Special Issue "What Limits Working Memory Performance?"

A special issue of Entropy (ISSN 1099-4300). This special issue belongs to the section "Information Theory, Probability and Statistics".

Deadline for manuscript submissions: 31 October 2020.

Special Issue Editors

Prof. Dr. Alessandro Treves
Website
Guest Editor
SISSA- Cognitive Neuroscience, Via Bonomea 265, I-34136 Trieste, Italy
Interests: neural computation; memory; brain organization
Dr. Yair Lakretz
Website
Guest Editor
Cognitive Neuroimaging Unit NeuroSpin center 91191, Gif-sur-Yvette, France
Interests: natural language processing; neural computation; machine learning

Special Issue Information

Dear Colleauges,

Common wisdom tells us that working memory is severely limited in capacity—for example, the “magical” number seven for digit span; perhaps because of hard biophysical constraints—as suggested by the typical few seconds of retention time for verbal material. Experimental evidence is however complex, and is in complex relation to information theory. George Miller noted that while humans can typically convey only about log2(7) bits in unidimensional judgements, our short-term memory span can be much longer, if information is organized in chunks. Venerable mnemonic techniques, like the method of loci, can help us to train ourselves to recode and reach well beyond our naive short-term information capacity. So, is a general information-theoretic account of working memory possible? How constrained would it be by cortical circuitry? Any theoretical and theory-framed experimental contribution to these questions is welcome to the SI, including evidence obtained in animal studies or with the simulation of plausible memory networks.

Prof. Dr. Alessandro Treves
Dr. Yair Lakretz
Guest Editors

Manuscript Submission Information

Manuscripts should be submitted online at www.mdpi.com by registering and logging in to this website. Once you are registered, click here to go to the submission form. Manuscripts can be submitted until the deadline. All papers will be peer-reviewed. Accepted papers will be published continuously in the journal (as soon as accepted) and will be listed together on the special issue website. Research articles, review articles as well as short communications are invited. For planned papers, a title and short abstract (about 100 words) can be sent to the Editorial Office for announcement on this website.

Submitted manuscripts should not have been published previously, nor be under consideration for publication elsewhere (except conference proceedings papers). All manuscripts are thoroughly refereed through a single-blind peer-review process. A guide for authors and other relevant information for submission of manuscripts is available on the Instructions for Authors page. Entropy is an international peer-reviewed open access monthly journal published by MDPI.

Please visit the Instructions for Authors page before submitting a manuscript. The Article Processing Charge (APC) for publication in this open access journal is 1600 CHF (Swiss Francs). Submitted papers should be well formatted and use good English. Authors may use MDPI's English editing service prior to publication or during author revisions.

Keywords

  • phonological output buffer
  • information bottleneck
  • short-term plasticity
  • long-range dependencies
  • visuospatial sketchpad
  • articulatory loop

Published Papers (2 papers)

Order results
Result details
Select all
Export citation of selected articles as:

Research

Jump to: Other

Open AccessArticle
Working Memory Training: Assessing the Efficiency of Mnemonic Strategies
Entropy 2020, 22(5), 577; https://doi.org/10.3390/e22050577 - 20 May 2020
Abstract
Recently, there has been increasing interest in techniques for enhancing working memory (WM), casting a new light on the classical picture of a rigid system. One reason is that WM performance has been associated with intelligence and reasoning, while its impairment showed correlations [...] Read more.
Recently, there has been increasing interest in techniques for enhancing working memory (WM), casting a new light on the classical picture of a rigid system. One reason is that WM performance has been associated with intelligence and reasoning, while its impairment showed correlations with cognitive deficits, hence the possibility of training it is highly appealing. However, results on WM changes following training are controversial, leaving it unclear whether it can really be potentiated. This study aims at assessing changes in WM performance by comparing it with and without training by a professional mnemonist. Two groups, experimental and control, participated in the study, organized in two phases. In the morning, both groups were familiarized with stimuli through an N-back task, and then attended a 2-hour lecture. For the experimental group, the lecture, given by the mnemonist, introduced memory encoding techniques; for the control group, it was a standard academic lecture about memory systems. In the afternoon, both groups were administered five tests, in which they had to remember the position of 16 items, when asked in random order. The results show much better performance in trained subjects, indicating the need to consider such possibility of enhancement, alongside general information-theoretic constraints, when theorizing about WM span. Full article
(This article belongs to the Special Issue What Limits Working Memory Performance?)
Show Figures

