Representation Learning Research Articles

Abstract 4998: Deriving patient similarity networks in myeloid neoplasms using multi-modal representation learning

Abstract Myeloid neoplasms (MNs), such as Myelodysplastic Syndromes (MDS) and Acute Myeloid Leukemia (AML), are clinically overlapping hematologic malignancies. Their assessment includes the evaluation of bone marrow (BM) morphology, immunophenotyping by flow cytometry, cytogenetics, and gene mutation profiling. Established diagnostic criteria between the MN subtypes are primarily guided by empirical unimodal thresholds (i.e. 20% blasts differentiating MDS from AML) that may not always reflect the underlying disease biology. This study integrates different data modalities and produces patient-specific multi-modal views aiming to uncover associations between molecular subtypes and clinical phenotypes at the cellular and morphological level, show distinct biological and clinically relevant subgroups in MNs, and help inform future guidelines for disease classification. The study cohort consists of 523 MSKCC patients across the spectrum of MNs. The dataset includes mutation and cytogenetic profiles, raw hematoxylin and eosin (H&E) BM whole slide images (WSIs), raw single-cell flow cytometry measurements, and differential cell type counts. Data for all modalities are present for 269 patients. The flow cytometry and imaging data views were the products of in-house computational workflows. In flow cytometry, single cells were grouped in clusters by the Leiden algorithm. Then, the cell abundance distributions across these clusters were used to form patient-level representations. Our imaging framework first identified the cell-rich regions of the WSIs by training a tile-level vision transformer (ROCAUC: 0.95). After obtaining the encodings of the cell-rich regions from the QuiltNet foundation model (Ikezogwo et al. NeurIPS 2023), we trained a graph neural network (SlideGraph+, Lu et al. Medical Image Analysis 2022) to predict diagnostic labels (ROCAUC: 0.79), and we subsequently retrieved the WSI deep embeddings. Different combinations of the unimodal views (genomics, flow cytometry, WSIs, differential cell types) were integrated into multi-modal embeddings through the Multiple Similarity Network Embedding (Xu et al. Methods 2021) method. Low dimensional projections and transformation of these embeddings to patient similarity graphs show the formation of homogeneous subgroups driven by somatic mutation status (i.e. TP53 mutated MNs), clusters defined by contributions of features across modalities, and the presence of out-of-group patients indicative of disease heterogeneity. In summary, this work produces multi-modal patient embeddings in MNs and constructs patient networks indicative of both inter-patient similarity and heterogeneity. Characterizing the data-derived representations of disease profiles and the underlying inter-modality relationships will be critical for the future definition of diagnostic criteria that incorporate multi-modal measurements. Citation Format: Georgios Asimomitis, Jake Sauter, Darin Moore, Himanshu Bhurtel, Katherine Lopez, Armaan Kohli, Kevin Boehm, Christopher Fong, Arfath Pasha, Luke Geneslaw, Anika Begum, Tom Pollard, Jacob Glass, Gregory M. Goldgof, Nicholas Barasch, Ahmet Dogan, Mikhail Roshal, Elli Papaemmanuil. Deriving patient similarity networks in myeloid neoplasms using multi-modal representation learning [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 4998.

Abstract 5084: Evaluation of single-cell foundation models for cancer outcome predictions

