Intrinsic Bias Research Articles

Artificial Intelligence Tools and Bias in Journalism-related Content Generation: Comparison Between Chat GPT-3.5, GPT-4 and Bing

This study explores the biases present in artificial intelligence (AI) tools, focusing on GPT-3.5, GPT-4, and Bing. The performance of the tools has been compared with a group of experts in linguistics, and journalists specialized in breaking news and international affairs. It reveals that GPT-3.5, widely accessible and free, exhibits a higher tendency rate in its word generation, suggesting an intrinsic bias within the tool itself rather than in the input data. Comparatively, GPT-4 and Bing demonstrate differing patterns in term generation and subjectivity, with GPT-4 aligning more closely with expert opinions and producing fewer opinative words. The research highlights the extensive use of generative AI in media and among the general populace, emphasizing the need for careful reliance on AI-generated content. The findings stress the risks of misinformation and biased reporting inherent in unexamined AI outputs. The challenge for journalists and information professionals is to ensure accuracy and ethical judgment in content creation to maintain the quality and diversity of content in journalistic practices.

Androcentrism in Arabic Educational Materials: A Linguistic Analysis of GCC Countries’ Selected School Textbooks

This study seeks to investigate the existence of androcentrism in school textbooks used in the six Gulf Cooperation Council (GCC) countries. It attempts to unveil potential implicit biases ingrained in the language of these textbooks. The data for this study is derived from a corpus comprising six textbooks, encompassing a variety of courses representing the GCC secondary level education systems. The employed methodological approach entailed content analysis, wherein categories were discerned directly from the textual data during a comprehensive examination of the textbooks. The established framework facilitated the classification of gendered linguistic references into three distinct categories: nouns, verbs, and pronouns. The principal findings highlight preferential representation of male students and, consequently, underrepresentation of female students, permeating almost the entire research corpus. Contrary to assumptions suggesting an intrinsic gender bias within the Arabic language, this investigation argues that the locus of androcentrism seems to be rooted in the authors of the scrutinized textbooks. This discernment contributes valuable insights challenging prevailing notions and accentuates the role of textbook authors in shaping linguistic biases.

The microbial composition of pancreatic ductal adenocarcinoma: A systematic review of 16S rRNA gene sequencing.

Pancreatic cancer, specifically pancreatic ductal adenocarcinoma (PDAC), continues to pose a significant clinical and scientific challenge. The most significant finding of recent years is that PDAC tumours harbour their specific microbiome, which differs amongst tumour entities and is distinct from healthy tissue. This review aims to evaluate and summarise all PDAC studies that have used the next-generation technique, 16S rRNA gene amplicon sequencing within each bodily compartment. As well as establishing a causal relationship between PDAC and the microbiome. This systematic review was carried out according to the Preferred Reporting Items for Systematic Reviews and Meta-analysis (PRISMA) guidelines. A comprehensive search strategy was designed, and 1727 studies were analysed. In total, 38 studies were selected for qualitative analysis and summarised significant PDAC bacterial signatures. Despite the growing amount of data provided, we are not able to state a universal 16S rRNA gene microbial signature that can be used for PDAC screening. This is most certainly due to the heterogeneity of the presentation of results, lack of available datasets and the intrinsic selection bias between studies. Several key studies have begun to shed light on causality and the influence the microbiome constituents and their produced metabolites could play in tumorigenesis and influencing outcomes. The challenge in this field is to shape the available microbial data into targetable signatures. Making sequenced data readily available is critical, coupled with the coordinated standardisation of data and the need for consensus guidelines in studies investigating the microbiome in PDAC.

A massively parallel in vivo assay of TdT mutants yields variants with altered nucleotide insertion biases.

