We introduce the attention-indexed model (AIM), a theoretical framework for analyzing learning in deep attention layers. Inspired by multi-index models, AIM captures how token-level outputs emerge from layered bilinear interactions over high-dimensional embeddings. Unlike prior tractable attention models, AIM allows full-width key and query matrices, aligning more closely with practical transformers. Using tools from statistical mechanics and random matrix theory, we derive closed-form predictions for Bayes-optimal generalization error and identify sharp phase transitions as a function of sample complexity, model width, and sequence length. We propose a matching approximate message passing algorithm and show that gradient descent can reach optimal performance. AIM offers a solvable playground for understanding learning in modern attention architectures.
In this paper, we proposed a Multi-modal Collaborative Optimization and Expansion Network (MCO-E Net), to use event modalities to resist challenges such as low light, high exposure, and high dynamic range in single-eye expression recognition tasks. The MCO-E Net introduces two innovative designs: Multi-modal Collaborative Optimization Mamba (MCO-Mamba) and Heterogeneous Collaborative and Expansion Mixture-of-Experts (HCE-MoE). MCO-Mamba, building upon Mamba, leverages dual-modal information to jointly optimize the model, facilitating collaborative interaction and fusion of modal semantics. This approach encourages the model to balance the learning of both modalities and harness their respective strengths. HCE-MoE, on the other hand, employs a dynamic routing mechanism to distribute structurally varied experts (deep, attention, and focal), fostering collaborative learning of complementary semantics. This heterogeneous architecture systematically integrates diverse feature extraction paradigms to comprehensively capture expression semantics. Extensive experiments demonstrate that our proposed network achieves competitive performance in the task of single-eye expression recognition, especially under poor lighting conditions.
In this manuscript, we study the learning of deep attention neural networks, defined as the composition of multiple self-attention layers, with tied and low-rank weights. We first establish a mapping of such models to sequence multi-index models, a generalization of the widely studied multi-index model to sequential covariates, for which we establish a number of general results. In the context of Bayesian-optimal learning, in the limit of large dimension $D$ and commensurably large number of samples $N$, we derive a sharp asymptotic characterization of the optimal performance as well as the performance of the best-known polynomial-time algorithm for this setting --namely approximate message-passing--, and characterize sharp thresholds on the minimal sample complexity required for better-than-random prediction performance. Our analysis uncovers, in particular, how the different layers are learned sequentially. Finally, we discuss how this sequential learning can also be observed in a realistic setup.
Learning representations on large graphs is a long-standing challenge due to the inter-dependence nature. Transformers recently have shown promising performance on small graphs thanks to its global attention for capturing all-pair interactions beyond observed structures. Existing approaches tend to inherit the spirit of Transformers in language and vision tasks, and embrace complicated architectures by stacking deep attention-based propagation layers. In this paper, we attempt to evaluate the necessity of adopting multi-layer attentions in Transformers on graphs, which considerably restricts the efficiency. Specifically, we analyze a generic hybrid propagation layer, comprised of all-pair attention and graph-based propagation, and show that multi-layer propagation can be reduced to one-layer propagation, with the same capability for representation learning. It suggests a new technical path for building powerful and efficient Transformers on graphs, particularly through simplifying model architectures without sacrificing expressiveness. As exemplified by this work, we propose a Simplified Single-layer Graph Transformers (SGFormer), whose main component is a single-layer global attention that scales linearly w.r.t. graph sizes and requires none of any approximation for accommodating all-pair interactions. Empirically, SGFormer successfully scales to the web-scale graph ogbn-papers100M, yielding orders-of-magnitude inference acceleration over peer Transformers on medium-sized graphs, and demonstrates competitiveness with limited labeled data.
Adverse conditions like snow, rain, nighttime, and fog, pose challenges for autonomous driving perception systems. Existing methods have limited effectiveness in improving essential computer vision tasks, such as semantic segmentation, and often focus on only one specific condition, such as removing rain or translating nighttime images into daytime ones. To address these limitations, we propose a method to improve the visual quality and clarity degraded by such adverse conditions. Our method, AllWeather-Net, utilizes a novel hierarchical architecture to enhance images across all adverse conditions. This architecture incorporates information at three semantic levels: scene, object, and texture, by discriminating patches at each level. Furthermore, we introduce a Scaled Illumination-aware Attention Mechanism (SIAM) that guides the learning towards road elements critical for autonomous driving perception. SIAM exhibits robustness, remaining unaffected by changes in weather conditions or environmental scenes. AllWeather-Net effectively transforms images into normal weather and daytime scenes, demonstrating superior image enhancement results and subsequently enhancing the performance of semantic segmentation, with up to a 5.3% improvement in mIoU in the trained domain. We also show our model's generalization ability by applying it to unseen domains without re-training, achieving up to 3.9% mIoU improvement. Code can be accessed at: https://github.com/Jumponthemoon/AllWeatherNet.
