Davinci the Dualist: The Mind-Body Divide in Large Language Models and in Human Learners

A large literature suggests that people are intuitive Dualists-they consider the mind ethereal, distinct from the body. Furthermore, Dualism emerges, in part, via learning (e.g., Barlev & Shtulman, 2021). Human learners, however, are also endowed with innate systems of core knowledge, and recent results suggest that core knowledge begets Dualism (Berent, 2023a; Berent et al., 2022). The resulting question, then, is whether the acquisition of Dualism requires core knowledge, or whether Dualism is learnable from experience alone, via domain-general mechanism. Since human learners are equipped with both systems, the evidence from humans cannot decide this question. Accordingly, here, we probe for a mind-body divide in Davinci-a large language model (LLM) that is devoid of core knowledge. We show that Davinci still leans towards Dualism, and that this bias increases systematically with the learner's inductive potential. Thus, davinci (which forms part of the GPT-3 suite) exhibits mild Dualist tendencies, whereas its descendent, text-davinci-003 (a GPT-3.5 model), shows a stronger bias. It selectively considers thoughts (epistemic states) as disembodied-as unlikely to show up in the body (in the brain). Unlike humans, GPT 3.5 categorically rejected the persistence of the psyche after death. Still, when probed about life, GPT 3.5 showed robust Dualist tendencies. These results demonstrate that the mind-body divide is partly learnable from experience. While results from LLMs cannot fully determine how humans acquire Dualism, they do place a higher burden of proof on nativist theories that trace Dualism to innate core cognition (Berent, 2023a; Berent et al., 2022).

Superstitious learning of abstract order from random reinforcement

Humans and other animals often infer spurious associations among unrelated events. However, such superstitious learning is usually accounted for by conditioned associations, raising the question of whether an animal could develop more complex cognitive structures independent of reinforcement. Here, we tasked monkeys with discovering the serial order of two pictorial sets: a "learnable" set in which the stimuli were implicitly ordered and monkeys were rewarded for choosing the higher-rank stimulus and an "unlearnable" set in which stimuli were unordered and feedback was random regardless of the choice. We replicated prior results that monkeys reliably learned the implicit order of the learnable set. Surprisingly, the monkeys behaved as though some ordering also existed in the unlearnable set, showing consistent choice preference that transferred to novel untrained pairs in this set, even under a preference-discouraging reward schedule that gave rewards more frequently to the stimulus that was selected less often. In simulations, a model-free reinforcement learning algorithm (Q-learning) displayed a degree of consistent ordering among the unlearnable set but, unlike the monkeys, failed to do so under the preference-discouraging reward schedule. Our results suggest that monkeys infer abstract structures from objectively random events using heuristics that extend beyond stimulus-outcome conditional learning to more cognitive model-based learning mechanisms.

Relationship Between Efficiency, Effectiveness, and Learnability of Home Connected Medical Device in Ambulatory Surgery

Background: Currently, usability assessments of home connected medical devices do not systematically take into account learnability metrics. In case of the Smart Angel device-designed for monitoring ambulatory surgery patients-users are trained at the hospital and have to use the device at home to monitor their health remotely. Objective: The aim of this study was to better understand the relationships between two metrics of usability-efficiency and effectiveness-and learnability of the Smart Angel device. Materials and Methods: Twenty-eight participants were trained in a simulated hospital (SimUsanté), and then we filmed the participant using the device three times. Between each session, the participant had to complete questionnaires (sociodemographic and health literacy). Results: The results of a between-subject analysis [χ²(2) = 18.969, p < 0.001] and a within-subject analysis [F(2.28) = 13.34, p < 0.001, η² = 0.35] showed that efficiency (manipulation time) significantly improved with learnability (number of sessions). Conversely, effectiveness (number of manipulation errors) stagnated over the three sessions with a between-subject analysis [F(2.75) = 1.628, p = 0.203], while the within-subject analysis revealed that users seemed to significantly correct their errors with the number of sessions [F(2.28) = 6, p = 0.005, η² = 0.19]. By analyzing the errors, we observed that some errors could appear at any time (e.g., moving during the measurement) and others were systematic (e.g., the wrist blood pressure on the right arm). Conclusions: While the "first attempt" at using a home medical device is a major revealing step for the study of manipulation errors, learnability is an equally useful metric to include in usability studies, as well as in the very definition itself of usability.

