Human object recognition is robust to challenging conditions, such as when one's view of an object is fragmented due to an occluding foreground object. In comparison, deep neural networks (DNNs) are typically more susceptible to occlusion, suggesting that human vision relies on distinct mechanisms. Here, we investigated the role of visual diet in the emergence of these mechanisms by asking whether human-like robustness might arise in DNNs when trained with image datasets that better reflect the properties of occlusion in natural vision. We trained convolutional and transformer DNNs to classify clear images only, images augmented with artificial occluders (i.e., geometric shapes) or natural occluders (objects segmented from photographs). We then evaluated DNN occlusion robustness and compared their performance profiles with 30 human participants. We found that DNNs trained with artificial occluders remained vulnerable to natural occlusion and exhibited less human-like performance than those trained with natural occlusion. Our findings suggest that human robustness to visual occlusion arises from learning to disentangle natural objects from each other rather than simply learning to recognize objects from partial views. They also imply that commonly used forms of artificial occlusion are unsuitable for the evaluation or promotion of robustness to real-world occlusion in DNNs.
HLM-MOCAP is a motion capture dataset of context-dependent human arm movements designed to support the generation of human-like robot trajectories in industrial and collaborative settings. Data were recorded in a laboratory environment using an eight-camera infrared motion capture system (Vicon MX T10) operating at 200 Hz with 1120 × 896 px resolution and passive 1.5 mm retroreflective markers. Forty healthy adults (29 male, 11 female; broad age range from late adolescence to adulthood) performed a set of seven upper-limb tasks at a standardized workstation with marked start and target locations and physical objects of varying weight and size. The tasks comprise point-to-point transport to different distances and directions, sequential reaching to multiple targets, two trajectory-tracing tasks (zigzag and circle) with the index finger, a grasp-and-place task with arm crossing, a reaching task with load, and a high-precision placement task using small screws. Each participant completed 54 movement executions (with task-specific repetition counts), yielding 2160 recorded movement trials in total. The dataset includes raw 3D marker trajectories for all trials, segmented to active movement phases based on an objective velocity threshold, and processed versions after gap filling of short occlusions, time normalization to 0-100 % movement progress, and Savitzky-Golay filtering for derivation of stable velocity and acceleration profiles. Data are organized as comma-separated files with participant identifiers, task labels, target indices, object properties, and repetition indices, accompanied by scripts for loading and basic preprocessing. The resource can be reused for developing and benchmarking methods for human-like robot motion planning, learning from demonstration, and trajectory generation, as well as for analyzing how distance, sequence structure, load, and precision demands shape human arm kinematics in the context of human-robot collaboration.
While multimodal fake news detection methods have made progress in aligning multimodal semantics, they still face significant challenges in analyzing background context, emotional tone, and the overall plausibility of news content. To address these limitations, we propose a novel human-like collaborative framework for multimodal fake news detection, which integrates large and small models. Specifically, we exploit large vision-language models (LVLMs) to perform deep semantic analysis and reflective summarization of news cues. By leveraging the contextual understanding, knowledge recall, and logical reasoning capabilities of large models, the proposed approach improves the accuracy and reliability of fake news detection. It comprises three key components: 1) designing a chain-of-thought (CoT) prompting strategy for the LVLM to analyze news content, including evaluating image credibility, identifying potential tampering, extracting linguistic styles, detecting emotional tones, uncovering logical connections within the text, and verifying factual accuracy; 2) independently reflecting on and summarizing the lengthy analytical outputs from both image and text modalities to reduce redundancy. The resulting summary is then encoded into compact representations using pretrained text encoders and integrated with the original multimodal features; and 3) proposing a progressive fusion mechanism that enables collaboration between large and small models, allowing effective utilization of deeply fused features at the surface level. Extensive experiments conducted on three benchmark multimodal fake news datasets demonstrate the effectiveness and robustness of the proposed method, consistently outperforming state-of-the-art baselines in multimodal fake news detection tasks. The code is available at https://github.com/xxx.
