Cognitive Systems Research最新文献

Researchers studying the correspondences between Deep Neural Networks (DNNs) and humans often give little consideration to severe testing when drawing conclusions from empirical findings, and this is impeding progress in building better models of minds. We first detail what we mean by severe testing and highlight how this is especially important when working with opaque models with many free parameters that may solve a given task in multiple different ways. Second, we provide multiple examples of researchers making strong claims regarding DNN-human similarities without engaging in severe testing of their hypotheses. Third, we consider why severe testing is undervalued. We provide evidence that part of the fault lies with the review process. There is now a widespread appreciation in many areas of science that a bias for publishing positive results (among other practices) is leading to a credibility crisis, but there seems less awareness of the problem here.

The rapid development and widespread adoption of Artificial Intelligence (AI) technologies have made the development of AI-specific design methods an important topic to advance. In recent decades, the centre of gravity in AI has shifted away from cognitive science and related fields like psychology. However, there is a clear need and potential for added value in returning to stronger interaction. One potential challenge for this interaction may be the lack of common conceptual grounds and design languages.

In this article, we aim to contribute to the development of conceptual interfaces for human-based AI-specific design methods through the idea of cognitive mimetics. We begin by introducing basic concepts from mimetic design and interpret them in the context of this thematic area. These provide some of the basic building blocks for a design language and bring to the surface key questions. These in turn provide a ground for explicating cognitive mimetics. In the second part of this paper, we focus on specifying a key aspect in cognitive mimetics: the contents of information processes.

Others engaged in this field can derive value from using or developing the basic conceptual machinery to specify their own approaches in this interdisciplinary field that is still shaping itself. Furthermore, those who resonate with the idea of cognitive mimetics, as specified here, can join in taking this particular approach further.

The impressive recent performance of large language models has led many to wonder to what extent they can serve as models of general intelligence or are similar to human cognition. We address this issue by applying GPT-3.5 and GPT-4 to a classic problem in human inductive reasoning known as property induction. Over two experiments, we elicit human judgments on a range of property induction tasks spanning multiple domains. Although GPT-3.5 struggles to capture many aspects of human behavior, GPT-4 is much more successful: for the most part, its performance qualitatively matches that of humans, and the only notable exception is its failure to capture the phenomenon of premise non-monotonicity. Our work demonstrates that property induction allows for interesting comparisons between human and machine intelligence and provides two large datasets that can serve as benchmarks for future work in this vein.

Neuroscientists apply a range of analysis tools to recorded neural activity in order to glean insights into how neural circuits drive behavior in organisms. Despite the fact that these tools shape the progress of the field as a whole, we have little empirical proof that they are effective at identifying the mechanisms of interest. At the same time, deep learning systems are trained to produce intelligent behavior using neural networks, and the resulting models are impressive but also largely impenetrable. Can the tools of neuroscience be applied to artificial neural networks (ANNs) and if so what would this process tell us about ANNs, brains, and – most importantly – the tools themselves? Here we argue that applying analysis methods from neuroscience to ANNs will provide a much-needed test of the abilities of these tools. It would also encourage the development of a unified field of neural systems understanding, which can identify shared concepts and methods for studying distributed information processing in artificial and biological systems. To support this argument, we review methods commonly used in neuroscience, along with work that has demonstrated how these methods can be applied to ANNs and what we learn from this, and related efforts from interpretable AI.

Perceiving human emotions is crucial in the realm of affective computing. As a nonverbal biological feature, gait plays a significant role in this field, owing to its resistance to manipulation or replication. In this paper, we propose a gait-based emotion perception framework called Dependency-Difference Gait (DDG), which can extract emotional features from gait patterns comprehensively and efficiently. We also introduce a method of spatial–temporal difference representation, which constructs the static spatial difference information within frames and dynamic temporal difference information between frames. We abstract these details as difference information and fuse them with the dependency information extracted from the original sequence. Our approach not only breaks the limitations of hand-crafted features, but also enables the extraction of a broader spectrum of emotional features. Additionally, we present the Emotional Information Attention (EIA) mechanism, allowing DDG to focus on key joints and frames based on the quantity of emotional information. Experimental and visualization results substantiate the effectiveness of the DDG and EIA. In the quality analysis, we find that selecting a few number of joints with a substantial amount of emotional information is beneficial for emotion classification. However, selecting a few frames can disrupt the temporal structure of the sequence, resulting in suboptimal performance.