Introduction: Tumor single-cell data provide unique insights into the cellular dynamics of cancer progression and therapeutic outcomes. Inspired by advancements in natural language processing (NLP), foundation models are increasingly used in single-cell analyses to uncover biological insights and identify clinical biomarkers. However, the performance of these models on cancer-specific tasks remains underexplored. We present an extensive evaluation of foundation models trained on millions of single-cell profiles for cancer-focused investigations. Methods: We evaluated nine predictive models-three traditional methods and six foundation models (four Geneformer and two scGPT models), across various downstream tasks. Traditional models used embeddings from high variable genes (HVG), principal components (PCA), and generative model embedding (scVI), each paired with a random forest classifier for downstream classification. The evaluated foundation models varied in architecture sizes, pretraining dataset sizes (30M to 95M cell profiles), and training datasets, which were either generic or cancer-specific. We assessed the quality of representations by measuring their predictive performance on lung cancer treatment status (TKI treated vs. treatment-naive patients) and breast cancer subtypes (ER+ vs. TNBC). We examined how model characteristics influenced their performance in these domains for all tasks. Results: Zero-shot evaluations of the foundation models showed superior single-cell data clustering, measured by average silhouette width (ASW), compared to traditional methods (HVG, PCA, scVI), suggesting that pre-trained embeddings from millions of single-cell profiles may offer enhanced representations. We thus evaluated the learned representation on downstream predictive tasks, showing that foundation models with a larger pretraining dataset (95M cells), cancer-specific training, larger input size (4096 genes), and deeper model (12 layers) achieved the best performance among foundation models (F1=0.97) for predicting treatment status in lung cancer. In contrast, smaller and generic foundation models had significantly lower performance, however, additional self-supervised training on task-specific data rescued their performance to match larger foundation models. Similar trends were observed in predicting breast cancer subtypes. Notably, the best foundation model’s performance was not significantly different from the best traditional method (PCA) in predicting breast cancer subtypes. Conclusions: These results highlight key factors driving foundation model performance on cancer-specific tasks. While current foundation models show inconsistent contributions to cancer-specific questions, observed performance trends highlight their potential to enhance single-cell data representation and drive significant advancements in cancer research and precision medicine. Citation Format: Haitham Elmarakeby, Ahmed Roman, Shreya Johri, Eliezer M. Van Allen. Evaluation of single-cell foundation models for cancer outcome predictions [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 5084.

MambaOSR: Leveraging Spatial-Frequency Mamba for Distortion-Guided Omnidirectional Image Super-Resolution

Omnidirectional image super-resolution (ODISR) is critical for VR/AR applications, as high-quality 360° visual content significantly enhances immersive experiences. However, existing ODISR methods suffer from limited receptive fields and high computational complexity, which restricts their ability to model long-range dependencies and extract global structural features. Consequently, these limitations hinder the effective reconstruction of high-frequency details. To address these issues, we propose a novel Mamba-based ODISR network, termed MambaOSR, which consists of three key modules working collaboratively for accurate reconstruction. Specifically, we first introduce a spatial-frequency visual state space model (SF-VSSM) to capture global contextual information via dual-domain representation learning, thereby enhancing the preservation of high-frequency details. Subsequently, we design a distortion-guided module (DGM) that leverages distortion map priors to adaptively model geometric distortions, effectively suppressing artifacts resulting from equirectangular projections. Finally, we develop a multi-scale feature fusion module (MFFM) that integrates complementary features across multiple scales, further improving reconstruction quality. Extensive experiments conducted on the SUN360 dataset demonstrate that our proposed MambaOSR achieves a 0.16 dB improvement in WS-PSNR and increases the mutual information by 1.99% compared with state-of-the-art methods, significantly enhancing both visual quality and the information richness of omnidirectional images.

Transformer-based approach for printing quality recognition in fused filament fabrication

Ensuring high-quality prints in additive manufacturing is a critical challenge due to the variability in materials, process parameters, and equipment. Machine learning models are increasingly being employed for real-time quality monitoring, enabling the detection and classification of defects such as under-extrusion and over-extrusion. Vision Transformers (ViTs), with their global self-attention mechanisms, offer a promising alternative to traditional convolutional neural networks (CNNs). This paper presents a transformer-based approach for print quality recognition in additive manufacturing technologies, with a focus on fused filament fabrication (FFF), leveraging advanced self-supervised representation learning techniques to enhance the robustness and generalizability of ViTs. We show that the ViT model effectively classifies printing quality into different levels of extrusion, achieving exceptional performance across varying dataset scales and noise levels. Training evaluations show a steady decrease in cross-entropy loss, with prediction accuracy, precision, recall, and the harmonic mean of precision and recall (F1) scores reaching close to 1 within 40 epochs, demonstrating excellent performance across all classes. The macro and micro F1 scores further emphasize the ability of ViT to handle both class imbalance and instance-level accuracy effectively. Our results also demonstrate that ViT outperforms CNN in all scenarios, particularly in noisy conditions and with small datasets. Comparative analysis reveals ViT advantages, particularly in leveraging global self-attention and robust feature extraction methods, enhancing its ability to generalize effectively and remain resilient with limited data. These findings underline the potential of the transformer-based approach as a scalable, interpretable, and reliable solution to real-time quality monitoring in FFF, addressing key challenges in additive manufacturing defect detection and ensuring process efficiency.