Collagen, a key extracellular matrix protein, regulates cell adhesion, proliferation, and tissue integrity for skin health. In this study, a novel human-like type III collagen was engineered by substituting hydrophobic GXY triplets in the core fragment (Gly201-Asn498) of the human α1(III)type chain with hydrophilic motifs (GSP, GQP, GEP, and GSQ) and incorporating integrin-binding sequences (GER and RGD) at the C-terminus of the two consecutive sequence units. The codon-optimized gene was cloned into the pPIC9K vector, and transformed into Komagataella phaffii (Pichia pastoris) GS115. Multicopy transformants were selected by various G418 concentrations. Furthermore, a high yield of recombinant collagen at 19.4 g/L was achieved by fed-batch fermentation in a 5-L bioreactor. The purified protein exhibited excellent thermostability, remaining fully soluble after heating at 121 °C for 30 min. The collagen migrated on 10% SDS-PAGE with an apparent molecular weight approximately 2.18-fold greater than its theoretical mass, suggesting extensive hydration. Functionally, it significantly promoted the proliferation of human skin fibroblasts (HSF) and immortalized keratinocytes (HaCaT). In HSF cells, it upregulated the expression of COL1A1, COL3A1, and TIMP1, while in HaCaT cells, it enhanced transcription of skin barrier-related genes, including KRT1, KRT5, KRT10, KRT14, IVL, LOR, and FLG. This designed collagen integrates high-yield production, thermostability, and dual bioactivity for dermal regeneration and epidermal barrier reinforcement, showing a promising application for skincare and tissue engineering.
Symmetry, a fundamental concept in nature, science and art, has challenged computer vision researchers because it occurs in various forms and human symmetry perception can deviate from the mathematical definition. Previous symmetry detection datasets are limited by the number of annotators and by missing the nuances of human perception. We introduce PIX2PER, a novel dataset for reflection symmetry in natural scenes and artworks. We also introduce WF1, a modified version of the widely-used F1 detection performance score, by adding weights to precision and recall to accommodate for the perceived symmetry strength. Created by adding weights to precision and recall to accommodate for the perceived symmetry strength. We perform a comparative analysis of existing models for symmetry detection on this human-centric dataset. Additionally, we present a fully synthetic dataset for pretraining symmetry detection models. When finetuning this pretrained model with human data, performance increases significantly. This research introduces and evaluates ways of improving symmetry detection and contributes to the development of computer vision models that more effectively represent human perception.
暂无摘要（点击查看详情）
Human l-asparaginase 1 (hASNase1) is a promising next-generation candidate for Acute Lymphoblastic Leukemia (ALL) therapy due to its reduced immunogenicity and superior physiological compatibility. However, its recombinant production remains challenging, as expression in Escherichia coli typically results in insoluble aggregates and strong association with host chaperones, limiting biochemical characterization. Here, we established an optimized workflow combining mild induction conditions with tailored solubilization and purification strategies. Incubation with l-asparagine followed by ATP-mediated washing during affinity chromatography significantly improved protein recovery while reducing chaperone co-purification. This approach enabled the isolation of catalytically active hASNase1 displaying allosteric regulation and cooperative substrate binding, consistent with its proposed tetrameric organization. Complementary in silico analyses identified aggregation-prone regions, potentially providing structural insight into the observed expression challenges. Collectively, this study contributes to the development of improved recombinant production strategies for challenging human enzymes in prokaryotic systems.