Figure 1

Other

Jump to: Research

Open AccessOpinion
What Limits Our Capacity to Process Nested Long-Range Dependencies in Sentence Comprehension?
Entropy 2020, 22(4), 446; https://doi.org/10.3390/e22040446 - 16 Apr 2020
Abstract
Sentence comprehension requires inferring, from a sequence of words, the structure of syntactic relationships that bind these words into a semantic representation. Our limited ability to build some specific syntactic structures, such as nested center-embedded clauses (e.g., “The dog that the cat that [...] Read more.
Sentence comprehension requires inferring, from a sequence of words, the structure of syntactic relationships that bind these words into a semantic representation. Our limited ability to build some specific syntactic structures, such as nested center-embedded clauses (e.g., “The dog that the cat that the mouse bit chased ran away”), suggests a striking capacity limitation of sentence processing, and thus offers a window to understand how the human brain processes sentences. Here, we review the main hypotheses proposed in psycholinguistics to explain such capacity limitation. We then introduce an alternative approach, derived from our recent work on artificial neural networks optimized for language modeling, and predict that capacity limitation derives from the emergence of sparse and feature-specific syntactic units. Unlike psycholinguistic theories, our neural network-based framework provides precise capacity-limit predictions without making any a priori assumptions about the form of the grammar or parser. Finally, we discuss how our framework may clarify the mechanistic underpinning of language processing and its limitations in the human brain. Full article
(This article belongs to the Special Issue What Limits Working Memory Performance?)
Show Figures

Figure 1

Planned Papers

The below list represents only planned manuscripts. Some of these manuscripts have not been received by the Editorial Office yet. Papers submitted to MDPI journals are subject to peer-review.

Title: Working memory training: assessing efficiency of mnemonist strategies
Authors: Serena Di Santo; Vanni De Luca; Sara Andreetta
Affiliation: 1. SISSA - Cognitive Neuroscience, Trieste 2. Scuola Peripatetica d’Arte Mnemonica
Abstract: In recent years there has been increasing interest in exploring techniques to enhance working memory performance. One reason is that working memory (WM) capacity has been associated with intelligence and reasoning (Conway et al., 2003), while its limited performance seems to be correlated with cognitive deficits, such as attention-deficit hyperactivity disorder (Martinussen et al., 2005), among others. On a theoretical level, demonstrating a significant effect of training challenges rigid information-theoretic or biophysical accounts of what determines the surprisingly low working memory capacity. However, results on WM performance following specific training are controversial. Therefore, the question whether WM can be trained or not, has not been satisfactorily answered yet. The main issue is methodological: studies that describe WM training often lack theory-driven, rigorous systematic approaches (von Bastian & Oberauer, 2013, Shipstead et al., 2012). This study aims at assessing changes in WM performance by comparing it before and after training administered by a professional mnemonist. In the paradigm, participants have been divided into two groups: experimental and control. The experiment is organized in two phases: in the morning, both groups familiarize with a set of stimuli, by performing a simple n-back task, and then listen to a 2-hour lecture. For the experimental group, it is a lecture by the mnemonist, introducing memory encoding techniques; for the control group, it is a standard academic lecture about memory systems in the brain. In the afternoon, both groups are administered 5 tests, in each of which they have to remember the position of 16 items, when asked in random order. In the first two tests the items are presented sequentially, auditorily in the test with words (SW) and visually in the one with images (SI). In the 3 later tests, items are instead presented simultaneously in a 4x4 grid, and they are digits from 1 to 9 (obviously with repetitions, GD), written words (GW) or images (GI) from the same set as before. When visually interrogated about e.g. the 13th word or the first image in the third row, subjects can pick the answer from a pool of 48 options, including the 16 actually used; or from the 9 digits. From an information-theoretic point of view, therefore, 4 of the tasks are broadly equivalent, while the digit task, GD, differs from the others mainly because retrieving each item amounts to providing only log2(9) bits, or roughly 60% than those in the other tests. The spatio-temporal and qualitative features of the tasks, however, are distinct. Results show that control subjects vary significantly in their average performance, from about 4 correctly retrieved items in test SW to about 9 (SI), 12 (GD) and 11 (GW and GI). Ad hoc trained subjects were on average much better, from almost 14 correctly retrieved items in test SW to about 15 in the others, in which over half the trained participants did retrieve all items correctly. The specific training, therefore, was markedly effective in bringing participants close to ceiling performance across tests or, in a different interpretation, in largely removing the differential difficulty of the tests due to their spatio-temporal and qualitative features. These results thus indicate the need to consider such features, alongside general information-theoretic quantities, when analyzing working memory span.