Hierarchical embedding attention for overall survival prediction in lung cancer from unstructured EHRs

The automated processing of Electronic Health Records (EHRs) poses a significant challenge due to their unstructured nature, rich in valuable, yet disorganized information. Natural Language Processing (NLP), particularly Named Entity Recognition (NER), has been instrumental in extracting structured information from EHR data. However, existing literature primarly focuses on extracting handcrafted clinical features through NLP and NER methods without delving into their learned representations. In this work, we explore the untapped potential of these representations by considering their contextual richness and entity-specific information. Our proposed methodology extracts representations generated by a transformer-based NER model on EHRs data, combines them using a hierarchical attention mechanism, and employs the obtained enriched representation as input for a clinical prediction model. Specifically, this study addresses Overall Survival (OS) in Non-Small Cell Lung Cancer (NSCLC) using unstructured EHRs data collected from an Italian clinical centre encompassing 838 records from 231 lung cancer patients. Whilst our study is applied on EHRs written in Italian, it serves as use case to prove the effectiveness of extracting and employing high level textual representations that capture relevant information as named entities. Our methodology is interpretable because the hierarchical attention mechanism highlights the information in EHRs that the model considers the most crucial during the decision-making process. We validated this interpretability by measuring the agreement of domain experts on the importance assigned by the hierarchical attention mechanism to EHRs information through a questionnaire. Results demonstrate the effectiveness of our method, showcasing statistically significant improvements over traditional manually extracted clinical features.

HiDe-PET: Continual Learning via Hierarchical Decomposition of Parameter-Efficient Tuning.

The deployment of pre-trained models (PTMs) has greatly advanced the field of continual learning (CL), enabling positive knowledge transfer and resilience to catastrophic forgetting. To sustain these advantages for sequentially arriving tasks, a promising direction involves keeping the pre-trained backbone frozen while employing parameter-efficient tuning (PET) techniques to instruct representation learning. Despite the popularity of Prompt-based PET for CL, its empirical design often leads to sub-optimal performance in our evaluation of different PTMs and target tasks. To this end, we propose a unified framework for CL with PTMs and PET that provides both theoretical and empirical advancements. We first perform an in-depth theoretical analysis of the CL objective in a pre-training context, decomposing it into hierarchical components namely within-task prediction, task-identity inference and task-adaptive prediction. We then present Hierarchical Decomposition PET (HiDe-PET), an innovative approach that explicitly optimizes the decomposed objective through incorporating task-specific and task-shared knowledge via mainstream PET techniques along with efficient recovery of pre-trained representations. Leveraging this framework, we delve into the distinct impacts of implementation strategy, PET technique and PET architecture, as well as adaptive knowledge accumulation amidst pronounced distribution changes. Finally, across various CL scenarios, our approach demonstrates remarkably superior performance over a broad spectrum of recent strong baselines. Our code is available at https://github.com/thu-ml/HiDe-PET.

GRLGRN: graph representation-based learning to infer gene regulatory networks from single-cell RNA-seq data

BackgroundA gene regulatory network (GRN) is a graph-level representation that describes the regulatory relationships between transcription factors and target genes in cells. The reconstruction of GRNs can help investigate cellular dynamics, drug design, and metabolic systems, and the rapid development of single-cell RNA sequencing (scRNA-seq) technology provides important opportunities while posing significant challenges for reconstructing GRNs. A number of methods for inferring GRNs have been proposed in recent years based on traditional machine learning and deep learning algorithms. However, inferring the GRN from scRNA-seq data remains challenging owing to cellular heterogeneity, measurement noise, and data dropout.ResultsIn this study, we propose a deep learning model called graph representational learning GRN (GRLGRN) to infer the latent regulatory dependencies between genes based on a prior GRN and data on the profiles of single-cell gene expressions. GRLGRN uses a graph transformer network to extract implicit links from the prior GRN, and encodes the features of genes by using both an adjacency matrix of implicit links and a matrix of the profile of gene expression. Moreover, it uses attention mechanisms to improve feature extraction, and feeds the refined gene embeddings into an output module to infer gene regulatory relationships. To evaluate the performance of GRLGRN, we compared it with prevalent models and performed ablation experiments on seven cell-line datasets with three ground-truth networks. The results showed that GRLGRN achieved the best predictions in AUROC and AUPRC on 78.6% and 80.9% of the datasets, and achieved an average improvement of 7.3% in AUROC and 30.7% in AUPRC. The interpretation discussion and the network visualization were conducted.ConclusionsThe experimental results and case studies illustrate the considerable performance of GRLGRN in predicting gene interactions and provide interpretability for the prediction tasks, such as identifying hub genes in the network and uncovering implicit links.