This study examines the psychological mechanisms through which anthropomorphic artificial intelligence (AI) relates to consumer adoption intentions in fragile, low-trust economies. Integrating the Stimulus-Organism-Response framework with the Computers Are Social Actors paradigm, Institutional Trust Theory, and Privacy Calculus Theory, we investigate how human-like AI design shapes cognitive and affective responses within Sierra Leone's banking sector. Using survey data from 277 banking customers and partial least squares structural equation modeling, we find that AI anthropomorphism exhibits no direct association with adoption intention (β = -0.013, p = 0.760). Instead, its influence is entirely indirect-transmitted in parallel through perceived social presence (β = 0.144, 95% CI [0.062, 0.226]) and trust in the AI system (β = 0.139, 95% CI [0.068, 0.210]). Critically, customer skepticism-shaped by institutional fragility-functions as a boundary condition that substantially attenuates both pathways: among highly skeptical users (+1 SD), anthropomorphism's conditional effect on social presence becomes non-significant (β = 0.098, p = 0.124) compared to low-skepticism users (β = 0.412, p < 0.001), while its effect on trust is reduced by more than half (β = 0.118 vs. 0.284). These findings identify a critical boundary condition on human-like AI design: in low-trust environments, anthropomorphism operates not as a standalone adoption driver but as a relational amplifier whose efficacy depends on foundational trust and is substantially weakened when skepticism is high. The study challenges universalist assumptions in human-AI interaction research and underscores the need for institutionally sensitive design approaches in fragile economies.
Metabolic responses in mice at thermoneutrality (TN; 30°C) closely reflect those in humans rather than at conventional ambient temperature (CT; 22°C). However, the effect of TN on bile acid (BA) metabolism remains unclear. In this study, mice were maintained at CT or TN for 6 weeks, and BA profiles were quantitatively assessed at several sites, including enterohepatic and systemic circulation. TN markedly increased feed efficiency, adipose tissue mass, and hepatic and circulating triglyceride levels. Moreover, the ratio of 12-hydroxylated to non-12-hydroxylated BAs was elevated in the liver, small intestinal contents, and feces in mice maintained at TN. The ratios of primary to secondary BAs and of conjugated to unconjugated BAs were consistently reduced across sites, suggesting enhanced microbiota-dependent BA metabolism at TN. Collectively, these findings indicate that TN shifts murine BA composition toward a more human-like profile, highlighting the significance of TN models in terms of BA metabolism.
The rapid integration of generative artificial intelligence (GenAI) tools, such as ChatGPT, into educational contexts has raised important questions regarding how adolescents conceptualize and make sense of these technologies. Understanding students' perceptions is essential for developing age-appropriate, ethical, and pedagogically sound approaches to AI use in secondary education. This descriptive qualitative study employed a phenomenological approach and metaphor analysis to explore secondary school students' perceptions of generative artificial intelligence. The study sample consisted of 332 students aged 14-18 years from four secondary schools in Türkiye. Data were collected using an open-ended prompt ("Generative artificial intelligence is like … because …") and analyzed through content analysis. Metaphors were categorized based on shared semantic and conceptual features, and inter-rater reliability was established using Cohen's kappa (κ = 0.92). Analysis revealed ten metaphor categories clustered under five overarching themes: generative artificial intelligence as (1) a source of knowledge, (2) a teaching and guiding entity, (3) a supportive and assisting tool, (4) a reflection of human intelligence, and (5) a dual-purpose (beneficial-risky) technology. Students most frequently conceptualized GenAI as a comprehensive knowledge source (e.g., book, encyclopedia) and as a human-like cognitive entity (e.g., brain, wise person). At the same time, metaphors reflecting ethical awareness and potential risks, such as misuse and overreliance, were also identified. The findings indicate that secondary school students hold multifaceted and nuanced perceptions of generative artificial intelligence, encompassing both educational opportunities and ethical concerns. These results highlight the importance of integrating AI literacy into secondary education in ways that promote critical thinking, responsible use, and awareness of GenAI's limitations alongside its potential benefits. It was determined that secondary school students perceive generative artificial intelligence ambivalently as both a useful tool and a source of ethical and emotional concern, highlighting the need for developmentally appropriate artificial intelligence literacy approaches. • GenAI tools such as ChatGPT are increasingly integrated into educational contexts and have the potential to support personalized learning, information access, and student engagement. • Existing research has primarily focused on educators' perspectives or higher education settings, while studies examining adolescents' perceptions of GenAI remain limited. • This study provides empirical evidence on secondary school students' metaphorical perceptions of generative artificial intelligence within a K-12 context. • Findings reveal that adolescents conceptualize GenAI in multifaceted ways, including as a knowledge source, teaching and guiding entity, supportive tool, reflection of human intelligence, and a dual-purpose (beneficial-risky) technology.