Linguistic experience varies across individuals and is impacted by both demography and personal preferences, leading to differences in word meanings across languages (Thompson et al., 2020) and people (Johns, 2022). An active area of study in the cognitive sciences that examines the impact of varied knowledge across individuals is the wisdom of the crowd effect, where it is found that the aggregate judgement of a group of individuals is often better than the judgement of the best individual in the group (Surowiecki, 2004). The goal of this article was to determine if there is a wisdom of the crowd effect in lexical semantic memory, such that the aggregated word similarity values from many individual language users exceeds the fit of the best fitting individual. This was accomplished by training 500 different distributional models from 500 high-level commenters on the internet forum Reddit. By deriving aggregated word similarity values from these individuals, a strong wisdom of the crowd effect was found where the aggregated similarity values far exceeded the performance of the best fitting individual for each dataset tested. Additionally, it was found that even aggregating only a small number of users provided a large increase in fit relative to the individual corpora, but with the best fitting measure including word similarity values from all possible users. The results of this article provide an avenue for future distributional model development by demonstrating that the best pathway towards better distributional models may lie in the aggregation of multiple representations attained from individual users of a language.

In this work we investigate the phenomenon of emotional mirroring using remotely diagnosed dynamic parameters of facial expressions. The research is based on the fact that mirroring is the subconscious adjustment and copying of the dynamics of another person. We considered a reflection of face expression as a reproduction of emotions of one person by another. To obtain this behavior we used an induced cognitive–emotional conflict in the process of telecommunication dialogue. The conflict was initiated by a psychologist or by short videoclips with surprise endings. Since the communication in a telecommunication form limits non-verbal information about the interlocutor with respect to the normal dialogue, we have also investigated the hypothesis of whether the phenomenon of mirroring is detectable in such conditions. We developed a computer program using VGG16-based artificial neural network to mark people’s emotional reactions in video data automatically. The processed material consisted of 24 interview recordings with the participants of both genders and three qualified expert psychologists. We used different types of interviews: interviews based on self-attitude techniques, problematic interviews based on transactional analysis, free reasoning about controversial and topical situations. The communication topics were selected with respect to the age and other indicators of the group of participants. It was found that the parameters of facial expressions of the participant and the experimenter (psychologist) identified by the program strongly correlate with emotions such as happiness, sadness and surprise. Notable negative correlations were found between the parameters of the happiness of participant and fear of psychologist, sad of the participant and happiness of the psychologist, sad of psychologist and surprise of the participant. A direct relationship between sad of participant and fear of psychologist was detected. All of the identified correlations appear both in the situation with and without cognitive–emotional conflict. However, the degree of their manifestation was quite different for these two cases.

This paper describes an extension of a safety culture within hospital organizations providing more transparency and acknowledgement of all actors, and in particular the parents. It contributes a model architecture to support a hospital to develop such an extended safety culture. It is illustrated for prevention of postpartum depression. Postpartum depression is a commonly known consequence of childbirth for both mothers and fathers. In this research, we computationally analyze the risk factors and lack of support received by fathers. Therefore, we use shared mental models to model the effects of poor and additional communication by healthcare practitioners to mitigate the development of postpartum depression in both the mother and the father. Both individual mental models and shared mental models are considered in the design of the computational model. The paper illustrates the benefits of simple support in terms of communication during childbirth, which has lasting effects, even outside the hospital. For the impact of additional communication, a Virtual Safety Coach is designed that intervenes when necessary to provide support, i.e., when a health care practitioner doesn’t. Moreover, organizational learning is also modelled to improve the mental models of both the Safety Coach and the Health Care Practitioner.

This paper introduces a novel controlled adaptive mental causal network model addressing how dreams overnight can influence creativity in waking life. The network model depicts in a causal, dynamic, and generic manner which adaptive mental processes underlie the connection between dreams and creativity and is shown to be validated with the existing cognitive neuroscience literature.

In a 2009 paper, Dussutour et al. proposed that big headed ants (Pheidole megacephala) employ two attractant pheromones during foraging: one for exploration and another during food gathering. This claim was consistent with, and argued to be supported by, laboratory studies of ant exploration and food-gathering in a Y-maze apparatus. The authors measured foraging activity and colony foraging choice in terms of the number of ants choosing different branches over time, where experimental conditions modified the history of food availability at the end of each branch. They built a two-pheromone mathematical model to account for observed rates and proportions of ants traversing the left versus right branch. Here we show that the main reported experimental observations can be explained by a one-pheromone model. Our findings show that it is plausible, but unnecessary, to hypothesize that these ants employ two distinct pheromones in order to account for the two principal results of the Dussutour et al. study, and therefore, the study falls short of dispositive evidence for a two-pheromone model. More broadly, we highlight that patterns of animal behavior can be ambiguous with respect to sensory and cognitive mechanisms, hopefully motivating future modeling efforts that perform formal comparison across models with different structure.