Title: Professional or amateur? The phonological output buffer as a working memory operator
Authors: Neta Haluts; Massimiliano Trippa; Naama Friedmann; Alessandro Treves
Affiliation: 1. Sagol School of Neuroscience and School of Education, Tel Aviv University 2. SISSA - Cognitive Neuroscience, Trieste
Abstract: The Phonological Output Buffer (POB) is a conceptual construct which appears necessary to explain patterns of deficits observed in certain linguistic tasks, such as reading aloud. The occurrence in some subjects of phoneme omission, addition, replacement and misplacement, when producing words, reading or repeating words or non-words, with concurrent preservation of the lexicon and of grapheme-to-phoneme conversion, leads to posit a stage downstream, or a device, where phonemes are assembled in the correct order, perhaps modified as required, and kept in short-term memory (STM) until the word or non-word is uttered by the subject. It is tempting to think of the POB as a specialized device, located somewhere in the brain, which has evolved to facilitate the human language faculty. The notion of an ad hoc phoneme assembly line is challenged, however, by the occurrence of similar deficits in the sequential production of other linguistic objects, including numbers, morphemes, function words. Here we review an argument for the generality of the neural mechanisms underlying the POB, which comes from a different perspective: that of a Sign Language. The analysis of errors in the production and repetition of signed strings, by signers with a specific impairment as well as by control subjects, leads to posit a nearly identical downstream buffer, operating on objects with different statistics. Thus syllables are effectively conflated with words, and the features defining phonemes, and reflected in POB malfunction, such as place of articulation/obstruction, manner and voicing, are replaced by sign features such as handshape, movement, place of articulation/movement. In information-theoretic terms one may ask, given that the capacity of spoken and sign languages to convey bits of meaning per second appears similar, as corroborated by the practice of simultaneous translation, whether and in what sense one code may be more redundant or less error-prone than the other. The average production rate of syllables varies across spoken languages around 6/sec, anticorrelated with the average entropy per syllable, to convey overall around 35-40 bits/sec. Thus the POB is expected to churn out in the order of two 3-syllabic words per second. Interestingly, about two per second is also an average observed production rate for sign languages. The average sign “span” can also be measured and turns out to be 4-5 signs, lower but similar to the average 3-syllable word span, leading to the idea that a POB should normally operate correctly for at least 2.5sec, whatever the material it operates on. The homology and the learnability of the two buffers suggests that their fundamental characteristics are not particular to the objects they operate on, but reflect general entropic/statistical constraints and the neural wetware in which they are implemented. Beyond the overall similarity, however, the efficiency of the spoken and signed code may be quantitatively different – it has been suggested that handshapes come much closer to a maximum entropy code than phonemes, for example – and the required dynamics is qualitatively different, - due to the largely simultaneous rather than sequential production of the constituents of a signed word. We therefore sketch an analysis of these constraints, using a Potts network model of cortical dynamics, a portion of which is taken to represent the objects held in the output buffer. We focus on the production of 3-syllable words and of signs with 3 simultaneous elements, such as handshape, place and movement, making simple assumptions about their statistics. A fundamental constraint arises from the imprecision and the ill-determined time course of available neural STM mechanisms. In a network not designed ad hoc, this constraint leads, as the model indicates, to marked sensitivity to the degree of overlap between the representations of the objects in memory. A sensitivity that might explain the differential effects of training in distinct working memory systems, including the output buffers.

Title: Gating mechanisms in the brain for structure-sensitive language processing
Authors: Yair Lakretz; Jean-Rémi King
Affiliation: 1. UNICOG NeuroSpin, Paris, France 2. FaceBook AI Research; CNRS, Paris, France
Abstract: Natural language comprises complex structures whose processing requires structure-sensitivity. This processing occurs at a wide variety of time scales: ranging from the millisecond level - combining phonemes into distinctive words, to multi-second time scales - combining words into sentences, and higher time scales for narration. A central example of these linguistic structures is that of long-range dependencies between words. These are found in various linguistic phenomena such as subject-verb agreement (e.g., 'The keys to the cabinet are..') and reflexive anaphora (e.g., ‘The men that the child watches shave themselves.’). Online processing of long-range dependencies requires the storage and retrieval of various aspects of the linguistic input in and from short-term memory, such as carrying the grammatical number of the subject in the above examples (in bold) across interfering nouns (underscored). To this day, the possible mechanisms of such linguistic representations in the brain remain largely unknown. Recent advances in natural language processing in artificial recurrent neural networks offer compelling suggestions about how linguistic structures could be processed to incrementally generate the meaning of continuously unfolding sentences. To test whether these computational models offer grounds for a biological theory of language processing in the human brain, we review the related computational, cognitive and neuroscientific literatures. We discuss the functional advantages and limits of various types of (1) gating mechanisms (e.g., differential versus multiplicative), and (2) different forms of working-memory storage (i.e., silent versus explicit), and review their potential cellular, meso- and macroscopic neuronal bases. Overall, these empirical and theoretical findings highlight key challenges that need to be addressed to understand the biological underpinnings of structure-sensitive processing in the brain.

Back to TopTop