Chunking as a rational strategy for lossy data compression in visual working memory

Abstract

1 2The nature of capacity limits for visual working memory has been the 3 subject of an intense debate that has relied on models that assume items are 4 encoded independently. Here we propose that instead, similar features are jointly 5 encoded through a "chunking" process to optimize performance on visual working 6 memory tasks. We show that such chunking can: 1) facilitate performance 7 improvements for abstract capacity-limited systems, 2) be optimized through 8 reinforcement, 3) be implemented by center-surround dynamics, and 4) increase 9effective storage capacity at the expense of recall precision. Human performance 10 on a variant of a canonical working memory task demonstrated performance 11 advantages, precision detriments, inter-item dependencies, and trial-to-trial 12 behavioral adjustments diagnostic of performance optimization through center-13 surround chunking. Models incorporating center-surround chunking provided a 14 better quantitative description of human performance in our study as well as in a 15 meta-analytic dataset, and apparent differences in working memory capacity 16 across individuals were attributable to individual differences in the 17 implementation of chunking. Our results reveal a normative rationale for center-18 surround connectivity in working memory circuitry, call for re-evaluation of 19 memory performance differences that have previously been attributed to 20 differences in capacity, and support a more nuanced view of visual working 21 memory capacity limitations: strategic tradeoff between storage capacity and 22 memory precision through chunking contribute to flexible capacity limitations that 23 include both discrete and continuous aspects.

The bounded rationality of probability distortion

“…In contrast, BLO models a common mechanism underlying all probability distortion, where we identified two constraints-boundedness and compensation for representational uncertainty (variance)-that are pervasive in models of cognitive and perceptual tasks 37,44,66,67 . We found that BLO can be used to estimate an individual's probability distortion in one task and to some extent predict the same individual's performance in another task.…”

Structured event memory: a neuro-symbolic model of event cognition

“…Specifically, we assume an encoding process Pr(ỹ|y) creating the corrupted (encoded) memory traceỹ i = (x ′ ,ẽ ′ ,ñ) where ...,x d ],ẽ ′ andñ corresponds to the corrupted memory traces of the scene features, event label, and time index, respectively. The assumption that memory traces are corrupted versions of an original stimulus is common in computational models of memory (Hemmer & Steyvers, 2009;Huttenlocher et al, 1991;Shiffrin & Steyvers, 1997) and is analogous to a capacity-limited compression (Brady, Konkle, & Alvarez, 2009;Nassar, Helmers, & Frank, 2018).…”

Origin of perseveration in the trade-off between reward and complexity

“…To implement a policy computationally, we would need to describe it in some programming language, and the description length of that program (e.g., in bits or nats) imposes a demand on memory resources. Intuitively, if a policy can be "compressed" to a short description length, it will be easier to remember, much in the same way that the benefits of compression have been studied in memory for digit verbal and visual stimuli (Brady et al, 2009;Mathy and Feldman, 2012;Nassar et al, 2018). As we will formalize later, it turns out that perseveration arises naturally from the imperative to reduce policy complexity.…”