Cognitive Computation最新文献

Methods based on the physical Retinex model are effective in enhancing low-light images, adeptly handling the challenges posed by low signal-to-noise ratios and high noise in images captured under weak lighting conditions. However, traditional models based on manually designed Retinex priors do not adapt well to complex and varying degradation environments. DEANet (Jiang et al., Tsinghua Sci Technol. 2023;28(4):743–53 2023) combines frequency and Retinex to address the interference of high-frequency noise in low-light image restoration. Nonetheless, low-frequency noise still significantly impacts the restoration of low-light images. To overcome this issue, this paper integrates the physical Retinex model with deep learning to propose a joint network model, DEANet++, for enhancing low-light images. The model is divided into three modules: decomposition, enhancement, and adjustment. The decomposition module employs a data-driven approach based on Retinex theory to split the image; the enhancement module restores degradation and adjusts brightness in the decomposed images; and the adjustment module restores details and adjusts complex features in the enhanced images. Trained on the publicly available LOL dataset, DEANet++ not only surpasses the control group in both visual and quantitative aspects but also achieves superior results compared to other Retinex-based enhancement methods. Ablation studies and additional experiments highlight the importance of each component in this method.

From an unstructured text, relation extraction (RE) predicts semantic relationships between pairs of entities. The process of labeling tokens and phrases can be very expensive and require a great deal of time and effort. The low-resource relation extraction (LRE) problem comes into being and is challenging since there are only a limited number of annotated sentences available. Recent research has focused on minimizing the cross-entropy loss between pseudo labels and ground truth or on using external knowledge to make annotations for unlabeled data. Existing methods, however, fail to take into account the semantics of relation types and the information hidden within different relation groups. By drawing inspiration from the process of human interpretation of unstructured documents, we introduce a Template-based Contrastive Learning ( TempCL ). Through the use of template, we limit the model’s attention to the semantic information that is contained in a relation. Then, we employ a contrastive learning strategy using both group-wise and instance-wise perspectives to leverage shared semantic information within the same relation type to achieve a more coherent semantic representation. Particularly, the proposed group-wise contrastive learning minimizes the discrepancy between the template and original sentences in the same label group and maximizes the difference between those from separate label groups under limited annotation settings. Our experiment results on two public datasets show that our model TempCL achieves state-of-the-art results for low-resource relation extraction in comparison to baselines. The relative error reductions range from 0.68 to 1.32%. Our model encourages the feature to be aligned with both the original and template sentences. Using two contrastive losses, we exploit shared semantic information underlying sentences (both original and template) that have the same relation type. We demonstrate that our method reduces the noise caused by tokens that are unrelated and constrains the model’s attention to the tokens that are related.

Autism spectrum disorders (ASDs) pose complex challenges, characterized by atypical behaviors, sensory sensitivities, and difficulties in social interaction. Despite extensive research, their exact causes remain elusive, indicating a multifactorial interplay of genetic, environmental, and neurological factors. This complexity calls for innovative approaches to ASD understanding and management. Motivated by the need to address the nuanced and uncertain nature of ASD-related data, in this study, we introduce a novel hybrid model called rough spherical fuzzy bipolar soft sets (RSFBSSs) by integrating rough sets, spherical fuzzy sets, and bipolar soft sets, which accommodates imprecision inherent in clinical assessments. We build upon foundational concepts of RSFBSS theory, developing a comprehensive algorithm for uncertain multiple attribute decision-making (MADM). Leveraging this framework, we aim to assess ASD symptom severity in pediatric populations, considering diverse contributing factors to ASD pathogenesis. The RSFBSSs offer advantages over existing methodologies, providing a robust framework for handling complex ASD data. The algorithmic framework facilitates accurate and individualized assessments of ASD symptomatology. To validate our model’s efficacy, we conduct a comparative analysis with preexisting hybrid models, employing quantitative metrics and qualitative evaluations. Through this comprehensive evaluation, we demonstrate the superior performance and versatility of RSFBSSs, offering promising avenues for advancing ASD management.