VMKLA-UNet: vision Mamba with KAN linear attention U-Net

In the domain of medical image segmentation, while convolutional neural networks (CNNs) and Transformer-based architectures have attained notable success, they continue to face substantial challenges. CNNs are often limited in their ability to capture long-range dependencies, while Transformer models are frequently constrained by significant computational overhead. Recently, the Vision Mamba model, combined with KAN linear attention, has emerged as a highly promising alternative. In this study, we propose a novel model for medical image segmentation, termed VMKLA-UNet. The encoder of this architecture harnesses the VMamba framework, which employs a bidirectional state-space model for global visual context modeling and positional embedding, thus enabling efficient feature extraction and representation learning. For the decoder, we introduce the MKCSA architecture, which incorporates KAN linear attention—rooted in the Mamba framework—alongside a channel-spatial attention mechanism. KAN linear attention substantially mitigates computational complexity while enhancing the model’s capacity to focus on salient regions of interest, thereby facilitating efficient global context comprehension. The channel attention mechanism dynamically modulates the importance of each feature channel, accentuating critical features and bolstering the model’s ability to differentiate between various tissue types or lesion areas. Concurrently, the spatial attention mechanism refines the model’s focus on key regions within the image, enhancing segmentation boundary accuracy and detail resolution. This synergistic integration of channel and spatial attention mechanisms augments the model’s adaptability, leading to superior segmentation performance across diverse lesion types. Extensive experiments on public datasets, including Polyp, ISIC 2017, ISIC 2018, PH2, and Synapse, demonstrate that VMKLA-UNet consistently achieves high segmentation accuracy and robustness, establishing it as a highly effective solution for medical image segmentation tasks.

Scalable multi-modal representation learning networks

Multi-modal representation learning is recognized for its comprehensive interpretation across diverse modalities. Although existing approaches have yielded favorable results, they face challenges in high-order information preservation and out-of-sample data generalization. To tackle these issues, we propose a scalable multi-modal representation learning networks framework, which aims to learn optimal modality-specific projection matrices to project multi-modal features to a shared representation space. Specifically, weight guided modality-wise and row-sparsity driven feature-wise measures are considered to achieve adaptively hierarchical feature selection from the original data. Then, within the unified latent representation space, we employ hypergraph embedding to preserve the intricate high-order local geometric structures within the modality-specific high-dimensional spaces. Finally, we propose a proximal operator-inspired network architecture to resolve the optimization objectives, streamlining the process of feature auto-weighted selection and representation learning. The experimental results highlight the effectiveness and superiority of the proposed method, while online testing on out-of-sample data further demonstrates robust generalization. The code of the proposed method is publicly available at: https://github.com/ZihanFang11/SMMRL.

NeRF-Det++: Incorporating Semantic Cues and Perspective-aware Depth Supervision for Indoor Multi-View 3D Detection.

NeRF-Det has achieved impressive performance in indoor multi-view 3D detection by innovatively utilizing NeRF to enhance representation learning. Despite its notable performance, we uncover three decisive shortcomings in its current design, including semantic ambiguity, inappropriate sampling, and insufficient utilization of depth supervision. To combat the aforementioned problems, we present three corresponding solutions: 1) Semantic Enhancement. We project the freely available 3D segmentation annotations onto the 2D plane and leverage the corresponding 2D semantic maps as the supervision signal, significantly enhancing the semantic awareness of multi-view detectors. 2) Perspective-Aware Sampling. Instead of employing the uniform sampling strategy, we put forward the perspective-aware sampling policy that samples densely near the camera while sparsely in the distance, more effectively collecting the valuable geometric clues. 3) Ordinal Residual Depth Supervision. As opposed to directly regressing the depth values that are difficult to optimize, we divide the depth range of each scene into a fixed number of ordinal bins and reformulate the depth prediction as the combination of the classification of depth bins as well as the regression of the residual depth values, thereby benefiting the depth learning process. The resulting algorithm, NeRF-Det++, has exhibited appealing performance in the ScanNetV2 and ARKITScenes datasets. Notably, in ScanNetV2, NeRF-Det++ outperforms the competitive NeRF-Det by +1.9% in mAP@0.25 and +3.5% in mAP@0.50. The code will be publicly available at https://github.com/mrsempress/NeRF-Detplusplus.