L2 Acquisition of a Complex Stress Pattern: UG-Constrained Learning Paths in Khalkha Mongolian

This paper examines second language (L2) acquisition of stress in Khalkha Mongolian, which is one of the few Default-to-Opposite Edge stress systems of the world, and as such, demonstrates "conflicting directionality" regarding stress assignment, resulting in the leftmost edge of a word being more prominent in certain words and the rightmost edge in certain others. Given the additional fact that the language exhibits Non-finality effects, and that, unlike English, codas are not moraic, its acquisition presents unique difficulties and challenges for English-speaking learners of the language. Many of these challenges potentially lead these learners to make Universal Grammar (UG)-unconstrained (but cognitively reasonable) assumptions about how the phonology of Mongolian works, especially since the learners do not have all the Mongolian data available to them all at once. The learning scenario here, thus, provides unique opportunities to investigate whether L2 phonologies are constrained by the options made available by UG. The findings of a semi-controlled production experiment indicate that although learners do not necessarily converge on the prosodic representations employed by native speakers of the L2 (i.e., footless intonational prominence, at least for the leftmost/default edge 'stress'), and although certain changes to the grammar are very difficult to implement, such as switching from moraic codas to non-moraic codas, the learners nevertheless demonstrate a stage-like behavior where each step exhibits the parameter settings employed by a natural language, one that is neither like the L2 nor the L1. Conversely, despite the input leading them to do so, learners do not entertain UG-unconstrained prosodic representations, such as End-Rule-Middle or End-Rule-Variable; End-Rule is set either to Right or Left, as is expected in a system constrained by the options made available by UG. We conclude that the hypothesis space for interlanguage phonologies is determined by UG.

Learnability for the Information Bottleneck

The Information Bottleneck (IB) method provides an insightful and principled approach for balancing compression and prediction for representation learning. The IB objective I(X;Z)−βI(Y;Z) employs a Lagrange multiplier β to tune this trade-off. However, in practice, not only is β chosen empirically without theoretical guidance, there is also a lack of theoretical understanding between β, learnability, the intrinsic nature of the dataset and model capacity. In this paper, we show that if β is improperly chosen, learning cannot happen—the trivial representation P(Z|X)=P(Z) becomes the global minimum of the IB objective. We show how this can be avoided, by identifying a sharp phase transition between the unlearnable and the learnable which arises as β is varied. This phase transition defines the concept of IB-Learnability. We prove several sufficient conditions for IB-Learnability, which provides theoretical guidance for choosing a good β. We further show that IB-learnability is determined by the largest confident, typical and imbalanced subset of the examples (the conspicuous subset), and discuss its relation with model capacity. We give practical algorithms to estimate the minimum β for a given dataset. We also empirically demonstrate our theoretical conditions with analyses of synthetic datasets, MNIST and CIFAR10.

Segmentability Differences Between Child-Directed and Adult-Directed Speech: A Systematic Test With an Ecologically Valid Corpus

Previous computational modeling suggests it is much easier to segment words from child-directed speech (CDS) than adult-directed speech (ADS). However, this conclusion is based on data collected in the laboratory, with CDS from play sessions and ADS between a parent and an experimenter, which may not be representative of ecologically collected CDS and ADS. Fully naturalistic ADS and CDS collected with a nonintrusive recording device as the child went about her day were analyzed with a diverse set of algorithms. The difference between registers was small compared to differences between algorithms; it reduced when corpora were matched, and it even reversed under some conditions. These results highlight the interest of studying learnability using naturalistic corpora and diverse algorithmic definitions.

Evaluating the Usability of Authoring Environments for Serious Games

Background. The EMERGO method and online platform enable the development and delivery of scenario-based serious games that foster students to acquire professional competence. One of the main goals of the platform is to provide a user-friendly authoring environment for creating virtual environments where students can perform authentic tasks. Aim. We present the findings of an in-depth qualitative case study of the platform's authoring environment and compare our findings on usability with those found for comparable environments in literature. Method. We carried out semi-structured interviews, with two experienced game developers who have authored a game for higher education, and a literature review of comparable environments. Findings. The analysis shows that the usability of the authoring environment is problematic, especially regarding understandability and learnability, which is in line with findings of comparable environments. Other findings are that authoring is well integrated with the EMERGO method and that functionality and reliability of the authoring environment are valued. Practical implications. The lessons learned are presented in the form of general guidelines to improve the understandability and learnability of authoring environments for serious games.

