et al, 2019;Schapiro et al, 2013). Critically, transitional probabilities at community boundaries are uniform and instead parsing sequences necessitates a model that includes the associations between both adjacent and non-adjacent 5` preceding items, weighting those associations based on the distance between items (Benjamin et al, 2023;Lynn et al, 2020).…”

The successor representation subserves hierarchical abstraction for goal-directed behavior

“…Interestingly, it has been shown that the behaviors of both humans and rats exhibit some of the limitations in decision making specifically associated with the successor representation (26,27), which can capture some human and rat learning effects better than typical model-free or model-based RL algorithms do (28). Additionally, the successor representation can explain how people segment events based on the graph topology of the events, capturing higher-order structure in the environment (29)(30)(31). However, to our knowledge no one has yet investigated whether the successor representation is leveraged for the purpose of extracting hierarchies relevant for goal-directed behavior.…”