Terminal deoxynucleotidyl transferase (TdT) is a unique DNA polymerase capable of template-independent extension of DNA with random nucleotides. TdT's de novo DNA synthesis ability has found utility in DNA recording, DNA data storage, oligonucleotide synthesis, and nucleic acid labeling, but TdT's intrinsic nucleotide biases limit its versatility in such applications. Here, we describe a multiplexed assay for profiling and engineering the bias and overall activity of TdT variants in high throughput. In our assay, a library of TdTs is encoded next to a CRISPR-Cas9 target site in HEK293T cells. Upon transfection of Cas9 and sgRNA, the target site is cut, allowing TdT to intercept the double strand break and add nucleotides. Each resulting insertion is sequenced alongside the identity of the TdT variant that generated it. Using this assay, 25,623 unique TdT variants, constructed by site-saturation mutagenesis at strategic positions, were profiled. This resulted in the isolation of several altered-bias TdTs that expanded the capabilities of our TdT-based DNA recording system, Cell History Recording by Ordered Insertion (CHYRON), by increasing the information density of recording through an unbiased TdT and achieving dual-channel recording of two distinct inducers (hypoxia and Wnt) through two differently biased TdTs. Select TdT variants were also tested in vitro , revealing concordance between each variant's in vitro bias and the in vivo bias determined from the multiplexed high throughput assay. Overall, our work, and the multiplex assay it features, should support the continued development of TdT-based DNA recorders, in vitro applications of TdT, and further study of the biology of TdT.

Intrinsic superconducting phase battery

Leveraging the quantization properties inherent in superconductors, we present the development of an intrinsic superconducting phase battery. This advancement is achieved by integrating a ferromagnetic π-phase Josephson junction (π-JJ) within a superconducting ring. The core innovation lies in harnessing the potential energy of the π-JJ to generate an intrinsic circulating current, thereby inducing a phase difference as the current goes through the geometric inductance. This mechanism allows for tuning the phase bias φ (0 < φ < π) through an arrangement of the geometric inductance in the battery. We integrate the intrinsic phase batteries into superconducting quantum interference devices, where we verified the effectiveness of the induced phase bias. The polarity of the phase battery is determined by the direction of the intrinsic circulating current, which can be initialized by an external magnetic field. The design methodology for precise intrinsic phase bias has been established. Our findings not only show the feasibility of generating an intrinsic and adjustable phase bias using established fabrication techniques but also open new avenues for enhancing the design, efficiency, and functionality of superconducting electronics, promising to accelerate advancements in digital and quantum computing technologies.

How does prestige bias affect information recall during a pandemic?

The prestige theory of evolution states that our memory has an intrinsic bias to memorize information from someone of prestige. However, the evidence for information recall is mainly focused on content bias. Considering that the prestige bias can be advantageous in selecting information in contexts of uncertainty, this study assessed whether, in the scenario of the COVID-19 pandemic, the prestige bias would be favored over other models that do not possess the prestige spirit characteristics. The study was conducted through an online experiment, where participants were subjected to reading fictitious text, followed by a surprise recollection. Data were analyzed using a generalized linear mixed model, Poisson family, and logistic regression. The results showed that prestige is only prioritized in the recall due to the family model and does not present any difference from the other models tested. However, it influenced the recall of specific information, suggesting its role as a factor of cultural attraction. Furthermore, we observed that trust in science-oriented profiles can influence the recall of information during a health crisis. Finally, this study highlights the complexity of the functioning of the human mind and how several factors can act simultaneously in the recall of information.

SCCAM: Supervised Contrastive Convolutional Attention Mechanism for Ante-Hoc Interpretable Fault Diagnosis With Limited Fault Samples.

In real industrial processes, fault diagnosis methods are required to learn from limited fault samples since the procedures are mainly under normal conditions and the faults rarely occur. Although attention mechanisms have become increasingly popular for the task of fault diagnosis, the existing attention-based methods are still unsatisfying for the above practical applications. First, pure attention-based architectures like transformers need a substantial quantity of fault samples to offset the lack of inductive biases thus performing poorly under limited fault samples. Moreover, the poor fault classification dilemma further leads to the failure of the existing attention-based methods to identify the root causes. To develop a solution to the aforementioned problems, we innovatively propose a supervised contrastive convolutional attention mechanism (SCCAM) with ante-hoc interpretability, which solves the root cause analysis problem under limited fault samples for the first time. First, accurate classification results are obtained under limited fault samples. More specifically, we integrate the convolutional neural network (CNN) with attention mechanisms to provide strong intrinsic inductive biases of locality and spatial invariance, thereby strengthening the representational power under limited fault samples. In addition, we ulteriorly enhance the classification capability of the SCCAM method under limited fault samples by employing the supervised contrastive learning (SCL) loss. Second, a novel ante-hoc interpretable attention-based architecture is designed to directly obtain the root causes without expert knowledge. The convolutional block attention module (CBAM) is utilized to directly provide feature contributions behind each prediction thus achieving feature-level explanations. The proposed SCCAM method is testified on a continuous stirred tank heater (CSTH) and the Tennessee Eastman (TE) industrial process benchmark. Three common fault diagnosis scenarios are covered, including a balanced scenario for additional verification and two scenarios with limited fault samples (i.e., imbalanced scenario and long-tail scenario). The effectiveness of the presented SCCAM method is evidenced by the comprehensive results that show our method outperforms the state-of-the-art methods in terms of fault classification and root cause analysis.