Expressive piano performance poses extreme challenges for robotic manipulation, necessitating high-speed repetitive impacts, substantial force output, and coordinated multi-joint control under stringent dynamic constraints. However, existing robotic systems exhibit significant limitations in replicating human-level dexterity, as well as achievable motion speed and force output. This work presents a data-driven, bio-inspired dexterous robotic hand designed specifically for high-fidelity piano performance. We first extract kinematic primitives and stable inter-joint coupling patterns from large-scale motion capture data of professional pianists. These human motion priors are directly embedded into the mechanical architecture through morphological coupling and actuator allocation. Actuator selection is further guided by empirically measured human peak velocities and force profiles from biomechanics literature, ensuring sufficient bandwidth for high-speed repetitive motion and adequate force transmission. Experimental results demonstrate that the proposed hand replicates human-like joint coordination, achieves peak joint velocities of 53.88 rad/s, and provides sufficient fingertip force for authentic piano interaction. As a demonstration of its capabilities, the hand successfully performs a Grade 7 piano piece, Croatian Rhapsody, illustrating its potential for expressive musical performance. This research establishes a principled pathway from human motion statistics to embodied robotic intelligence, providing a high-performance hardware foundation for autonomous musical performance.
Ride comfort has become a crucial evaluation metric for autonomous vehicles. Existing studies on passenger comfort state estimation mainly rely on single-source vehicle data and traditional machine learning models for state estimation, which struggle to adequately capture local features in multi-source time-series signals and their nonlinear relationships with subjective perception, resulting in limited classification accuracy. To address this issue, this paper proposes a passenger ride comfort state (discomfort/no-discomfort) estimation method based on the fusion of human-vehicle data. Vehicle acceleration, passenger posture data, and individual characteristics are transformed into two-dimensional images using the recurrence plot (RP) technique to explicitly represent local temporal structures in the time-series signals, thereby improving data utilisation. Subsequently, a two-dimensional convolutional neural network is then used to train on the image data and identify comfort states. Experimental results verify that the performance of the proposed evaluation model outperforms traditional methods, with a state estimation accuracy of 94.04%. This study estimates passenger ride comfort by analysing vehicle vibrations via Recurrence Plots and CNNs. As autonomous driving systems increasingly approximate human-like driving styles, the method can provide potential technical support for improving ride comfort and alleviating motion sickness in autonomous vehicles.
Scanpath prediction in panoramic videos is a challenging task due to the spherical geometry and multimodality of the input, and the inherent uncertainty and diversity of the output. To give a complete treatment of these characteristics, we first present a simple criterion for scanpath prediction based on principles from lossy data compression. This criterion suggests minimizing the expected code length of quantized scanpaths, corresponding to fitting a discrete conditional probability model via maximum likelihood. We condition the probability model on two modalities: a viewport sequence as the deformation-reduced visual input and a set of relative past scanpaths projected onto respective viewports as the aligned path input. Furthermore, we parameterize it by a product of discretized Gaussian mixture models to capture the uncertainty and diversity of scanpaths from different humans. In doing so, the training of the probability model does not rely on the specification of "ground-truth" scanpaths for imitation learning. We also introduce a proportional-integral-derivative (PID) controller-based sampler to generate realistic human-like scanpaths from the learned probability model. Experimental results demonstrate that our method consistently produces better quantitative scanpath results in terms of prediction accuracy (by comparing to the assumed "ground-truths") and perceptual realism (through machine discrimination) over a wide range of prediction horizons. We additionally verify the perceptual realism improvement via a formal psychophysical experiment and the generalization improvement on several unseen panoramic video datasets.