Visual Reinforcement Learning Control With Instance-Reweighted Alignment and Instance-Dimension Uniformity.

Visual reinforcement learning (VRL) has demonstrated remarkable capabilities in learning behaviors directly from intricate high-dimensional visual inputs. Despite these advancements, existing VRL methods still encounter obstacles such as complete collapse and dimensional collapse, resulting in representation degradation and dimensional redundancy. Contrastive learning, while helping to mitigate the complete collapse issue, is prone to the class collision dilemma. To tackle the aforementioned challenges, this article proposes a novel VRL control method with instance-reweighted alignment and instance-dimension uniformity (IAIU). In this VRL control method, the instance-reweighted alignment representation learning is introduced by minimizing the Kullback-Leibler (KL) divergence between the distributions of predicted next state representations and their weighted actual counterparts. By doing so, we aim to align state representations within the same semantic class, thereby effectively alleviating class collision. Meanwhile, an instance-dimension uniformity regularization mechanism is adopted to suppress the collapse phenomenon. This is realized by leveraging the Hilbert-Schmidt independence criterion (HSIC) and standard orthogonal constraint at the instance and dimension levels, respectively, ensuring the extraction of task-relevant state representations. In essence, IAIU's dual-strategy of alignment and uniformity not only addresses the critical issue of class collision but also guarantees uniformity with respect to both the instances and dimensions. Simulation results from the distracting control suite (DCS) benchmark demonstrate IAIU's superior performance, with substantial enhancements in both representational ability and policy efficacy. The code is available at https://github.com/anonymousforcode/IAIU.

PixCon: Pixel-Level Contrastive Learning Revisited

Contrastive image representation learning has been essential for pre-training vision foundation models to deliver excellent transfer learning performance. It was originally developed based on instance discrimination, which focuses on instance-level recognition tasks. Lately, the focus has shifted to directly working on the dense spatial features to improve transfer performance on dense prediction tasks such as object detection and semantic segmentation, for which pixel-level and region-level contrastive learning methods have been proposed. Region-level methods usually employ region-mining algorithms to capture holistic regional semantics and address the issue of semantically inconsistent scene image crops, as they assume that pixel-level learning struggles with both. In this paper, we revisit pixel-level learning’s potential and show that (1) it can effectively and more efficiently learn holistic regional semantics and (2) it intrinsically provides tools to mitigate the impact of semantically inconsistent views involved with scene-level training images. We prove this by proposing PixCon, a pixel-level contrastive learning framework, and testing different positive matching strategies based on this framework to rediscover the potential of pixel-level learning. Additionally, we propose a novel semantic reweighting approach tailored for pixel-level learning-based scene image pre-training, which outperforms or matches the performance of previous region-level methods in object detection and semantic segmentation tasks across multiple benchmarks.

HNF-DDA: subgraph contrastive-driven transformer-style heterogeneous network embedding for drug–disease association prediction

BackgroundDrug–disease association (DDA) prediction aims to identify potential links between drugs and diseases, facilitating the discovery of new therapeutic potentials and reducing the cost and time associated with traditional drug development. However, existing DDA prediction methods often overlook the global relational information provided by other biological entities, and the complex association structure between drug diseases, limiting the potential correlations of drug and disease embeddings.ResultsIn this study, we propose HNF-DDA, a subgraph contrastive-driven transformer-style heterogeneous network embedding model for DDA prediction. Specifically, HNF-DDA adopts all-pairs message passing strategy to capture the global structure of the network, fully integrating multi-omics information. HNF-DDA also proposes the concept of subgraph contrastive learning to capture the local structure of drug-disease subgraphs, learning the high-order semantic information of nodes. Experimental results on two benchmark datasets demonstrate that HNF-DDA outperforms several state-of-the-art methods. Additionally, it shows superior performance across different dataset splitting schemes, indicating HNF-DDA’s capability to generalize to novel drug and disease categories. Case studies for breast cancer and prostate cancer reveal that 9 out of the top 10 predicted candidate drugs for breast cancer and 8 out of the top 10 for prostate cancer have documented therapeutic effects.ConclusionsHNF-DDA incorporates all-pairs message passing and subgraph capture strategies into heterogeneous network embedding, enabling effective learning of drug and disease representations enriched with heterogeneous information, while also demonstrating significant potential for applications in drug repositioning.