Pruning Self-Attentions Into Convolutional Layers in Single Path.

Vision Transformers (ViTs) have achieved impressive performance over various computer vision tasks. However, modeling global correlations with multi-head self-attention (MSA) layers leads to two widely recognized issues: the massive computational resource consumption and the lack of intrinsic inductive bias for modeling local visual patterns. To solve both issues, we devise a simple yet effective method named Single-Path Vision Transformer pruning (SPViT), to efficiently and automatically compress the pre-trained ViTs into compact models with proper locality added. Specifically, we first propose a novel weight-sharing scheme between MSA and convolutional operations, delivering a single-path space to encode all candidate operations. In this way, we cast the operation search problem as finding which subset of parameters to use in each MSA layer, which significantly reduces the computational cost and optimization difficulty, and the convolution kernels can be well initialized using pre-trained MSA parameters. Relying on the single-path space, we introduce learnable binary gates to encode the operation choices in MSA layers. Similarly, we further employ learnable gates to encode the fine-grained MLP expansion ratios of FFN layers. In this way, our SPViT optimizes the learnable gates to automatically explore from a vast and unified search space and flexibly adjust the MSA-FFN pruning proportions for each individual dense model. We conduct extensive experiments on two representative ViTs showing that our SPViT achieves a new SOTA for pruning on ImageNet-1 k. For example, our SPViT can trim 52.0% FLOPs for DeiT-B and get an impressive 0.6% top-1 accuracy gain simultaneously.

Breast density quantification in dual-energy mammography using virtual anthropomorphic phantoms.

Breast density is a significant risk factor for breast cancer and can impact the sensitivity of screening mammography. Area-based breast density measurements may not provide an accurate representation of the tissue distribution, therefore volumetric breast density (VBD) measurements are preferred. Dual-energy mammography enables volumetric measurements without additional assumptions about breast shape. In this work we evaluated the performance of a dual-energy decomposition technique for determining VBD by applying it to virtual anthropomorphic phantoms. The dual-energy decomposition formalism was used to quantify VBD on simulated dual-energy images of anthropomorphic virtual phantoms with known tissue distributions. We simulated 150 phantoms with volumes ranging from 50 to 709mL and VBD ranging from 15% to 60%. Using these results, we validated a correction for the presence of skin and assessed the method's intrinsic bias and variability. As a proof of concept, the method was applied to 14 sets of clinical dual-energy images, and the resulting breast densities were compared to magnetic resonance imaging (MRI) measurements. Virtual phantom VBD measurements exhibited a strong correlation (Pearson's ) with nominal values. The proposed skin correction eliminated the variability due to breast size and reduced the bias in VBD to a constant value of -2%. Disagreement between clinical VBD measurements using MRI and dual-energy mammography was under 10%, and the difference in the distributions was statistically non-significant. VBD measurements in both modalities had a moderate correlation (Spearman's =0.68). Our results in virtual phantoms indicate that the material decomposition method can produce accurate VBD measurements if the presence of a third material (skin) is considered. The results from our proof of concept showed agreement between MRI and dual-energy mammography VBD. Assessment of VBD using dual-energy images could provide complementary information in dual-energy mammography and tomosynthesis examinations.

Intrinsic Neural Excitability Biases Allocation and Overlap of Memory Engrams.

Memories are thought to be stored in neural ensembles known as engrams that are specifically reactivated during memory recall. Recent studies have found that memory engrams of two events that happened close in time tend to overlap in the hippocampus and the amygdala, and these overlaps have been shown to support memory linking. It has been hypothesized that engram overlaps arise from the mechanisms that regulate memory allocation itself, involving neural excitability, but the exact process remains unclear. Indeed, most theoretical studies focus on synaptic plasticity and little is known about the role of intrinsic plasticity, which could be mediated by neural excitability and serve as a complementary mechanism for forming memory engrams. Here, we developed a rate-based recurrent neural network that includes both synaptic plasticity and neural excitability. We obtained structural and functional overlap of memory engrams for contexts that are presented close in time, consistent with experimental and computational studies. We then investigated the role of excitability in memory allocation at the network level and unveiled competitive mechanisms driven by inhibition. This work suggests mechanisms underlying the role of intrinsic excitability in memory allocation and linking, and yields predictions regarding the formation and the overlap of memory engrams.