Mammalian cell cultures are widely used for producing complex biopharmaceuticals that require human-like post-translational modifications, such as antibody-based therapeutics. Traditionally, serum-supplemented media support high cell viability and productivity; however, regulatory and scientific requirements demand serum-free conditions for clinical-grade manufacture. Recently, a novel fusion protein, the anti-huCD20(hγ1)-IL2no-alpha immunocytokine (IC), was presented as a promising therapeutic alternative, mostly for relapsed or refractory (r/r) B-cell non-Hodgkin lymphoma (B-NHL) patients, considering currently approved therapies. Three Chinese hamster ovary clones (K1 strain) producing the anti-huCD20(hγ1)-IL2no-alpha IC were generated and adapted to serum-free suspension culture. We performed a kinetic characterization of one clone in two culture media with different nutritional compositions, evaluating cell growth, productivity, cell cycle progression and mTOR signaling. The IC was purified by Protein A, then evaluated for identity, aggregation profile, CD20 recognition, CTLL-2 cytokine activity, ex vivo B-cell depletion in PBMC from r/r B-NHL patients and antitumor efficacy in immunocompetent C57BL/6 mice bearing EL4-hCD20- cells. The results demonstrated noticeable differences in cell growth and productivity in both batch and pseudo-perfusion performance, likely due to an influence on cell-cycle progression and mTOR signaling. The purified IC maintained its structural integrity while exhibiting an improved aggregation profile compared to serum-containing cultures. Furthermore, key biological activities, including B-cell depletion and antitumoral effects, remained intact. This research highlights the successful serum-free production of a functional anti-huCD20(hγ1)-IL2no-alpha IC, reinforcing its potential for biopharmaceutical development.
Generative Artificial Intelligence (GenAI) refers to the aspect of Artificial Intelligence (AI) that deals with generating content such as texts, images, music, videos, etc. GenAI has seen a surge in interest recently, particularly with the emergence of tools like ChatGPT, which demonstrate the potential of AI to generate contextually relevant and human-like outputs across various domains. GenAI has also enhanced existing language, image, and speech models, improving their performance on domain-specific tasks. This improvement has led to better integration in existing systems, including content creation, image generation, and enhanced decision-making processes. GenAI has facilitated personalized content delivery, provided data-driven insights, and automated complex tasks, enhancing efficiency and precision across various domains. While the advancements in GenAI have been greatly beneficial in multiple domains, there are many instances of misuse and abuse. This paper aims to provide a comprehensive review of the history of GenAI, its use cases, recent developments, downsides, and solutions to address the problem of its misuse in various sectors.
To behave adaptively in complex environments, animals must selectively process the most important information in space while ignoring distractors. Here, we report that an evolutionarily ancient group of brainstem inhibitory neurons, called PLTi, is surprisingly critical for this function of selective spatial attention. In freely behaving mice performing a human-like spatial attention task, we found that bilateral silencing of PLTi severely disrupted target selection without causing perceptual or task-relevant motor impairments. PLTi's effects depended necessarily on goal-relevant, rather than just physical salience-based signals, together revealing it as a specialized site for priority-driven attentional target selection. PLTi's core contribution was in controlling accuracy and categorical precision of the decision boundary separating the target from lower-priority distractors. PLTi's control of neural representations of competing stimuli in the superior colliculus, an established attentional hub, revealed a potential mechanistic pathway. PLTi may, therefore, be a conserved brainstem site across vertebrates for winner-take-all-like spatial decisions.
Deep convolutional neural networks (DCNNs) have approached and even surpassed human-level performance on many visual recognition tasks, yet they remain strikingly vulnerable to near-imperceptible perturbations generated by adversarial attacks. Recent research demonstrates that aligning DCNN representations with human visual cortex activity improves adversarial robustness, but the mechanisms driving this advantage are yet to be understood. One hypothesis suggests that neural alignment confers robustness by biasing models away from brittle high-frequency details and towards the low spatial frequencies (LSF). However, recent work indicates that human object recognition critically depends on a narrow, mid-frequency "human channel". Interestingly, this band was partially preserved in prior LSF-focused studies. In this work, we explicitly investigate whether a spectral bias towards the LSF or the human channel is the primary driver of the adversarial robustness observed in neurally aligned DCNNs. We first show that DCNNs aligned to higher-order regions of the human ventral visual stream systematically increase their reliance on both the LSF and the human channel. However, directly steering DCNNs towards these bands revealed a clear dissociation. Biasing models towards the human channel, either alone or together with the LSF, does not improve robustness and can even impair it. LSF bias produced some robustness gains, but such improvements are modest despite inducing much larger shifts in spatial-frequency reliance than the neurally-aligned DCNNs. Spatial-frequency-biased models overall show little, if any, increase in similarity to human neural representational geometry. Together, our results suggest that altered spatial-frequency reliance is likely an emergent property of learning more human-like representations rather than the primary mechanism by which neural alignment confers adversarial robustness, and motivate the need for future research examining representational properties beyond spatial-frequency biases.