Malicious apps use a variety of methods to spread infections, take over computers and/or IoT devices, and steal sensitive data. Several detection techniques have been proposed to counter these attacks. Despite the promising results of recent malware detection strategies, particularly those addressing evolving threats, inefficiencies persist due to potential inconsistency in both the generated malicious malware and the pre-specified detection rules, as well as their crisp decision-making process. In this paper, we propose to address these issues by (i) considering the detection rules generation process as a Bi-Level Optimization Problem, where a competition between two levels (an upper level and a lower one) produces a set of effective detection rules capable of detecting new variants of existing and even unseen malware patterns. This bi-level strategy is subtly inspired by natural evolutionary processes, where organisms adapt and evolve through continuous interaction and competition within their environments. Furthermore, (ii) we leverage the fundamentals of Rough Set Theory, which reflects cognitive decision-making processes, to assess the true nature of artificially generated malicious patterns. This involves retaining only the consistent malicious patterns and detection rules and categorizing these rules into a three-way decision framework comprising accept, abstain, and reject options. Our novel malware detection technique outperforms several state-of-the-art methods on various Android malware datasets, accurately predicting new apps with a 96.76% accuracy rate. Moreover, our approach is versatile and effective in detecting patterns applicable to a variety of cybersecurity threats.

Does the model demonstrate exceptional proficiency in “item counting,” “color recognition,” or other Fundamental Visual Comprehension Capability (FVCC)? There have been remarkable advancements in the field of multimodal, the pretrained general Vision Language Models exhibit strong performance across a range of intricate Visual Language (VL) tasks and Multimodal Large Language Models (MLLMs) emerge novel visual reasoning abilities from several examples. But models tend to encounter difficulties when confronted with texts supplemented with specific details by simple visual phrases. Moreover, there is a scarcity of datasets in sufficient quantity, variety, and composability to enable the evaluation of each FVCC using statistical metrics. Accordingly, we decomposed the complete VL task into 9 M simple Visual Phrase Triplets (VPTs) across 16 categories representing 16 distinct FVCCs from the structural scene graph. Then, we reconstructed a Multilevel Scene Graph (MLSG) for each image and introduced our unbiased, balanced, and binary Visual Phrase Entailment benchmark with 20 times the data volume of SNLI-VE. The benchmark consisted of three exams and evaluated the performance of 8 widely used VLM and 10 MLLMs respectively. The results demonstrate the performance of each model across 16 classes in FVCC, as well as their lower and upper limits under conditions of increased text complexity or unnoised image input. Finally, we enhanced the efficiency of MLLM and evoked their In-Context Learning characteristics by appending multiple VPT generated QA pairs of identical types to the conversation history without tuning. The proposed structural VPTs and MLSG data hold promise for facilitating future explorations on FVCC.

Generative Artificial Intelligence models are Artificial Intelligence models that generate new content based on a prompt or input. The output content can be in various forms, including text, images, and video. Metaverse refers to a virtual world where users can interact with each other, objects and events in an immersive, realistic, and dynamic manner. A critical and foremost step in realizing the Metaverse is content creation for its different realms. Given Metaverse’s need for enormous content, Generative AI is a perfect technology for content creation. This paper explores how Generative AI models can help fulfil the potential of the Metaverse by assisting in the design and production of various aspects of the Metaverse and attracting users not just by creating dynamic, interactive, and personalised content at scale but also by producing various revenue-generating opportunities for users and organisations in the Metaverse. The paper analyses the Generative AI models by grouping them according to the type of content they generate, namely text, image, video, 3D visual, audio, and gaming. Various use cases in the Metaverse are explored and listed according to each type of AI Generated Content (AIGC). This paper also presents several applications and scenarios where the mixture of different Generative AI (GAI) models benefits the Metaverse. Further, this paper also enumerates the limitations and challenges of Generative AI models and the areas of future work. Despite the obstacles, Generative AI can realise the potential of the Metaverse by making it much more functional and interactive owing to the vast use cases of different types of AIGC in the Metaverse, and the age of virtual reality may not be too distant.

In the text mining area, prevalent deep learning models primarily focus on mapping input features to result of predicted outputs, which exhibit a deficiency in self-dialectical thinking process. Inspired by self-reflective mechanisms in human cognition, we propose a hypothesis that existing models emulate decision-making processes and automatically rectify erroneous predictions. The Self-adaptive Reflection Enhanced pre-trained deep learning Model (S-REM) is introduced to validate our hypotheses and to determine the types of knowledge that warrant reproduction. Based on the pretrained-model, S-REM introduces the local explanation for pseudo-label and the global explanation for all labels as the explanation knowledge. The keyword knowledge from TF-IDF model is also integrated to form a reflection knowledge. Based on the key explanation features, the pretrained-model reflects on the initial decision by two reflection methods and optimizes the prediction of deep learning models. Experiments with local and global reflection variants of S-REM on two text mining tasks across four datasets, encompassing three public and one private dataset were conducted. The outcomes demonstrate the efficacy of our method in improving the accuracy of state-of-the-art deep learning models. Furthermore, the method can serve as a foundational step towards developing explainable through integration with various deep learning models.