Structural Principles or Frequency of Use? An ERP Experiment on the Learnability of Consonant Clusters

Phonological knowledge of a language involves knowledge about which segments can be combined under what conditions. Languages vary in the quantity and quality of licensed combinations, in particular sequences of consonants, with Polish being a language with a large inventory of such combinations. The present paper reports on a two-session experiment in which Polish-speaking adult participants learned nonce words with final consonant clusters. The aim was to study the role of two factors which potentially play a role in the learning of phonotactic structures: the phonological principle of sonority (ordering sound segments within the syllable according to their inherent loudness) and the (non-) existence as a usage-based phenomenon. EEG responses in two different time windows (adversely to behavioral responses) show linguistic processing by native speakers of Polish to be sensitive to both distinctions, in spite of the fact that Polish is rich in sonority-violating clusters. In particular, a general learning effect in terms of an N400 effect was found which was demonstrated to be different for sonority-obeying clusters than for sonority-violating clusters. Furthermore, significant interactions of formedness and session, and of existence and session, demonstrate that both factors, the sonority principle and the frequency pattern, play a role in the learning process.

Multiple Grammars and the Logic of Learnability in Second Language Acquisition

The core notion of modern Universal Grammar is that language ability requires abstract representation in terms of hierarchy, movement operations, abstract features on words, and fixed mapping to meaning. These mental structures are a step toward integrating representational knowledge of all kinds into a larger model of cognitive psychology. Examining first and second language at once provides clues as to how abstractly we should represent this knowledge. The abstract nature of grammar allows both the formulation of many grammars and the possibility that a rule of one grammar could apply to another grammar. We argue that every language contains Multiple Grammars which may reflect different language families. We develop numerous examples of how the same abstract rules can apply in various languages and develop a theory of how language modules (case-marking, topicalization, and quantification) interact to predict L2 acquisition paths. In particular we show in depth how Germanic Verb-second operations, based on Verb-final structure, can apply in English. The argument is built around how and where V2 from German can apply in English, seeking to explain the crucial contrast: "nothing" yelled out Bill/(*)"nothing" yelled Bill out in terms of the necessary abstractness of the V2 rule.

Usability of clinical decision support system as a facilitator for learning the assistive technology adaptation process

The aim of this study was to evaluate the usability of Ontology Supported Computerized Assistive Technology Recommender (OSCAR), a Clinical Decision Support System (CDSS) for the assistive technology adaptation process, its impact on learning the matching process, and to determine the relationship between its usability and learnability. Two groups of expert and novice clinicians (total, n = 26) took part in this study. Each group filled out system usability scale (SUS) to evaluate OSCAR's usability. The novice group completed a learning questionnaire to assess OSCAR's effect on their ability to learn the matching process. Both groups rated OSCAR's usability as "very good", (M [SUS] = 80.7, SD = 11.6, median = 83.7) by the novices, and (M [SUS] = 81.2, SD = 6.8, median = 81.2) by the experts. The Mann-Whitney results indicated that no significant differences were found between the expert and novice groups in terms of OSCAR's usability. A significant positive correlation existed between the usability of OSCAR and the ability to learn the adaptation process (rs = 0.46, p = 0.04). Usability is an important factor in the acceptance of a system. The successful application of user-centered design principles during the development of OSCAR may serve as a case study that models the significant elements to be considered, theoretically and practically in developing other systems. Implications for Rehabilitation Creating a CDSS with a focus on its usability is an important factor for its acceptance by its users. Successful usability outcomes can impact the learning process of the subject matter in general, and the AT prescription process in particular. The successful application of User-Centered Design principles during the development of OSCAR may serve as a case study that models the significant elements to be considered, theoretically and practically. The study emphasizes the importance of close collaboration between the developers and the end users.