Collective memory in humans refers to individual memories that become shared within a community and contribute to social identity, coordination, and cultural continuity. Extensive research shows that collective memory emerges through language-mediated interaction, social influence, and distributed cognitive mechanisms, supporting cooperation, decision-making, and cultural transmission. While these processes are well documented in humans, it remains unclear whether, and in what sense, collective memory exists in nonhuman animals. Here we propose a systematic comparative framework for understanding collective memory across species. We first outline the cognitive foundations of collective memory in humans, highlighting the roles of episodic, semantic, autobiographical, and procedural memory, together with mechanisms such as memory conformity, social contagion, memory convergence, transactive memory systems, and ritualized practices. We then examine whether nonhuman animals possess functional precursors of these mechanisms. Across taxa, animals exhibit episodic-like and semantic-like memory, social learning, conformity, distributed information systems, and ritualized behaviors that allow groups to store, transmit, and update knowledge beyond individual capacities. These processes support stable traditions, coordinated action, and adaptive responses. Although nonhuman animals show no evidence of autonoetic consciousness or human-like autobiographical memory, we argue that collective memory should be understood as a graded, non-dichotomous phenomenon grounded in distributed cognition.
Artificial intelligence (AI) is changing society in major ways. AI can create content that entertains, but it can also mislead people. One part of this change is the ability to produce human-like AI content, including voices and faces. In psychology, AI-generated faces have started to be used to study face perception, thanks to their flexibility and convenience. This choice appears to be supported by recent evidence indicating that AI-generated faces are indistinguishable from human faces. However, other findings suggest that AI- or computer-generated faces may be evaluated differently than real faces, with AI-generated faces even leading to hyperrealism. To address this, we conducted a study to investigate whether human and AI-generated faces are recognized similarly in a memory task, and whether their recognition is related to our evaluation of these faces. Results show that recognition accuracy was significantly higher for AI faces than for real faces. Critically, classification (i.e., the ability to correctly classify a face as human or AI-generated) accuracy was significantly better than chance level for both real and AI-generated faces, with no difference between the two. These results suggest that humans can correctly discriminate AI-generated and human faces, and that AI-faces might be recognized better than human ones, suggesting that AI-generated faces might engage human perceptual and memory systems differently from real faces. However, recognition and classification accuracy were not significantly correlated, suggesting that participants might be partially unaware of their increased performance with AI-generated faces.
Can large language models (LLMs) be viewed as a cognitive model of human language? Do they possess human-like language competence? To address these questions, this study takes a multifaceted approach, comparing the performance of 10 recent LLMs (n = 4000 responses) and 94 humans (n = 3760 responses) on grammaticality judgments and sentence interpretations, focusing on five linguistic phenomena that involve missing material. The analyses show that while the LLMs appeared to differentiate between grammatical/possible and ungrammatical/impossible sentences/interpretations overall, they struggled with infrequent phenomena (e.g., Gapping, Sluicing), often rejecting grammatical sentences and accepting impossible interpretations. Notably, increased size seemed to improve their performance on grammaticality judgments, but neither size nor reasoning capability improved their performance on interpretation. In contrast, humans demonstrated a clear sensitivity to these distinctions. The findings seem to align with the view that LLMs, in their current form, lack language competence and do not provide a convincing explanation of human language.