Although deep neural networks have developed at a high level, the large computational requirement limits the deployment in end devices. To this end, a variety of model compression and acceleration techniques have been developed. Among these, knowledge distillation has emerged as a popular approach that involves training a small student model to mimic the performance of a larger teacher model. However, the student architectures used in existing knowledge distillation are not optimal and always have redundancy, which raises questions about the validity of this assumption in practice. This study aims to investigate this assumption and empirically demonstrate that student models could contain redundancy, which can be removed through pruning without significant performance degradation. Therefore, we propose a novel pruning method to eliminate redundancy in student models. Instead of using traditional post-training pruning methods, we perform pruning during knowledge distillation (PDD) to prevent any loss of important information from the teacher models to the student models. This is achieved by designing a differentiable mask for each convolutional layer, which can dynamically adjust the channels to be pruned based on the loss. Experimental results show that with ResNet20 as the student model and ResNet56 as the teacher model, a 39.53%-FLOPs reduction was achieved by removing 32.77% of parameters, while the top-1 accuracy on CIFAR10 increased by 0.17%. With VGG11 as the student model and VGG16 as the teacher model, a 74.96%-FLOPs reduction was achieved by removing 76.43% of parameters, with only a loss of 1.34% in the top-1 accuracy on CIFAR10. Our code is available at https://github.com/YihangZhou0424/PDD-Pruning-during-distillation.

Multimodal models are designed to process different types of data within a single generative framework. The prevalent strategy in previous methods involves learning joint representations that are shared across different modalities. These joint representations are typically obtained by concatenating the top of layers of modality-specific networks. Recently, significant advancements have been made in generating images from text and vice versa. Despite these successes, current models often overlook the role of fuzzy concepts, which are crucial given that human cognitive processes inherently involve a high degree of fuzziness. Recognizing and incorporating fuzzy concepts is therefore essential for enhancing the effectiveness of multimodal cognition models. In this paper, a novel framework, named the Fuzzy Concept Learning Model (FCLM), is proposed to process modalities based on fuzzy concepts. The high-level abstractions between different modalities in the FCLM are represented by the ‘fuzzy concept functions.’ After training, the FCLM is capable of generating images from attribute descriptions and inferring the attributes of input images. Additionally, it can formulate fuzzy concepts at various levels of abstraction. Extensive experiments were conducted on the dSprites and 3D Chairs datasets. Both qualitative and quantitative results from these experiments demonstrate the effectiveness and efficiency of the proposed framework. The FCLM integrates the fuzzy cognitive mechanism with the statistical characteristics of the environment. This innovative cognition-inspired framework offers a novel perspective for processing multimodal information.

Commonsense knowledge acquisition and representation is a core topic in artificial intelligence (AI), which is crucial for building more sophisticated and human-like AI systems. However, existing commonsense knowledge bases organize facts in an isolated manner like bag of facts, lacking the cognitive-level connections that humans commonly possess. People have the ability to efficiently organize vast amounts of knowledge by linking or generalizing concepts using a limited set of conceptual primitives that serve as the fundamental building blocks of reasoning. These conceptual primitives are basic, foundational elements of thought that humans use to make sense of the world. By combining and recombining these primitives, people can construct complex ideas, solve problems, and understand new concepts. To emulate this cognitive mechanism, we design a new commonsense knowledge base, termed PrimeNet, organized in a three-layer structure: a small core of conceptual primitives (e.g., FOOD), a bigger set of concepts that connect to such primitives (e.g., fruit), and an even larger layer of entities connecting to the concepts (e.g., banana). First, we collect commonsense knowledge and employ a gradual expansion strategy for knowledge integration. After refinement, PrimeNet contains 6 million edges between 2 million nodes, with 34 different types of relations. Then, we design a new conceptualization method by leveraging a probabilistic taxonomy, to build the concept layer of PrimeNet. Finally, we conduct primitive detection to build the primitive layer, where a lexical substitution task is used to identify related concepts, and large language models are employed to generate a rational primitive to label each concept cluster as well as verify the primitive detection process.