The atypical IκB family member Bcl3 determines differentiation and fate of intestinal RORγt+ regulatory T-cell subsets

Peripherally induced regulatory T cells (pTregs) expressing the retinoic acid receptor-related orphan-receptor gamma t (RORγt) are indispensable for intestinal immune homeostasis. NF-κB family members regulate the differentiation of thymic Tregs and promote their survival in the periphery. However, the Treg-intrinsic molecular mechanisms controling the size of the pTregs in the intestine and associated lymphoid organs remain unclear. Here, we provide direct evidence that Bcl3 limits the development of pTregs in a T cell-intrinsic manner. Moreover, the absence of Bcl3 allowed for the formation of an unusual intestinal Treg population co-expressing the transcription factors Helios and RORγt. The expanded RORγt+ Treg populations in the absence of Bcl3 displayed an activated phenotype and secreted high levels of the anti-inflammatory cytokines IL-10 and TGFβ. They were fully capable of suppressing effector T cells in a transfer colitis model despite an intrinsic bias to trans-differentiate toward Th17-like cells. Finally, we provide a Bcl3-dependent gene signature in pTregs including altered responsiveness to the cytokines IL-2, IL-6, and TNFα. Our results demonstrate that Bcl3 acts as a molecular switch to limit the expansion of different intestinal Treg subsets and may thus serve as a novel therapeutic target for inflammatory bowel disease by restoring intestinal immune tolerance.

JAML promotes the antitumor role of tumor-resident CD8+ T cells by facilitating their innate-like function in human lung cancer

Tissue-resident memory CD8+T cells (CD8+TRMs) are thought to play a crucial role in cancer immunosurveillance. However, the characteristics of CD8+TRMs in the tumor microenvironment (TME) of human non-small cell lung cancer (NSCLC) remain unclear. Here, we report that CD8+TRMs accumulate explicitly and exhibit a unique gene expression profile in the TME of NSCLC. Interestingly, these tumor-associated CD8+TRMs uniquely exhibit an innate-like phenotype. Importantly, we found that junction adhesion molecule-like (JAML) provides an alternative costimulatory signal to activate tumor-associated CD8+TRMs via combination with cancer cell-derived CXADR (CXADR Ig-like cell adhesion molecule). Furthermore, we demonstrated that activating JAML could promote the expression of TLR1/2 on CD8+TRMs, inhibit tumor progression and prolong the survival of tumor-bearing mice. Finally, we found that higher CD8+TRMs and JAML expression in the TME could predict favorable clinical outcomes in NSCLC patients. Our study reveals an intrinsic bias of CD8+TRMs for receiving the tumor-derived costimulatory signal in the TME, which sustains their innate-like function and antitumor role. These findings will shed more light on the biology of CD8+TRMs and aid in the development of potential targeted treatment strategies for NSCLC.

Slimmable transformer with hybrid axial-attention for medical image segmentation

The transformer architecture has achieved remarkable success in medical image analysis owing to its powerful capability for capturing long-range dependencies. However, due to the lack of intrinsic inductive bias in modeling visual structural information, the transformer generally requires a large-scale pre-training schedule, limiting the clinical applications over expensive small-scale medical data. To this end, we propose a slimmable transformer to explore intrinsic inductive bias via position information for medical image segmentation. Specifically, we empirically investigate how different position encoding strategies affect the prediction quality of the region of interest (ROI) and observe that ROIs are sensitive to different position encoding strategies. Motivated by this, we present a novel Hybrid Axial-Attention (HAA) that can be equipped with pixel-level spatial structure and relative position information as inductive bias. Moreover, we introduce a gating mechanism to achieve efficient feature selection and further improve the representation quality over small-scale datasets. Experiments on LGG and COVID-19 datasets prove the superiority of our method over the baseline and previous works. Internal workflow visualization with interpretability is conducted to validate our success better; the proposed slimmable transformer has the potential to be further developed into a visual software tool for improving computer-aided lesion diagnosis and treatment planning.