QATCH: Automatic Evaluation of SQL-Centric Tasks on Proprietary Data

Tabular Representation Learning (TRL) and Large Language Models (LLMs) have become established for tackling Question Answering (QA) and Semantic Parsing (SP) tasks on tabular data. State-of-the-art models are pre-trained and evaluated on large open-domain datasets. However, the performance on existing QA and SP benchmarks is not necessarily representative of that achieved on proprietary data as the characteristics of the input and the complexity of the posed queries show high variability. To tackle this challenge, our goal is to allow end-users to evaluate TRL and LLM performance on their own proprietary data. We present Query-Aided TRL CHecklist (QATCH), a toolbox to automatically generate a testing checklist tailored to QA and SP. QATCH provides a testing suite highlighting models’ strengths and weaknesses on relational tables unseen at training time. The proposed toolbox relies on a SQL query generator that crafts tests of varying types and complexity including, amongst others, tests on null values, projection, selections, joins, group by, and having clauses. QATCH also supports a set of general cross-task performance metrics providing more insights into SQL-related model capabilities than currently used metrics. The empirical results, achieved by state-of-the-art TRL models and LLMs, show substantial performance differences (1) between existing benchmarks and proprietary data, (2) across queries of different complexity.

Multi-modal Disease Prediction with Hierarchical Self-supervised Learning.

The proliferation of healthcare data sources, including diverse imaging modalities and biochemical measurements, has created unprecedented opportunities for comprehensive disease prediction. Multi-modal clinical data, encompassing medical imaging reports, biochemical assays, and longitudinal clinical records, provides a rich foundation for developing sophisticated diagnostic models. Graph Neural Networks (GNNs) have emerged as a leading methodological framework, distinguished by their capacity to model complex inter-patient relationships and capture community structures within patient data. Despite their promise, current GNN-based approaches exhibit limitations in handling noisy, low-quality data and often impose overly restrictive graph smoothness constraints. These limitations can obscure patient-specific variations and compromise model robustness. To overcome these challenges, we propose HierSSL (Hierarchical Self-Supervised Learning), a novel multi-modal disease prediction framework that enhances representational learning through dual-scale self-supervision mechanisms operating at both local and global levels. HierSSL's architecture specifically addresses two critical aspects: 1) the capture of local inter-modality dependencies and global community patterns, and 2) the optimization of multi-modal feature integration through an innovative combination of feature consistency constraints and graph contrastive learning. Empirical evaluation across two distinct disease prediction datasets demonstrates that HierSSL achieves statistically significant performance improvements compared to state-of-the-art methods, highlighting its efficacy in robust multi-modal data integration for disease prediction tasks.

A coregulatory influence map of glioblastoma heterogeneity and plasticity

We present GBM-cRegMap, an online resource providing a comprehensive coregulatory influence network perspective on glioblastoma (GBM) heterogeneity and plasticity. Using representation learning algorithms, we derived two components of this resource: GBM-CoRegNet, a highly specific coregulatory network of tumor cells, and GBM-CoRegMap, a unified network influence map based on 1612 tumors from 16 studies. As a widely applicable closed-loop system connecting cellular models and tumors, GBM-cRegMap will provide the GBM research community with an easy-to-use web tool (https://gbm.cregmap.com) that maps any existing or newly generated transcriptomic “query” data to a reference coregulatory network and a large-scale manifold of disease heterogeneity. Using GBM-cRegMap, we demonstrated the synergy between the two components by refining the molecular classification of GBM, identifying potential key regulators, and aligning the transcriptional profiles of tumors and in vitro models. Through the amalgamation of a vast dataset, we validated the proneural (PN)-mesenchymal (MES) axis and identified three subclasses of classical (CL) tumors: astrocyte-like (CL-A), epithelial basal-like (CL-B), and cilium-rich (CL-C). We revealed the CL-C subclass, an intermediate state demonstrating the plasticity of GBM cells along the PN-MES axis under chemotherapy. We identified key regulators, such as PAX8, and NKX2.5, potentially involved in temozolomide (TMZ) resistance. Notably, NKX2.5, more expressed in higher-grade gliomas, negatively impacts patient survival, and regulates genes involved in glucose metabolism.

