Mental compression of binary sequences in a language of thought

Abstract

The capacity to store information in working memory strongly depends upon the ability to recode the information in a compressed form. Here, we tested the theory that human adults encode binary sequences of stimuli in memory using a recursive compression algorithm akin to a “language of thought”, and capable of capturing nested patterns of repetitions and alternations. In five experiments, we probed memory for auditory or visual sequences using both subjective and objective measures. We used a sequence violation paradigm in which participants detected occasional violations in an otherwise fixed sequence. Both subjective ratings of complexity and objective sequence violation detection rates were well predicted by complexity, as measured by minimal description length (also known as Kolmogorov complexity) in the binary version of the “language of geometry”, a formal language previously found to account for the human encoding of complex spatial sequences in the proposed language. We contrasted the language model with a model based solely on surprise given the stimulus transition probabilities. While both models accounted for variance in the data, the language model dominated over the transition probability model for long sequences (with a number of elements far exceeding the limits of working memory). We use model comparison to show that the minimal description length in a recursive language provides a better fit than a variety of previous encoding models for sequences. The data support the hypothesis that, beyond the extraction of statistical knowledge, human sequence coding relies on an internal compression using language-like nested structures.

Rational arbitration between statistics and rules in human sequence processing

“…AB-AABB-AAABBB… = A n B n where n increases sequentially). A possibility, which has received some behavioural support, is that the human brain uses a “language of thought”, including such algebraic patterns, in order to compress deterministic sequences into a regular expression 29,35,74–76 . For simplicity, we manipulated here only repeated patterns of various lengths, but found, in line with a possible role of compression, that the barely compressible length-10 pattern AAABAABBAB was more often missed by subjects compared to the other two length-10 patterns, AABABABABB and AAAAABBBBB, which are more compressible ([A 2 , (BA) 2 , B 2 ] and [A 4 , B 4 ] respectively).…”

“…For simplicity, we manipulated here only repeated patterns of various lengths, but found, in line with a possible role of compression, that the barely compressible length-10 pattern AAABAABBAB was more often missed by subjects compared to the other two length-10 patterns, AABABABABB and AAAAABBBBB, which are more compressible ([A 2 , (BA) 2 , B 2 ] and [A 4 , B 4 ] respectively). Note, however, that this effect could also result from the difference in the patterns’ apparent statistical bias (which is respectively weak and strong), an alternative we have recently considered and discarded 76 .…”

“…which type(s) of rules) that is relevant for the brain, our model makes the general claim that there exist fundamentally two categories of regularity, statistics and rules, which afford respectively uncertain vs. sure predictions. The model also posits that the statistical bias and deterministic rule hypotheses are evaluated in parallel during sequence processing, leading to testable predictions 76 . The model also posits that humans can rationally compare those different hypotheses using a common probabilistic currency.…”

Mental compression of spatial sequences in human working memory using numerical and geometrical primitives

“…Furthermore, very recently, the same formal language of geometry that we tested here was also found to successfully capture the regularities in binary auditory and visual sequences made of only two arbitrary sounds or pictures (Planton et al, 2020). are such codes uniquely developed in human, as postulated by some researchers Fitch, 2014;Hauser et al, 2002), or can they also be observed in non-human primates?…”

“…The present research adds further evidence in favor of this framework. Furthermore, very recently, the same formal language of geometry that we tested here was also found to successfully capture the regularities in binary auditory and visual sequences made of only two arbitrary sounds or pictures (Planton et al, 2020). Future research should examine three key open questions.…”