Panoptic Scene Graph Generation with Semantics-Prototype Learning

Panoptic Scene Graph Generation (PSG) parses objects and predicts their relationships (predicate) to connect human language and visual scenes. However, different language preferences of annotators and semantic overlaps between predicates lead to biased predicate annotations in the dataset, i.e. different predicates for the same object pairs. Biased predicate annotations make PSG models struggle in constructing a clear decision plane among predicates, which greatly hinders the real application of PSG models. To address the intrinsic bias above, we propose a novel framework named ADTrans to adaptively transfer biased predicate annotations to informative and unified ones. To promise consistency and accuracy during the transfer process, we propose to observe the invariance degree of representations in each predicate class, and learn unbiased prototypes of predicates with different intensities. Meanwhile, we continuously measure the distribution changes between each presentation and its prototype, and constantly screen potentially biased data. Finally, with the unbiased predicate-prototype representation embedding space, biased annotations are easily identified. Experiments show that ADTrans significantly improves the performance of benchmark models, achieving a new state-of-the-art performance, and shows great generalization and effectiveness on multiple datasets. Our code is released at https://github.com/lili0415/PSG-biased-annotation.

Abstract 7435: PATTY: A bias estimation and correction model for bulk and single-cell CUT&Tag data

Abstract The accurate detection of transcription factor (TF) binding sites and histone modifications (HM) on the genome-wide scale is essential for studying functional epigenetics and gene regulation. Cleavage Under Targets & Tagmentation (CUT&Tag) is a low-cost and easy-to-implement epigenomic profiling method that can be performed on a low number of cells and on the single-cell level. CUT&Tag experiments use the hyperactive transposase Tn5 for tagmentation. We find that Tn5 is subject to intrinsic sequence insertion bias (intrinsic bias). Additionally, the preference of Tn5 insertion toward accessible chromatin also affects the distribution of CUT&Tag reads (open chromatin bias). Both types of biases can significantly confound bulk and single-cell CUT&Tag data analysis, which requires careful assessment and new analytical methods. To address this challenge, we present PATTY (Propensity Analyzer for Tn5 Transposase Yielded bias), a computational method for systematic characterization and correction of biases in CUT&Tag data. We demonstrate that histone modification signals (H3K27me3 and H3K27ac) detected from CUT&Tag data after bias correction using PATTY are better associated with orthogonal biological features such as gene expression. PATTY-corrected single-cell CUT&Tag signals for histone modification can better cluster cells into their true cell types. This new computational method can improve both bulk and single-cell CUT&Tag data analysis. Citation Format: Sheng'en Shawn Hu, Qingying Chen, Megan Grieco, Chongzhi Zang. PATTY: A bias estimation and correction model for bulk and single-cell CUT&Tag data [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 7435.

Identification of group differences in predictive anticipatory biasing of pain during uncertainty: preparing for the worst but hoping for the best.

Pain anticipation during conditions of uncertainty can unveil intrinsic biases, and understanding these biases can guide pain treatment interventions. This study used machine learning and functional magnetic resonance imaging to predict anticipatory responses in a pain anticipation experiment. One hundred forty-seven participants that included healthy controls (n = 57) and individuals with current and/or past mental health diagnosis (n = 90) received cues indicating upcoming pain stimuli: 2 cues predicted high and low temperatures, while a third cue introduced uncertainty. Accurate differentiation of neural patterns associated with specific anticipatory conditions was observed, involving activation in the anterior short gyrus of the insula and the nucleus accumbens. Three distinct response profiles emerged: subjects with a negative bias towards high pain anticipation, those with a positive bias towards low pain anticipation, and individuals whose predictions during uncertainty were unbiased. These profiles remained stable over one year, were consistent across diagnosed psychopathologies, and correlated with cognitive coping styles and underlying insula anatomy. The findings suggest that individualized and stable pain anticipation occurs in uncertain conditions.

Disparities in seizure outcomes revealed by large language models.