Preventing Posterior Collapse with DVAE for Text Modeling

This paper introduces a novel variational autoencoder model termed DVAE to prevent posterior collapse in text modeling. DVAE employs a dual-path architecture within its decoder: path A and path B. Path A makes the direct input of text instances into the decoder, whereas path B replaces a subset of word tokens in the text instances with a generic unknown token before their input into the decoder. A stopping strategy is implemented, wherein both paths are concurrently active during the early phases of training. As the model progresses towards convergence, path B is removed. To further refine the performance, a KL weight dropout method is employed, which randomly sets certain dimensions of the KL weight to zero during the annealing process. DVAE compels the latent variables to encode more information about the input texts through path B and fully utilize the expressiveness of the decoder, as well as avoiding the local optimum when path B is active through path A and the stopping strategy. Furthermore, the KL weight dropout method augments the number of active units within the latent variables. Experimental results show the excellent performance of DVAE in density estimation, representation learning, and text generation.

Cross-modal Contrastive Learning for Robust Visual Representation in Dynamic Environmental Conditions

This paper proposes a novel cross-modal contrastive learning framework for robust visual representation under dynamic environmental conditions. We address the challenge of maintaining consistent performance across varying environments by introducing a dual-stream architecture that leverages complementary information from visual and contextual modalities. Our framework incorporates three key components: (1) a cross-modal contrastive learning mechanism that establishes correspondences between modalities while preserving their semantic structure, (2) a feature alignment module with cross-modal attention that dynamically aligns features across modalities, and (3) an environmental adaptation strategy with adaptive normalization and memory-augmented learning to enhance robustness against environmental variations. Extensive experiments on three datasets (DynamicVQA, MultiEnv-ImageText, and RobustSceneX) demonstrate that our approach consistently outperforms existing methods, achieving an average improvement of 8.1% in mean Average Precision over state-of-the-art baselines. Ablation studies confirm the contribution of each component, with the full model exhibiting superior performance in cross-condition scenarios. Zero-shot transfer experiments further validate the generalizability of our learned representations to downstream tasks. Our work provides a comprehensive solution for robust visual representation learning in real-world applications where environmental conditions frequently change.

Small object detection in remote sensing images through multi-scale feature fusion

Abstract Due to the challenges posed by background noise and the limited information available for small targets in remote sensing images, the detection performance for such targets remains unsatisfactory. To address these issues and enhance detection accuracy, we propose an improved algorithm based on RTDETR, named Adaptive Selective Transformer. Firstly, in the feature extraction network, we introduce an adaptive convolutional feature enhancement module to improve the multi-scale feature extraction capability in low-resolution remote sensing images. Secondly, we design a multi-scale enhancement structure to extract detailed information from small target images through enhanced multi-scale representation learning, thereby generating target features with stronger discriminative power. Finally, we propose a hierarchical frequency attention mechanism to achieve localized enhancement of contextual awareness, effectively capturing high-frequency local feature information of small targets. Experimental results demonstrate that the Adaptive Selective Transformer achieves superior small target detection performance, validating the effectiveness of our modifications to the original RTDETR model.

Cognitive network enrichment, not degradation, explains the aging mental lexicon and links fluid and crystallized intelligence.

Cognition is a complex system of interacting components. Late-life cognitive decline is often explained as a degradation of the interconnectivity among these components. Evidence from the aging mental lexicon corroborates this interpretation, as older adults produce higher entropy responses in free association tasks, appear to have sparser free association networks, and judge objects to be less similar to one another than younger adults. Here, I demonstrate that all of these effects are produced by a model of cognitive network enrichment, which treats aging as an extension of lifelong learning. By increasing interconnectivity, learning increases competition for activation among potential targets, increasing entropy and reducing targeted activation. The impact of network enrichment is demonstrated using a general prediction error model (Rescorla-Wagner), which learns and enriches a cognitive network representation following lifelong experience with a network of associations in the environment. Sampling from the learned representation to produce behavior reproduces the above effects. A qualitative model comparison shows that various models of degradation fail to capture the above results for entropy and similarity. Both enriched and degraded representations can produce sparsening-free association networks, depending on the specific methodological details of data collection. This underscores the general problem of inferring representation from behavior without considering process. Further, extending cognitive network enrichment more broadly provides a lifelong developmental pathway for overattention to irrelevant stimuli and cognitive slowing-increasing interference, taxing resource limitations, and reducing targeted activation-offering a common cause for rising crystallized intelligence and declining fluid intelligence. (PsycInfo Database Record (c) 2025 APA, all rights reserved).