Large-language models (LLMs) can potentially revolutionize health care delivery and research, but risk propagating existing biases or introducing new ones. In epilepsy, social determinants of health are associated with disparities in care access, but their impact on seizure outcomes among those with access remains unclear. Here we (1) evaluated our validated, epilepsy-specific LLM for intrinsic bias, and (2) used LLM-extracted seizure outcomes to determine if different demographic groups have different seizure outcomes. We tested our LLM for differences and equivalences in prediction accuracy and confidence across demographic groups defined by race, ethnicity, sex, income, and health insurance, using manually annotated notes. Next, we used LLM-classified seizure freedom at each office visit to test for demographic outcome disparities, using univariable and multivariable analyses. We analyzed 84675 clinic visits from 25612 unique patients seen at our epilepsy center. We found little evidence of bias in the prediction accuracy or confidence of outcome classifications across demographic groups. Multivariable analysis indicated worse seizure outcomes for female patients (OR 1.33, P ≤ .001), those with public insurance (OR 1.53, P ≤ .001), and those from lower-income zip codes (OR ≥1.22, P ≤ .007). Black patients had worse outcomes than White patients in univariable but not multivariable analysis (OR 1.03, P = .66). We found little evidence that our LLM was intrinsically biased against any demographic group. Seizure freedom extracted by LLM revealed disparities in seizure outcomes across several demographic groups. These findings quantify the critical need to reduce disparities in the care of people with epilepsy.

Making the leap. Part I. Modelling the reconstructed lensing convergence PDF from cosmic shear with survey masks and systematics

The last few years have seen the development of a promising theoretical framework for statistics of the cosmic large-scale structure — the theory of large deviations (LDT) for modelling weak-lensing one-point statistics in the mildly nonlinear regime. The goal of this series of papers is to make the leap and lay out the steps to perform an actual data analysis with this theoretical tool. Building upon the LDT framework, in this work (paper I) we demonstrate how to accurately model the Probability Distribution Function (PDF) of a reconstructed Kaiser-Squires convergence field under a realistic mask, that of the third data release of the Dark Energy Survey (DES). We also present how weak lensing systematics and higher-order lensing corrections due to intrinsic alignments, shear biases, photo-z errors and baryonic feedback can be incorporated in the modelling of the reconstructed convergence PDF. In an upcoming work (paper II) we will then demonstrate the robustness of our modelling through simulated likelihood analyses, the final step required before applying our method to actual data.

Constraining ultralight axions with CSST weak gravitational lensing and galaxy clustering photometric surveys

ABSTRACT Ultralight axion (ULA) can be one of the potential candidates for dark matter. The extremely low mass of the ULA can lead to a de Broglie wavelength the size of galaxies which results in a suppression of the growth of structure on small scales. In this work, we forecast the constraint on the ULA particle mass ma and relative fraction to dark matter fa = Ωa/Ωd for the forthcoming Stage IV space-based optical survey equipment CSST (China Space Station Telescope). We focus on the CSST cosmic shear and galaxy clustering photometric surveys, and forecast the measurements of shear, galaxy, and galaxy–galaxy lensing power spectra (i.e. 3 × 2 pt). The effects of neutrino, baryonic feedback, and uncertainties of intrinsic alignment, shear calibration, galaxy bias, and photometric redshift are also included in the analysis. After performing a joint constraint on all the cosmological and systematical parameters based on the simulated data from the theoretical prediction, we obtain a lower limit of the ULA particle mass log10 (ma/eV) ≥ −22.5 and an upper limit of the ULA fraction fa ≤ 0.83 at 95 per cent confidence level, and log10 (ma/eV) ≥ −21.9 with fa ≤ 0.77 when ignoring the baryonic feedback. We find that the CSST photometric surveys can improve the constraint on the ULA mass by about one order of magnitude, compared to the current constraints using the same kind of observational data.

Leveraging saliency priors and explanations for enhanced consistent interpretability

Deep neural networks have emerged as highly effective tools for computer vision systems, showcasing remarkable performance. However, the intrinsic opacity, potential biases, and vulnerability to shortcut learning in these models pose significant concerns regarding their practical application. To tackle these issues, this work employs saliency prior and explanations to enhance the credibility, reliability, and interpretability of neural networks. Specifically, we employ a salient object detection algorithm to extract human-consistent priors from images for data augmentation. The identified saliency priors, along with explanations, serve as supervision signals directing the network’s focus to salient regions within the image. Additionally, contrastive self-supervised learning is incorporated to enable the model to discern the most discriminative concepts. Experimental results confirm the algorithm’s capability to align model explanations with human priors, thereby improving interpretability. Moreover, the proposed approach enhances model performance in data-limited and fine-grained classification scenarios. Importantly, our algorithm is label-independent, allowing for the integration of unlabeled data during training. In practice, this method contributes to improving the reliability and interpretability of intelligent models for downstream tasks. Our code is available here: https://github.com/DLAIResearch/SGC.