We consider function spaces defined by self-attention networks without normalization, and theoretically analyze their geometry. Since these networks are polynomial, we rely on tools from algebraic geometry. In particular, we study the identifiability of deep attention by providing a description of the generic fibers of the parametrization for an arbitrary number of layers and, as a consequence, compute the dimension of the function space. Additionally, for a single-layer model, we characterize the singular and boundary points. Finally, we formulate a conjectural extension of our results to normalized self-attention networks, prove it for a single layer, and numerically verify it in the deep case.
Autonomous Vehicle (AV) decision making in urban environments is inherently challenging due to the dynamic interactions with surrounding vehicles. For safe planning, AV must understand the weightage of various spatiotemporal interactions in a scene. Contemporary works use colossal transformer architectures to encode interactions mainly for trajectory prediction, resulting in increased computational complexity. To address this issue without compromising spatiotemporal understanding and performance, we propose the simple Deep Attention Driven Reinforcement Learning (DADRL) framework, which dynamically assigns and incorporates the significance of surrounding vehicles into the ego's RL driven decision making process. We introduce an AV centric spatiotemporal attention encoding (STAE) mechanism for learning the dynamic interactions with different surrounding vehicles. To understand map and route context, we employ a context encoder to extract features from context maps. The spatiotemporal representations combined with contextual encoding provide a comprehensive state representation. The resulting model is trained using the Soft Actor Critic (SAC) algorithm. We evaluate the proposed framework on the SMARTS urban benchmarking scenarios without traffic signals to demonstrate that DADRL outperforms recent state of the art methods. Furthermore, an ablation study underscores the importance of the context-encoder and spatio temporal attention encoder in achieving superior performance.
Tailoring treatment for individual patients is crucial yet challenging in order to achieve optimal healthcare outcomes. Recent advances in reinforcement learning offer promising personalized treatment recommendations; however, they rely solely on current patient observations (vital signs, demographics) as the patient's state, which may not accurately represent the true health status of the patient. This limitation hampers policy learning and evaluation, ultimately limiting treatment effectiveness. In this study, we propose the Deep Attention Q-Network for personalized treatment recommendations, utilizing the Transformer architecture within a deep reinforcement learning framework to efficiently incorporate all past patient observations. We evaluated the model on real-world sepsis and acute hypotension cohorts, demonstrating its superiority to state-of-the-art models. The source code for our model is available at https://github.com/stevenmsm/RL-ICU-DAQN.
Graph neural networks (GNNs) learn the representation of graph-structured data, and their expressiveness can be further enhanced by inferring node relations for propagation. Attention-based GNNs infer neighbor importance to manipulate the weight of its propagation. Despite their popularity, the discussion on deep graph attention and its unique challenges has been limited. In this work, we investigate some problematic phenomena related to deep graph attention, including vulnerability to over-smoothed features and smooth cumulative attention. Through theoretical and empirical analyses, we show that various attention-based GNNs suffer from these problems. Motivated by our findings, we propose AEROGNN, a novel GNN architecture designed for deep graph attention. AERO-GNN provably mitigates the proposed problems of deep graph attention, which is further empirically demonstrated with (a) its adaptive and less smooth attention functions and (b) higher performance at deep layers (up to 64). On 9 out of 12 node classification benchmarks, AERO-GNN outperforms the baseline GNNs, highlighting the advantages of deep graph attention. Our code is available at https://github.com/syleeheal/AERO-GNN.
Existing deep embedding clustering works only consider the deepest layer to learn a feature embedding and thus fail to well utilize the available discriminative information from cluster assignments, resulting performance limitation. To this end, we propose a novel method, namely deep attention-guided graph clustering with dual self-supervision (DAGC). Specifically, DAGC first utilizes a heterogeneity-wise fusion module to adaptively integrate the features of an auto-encoder and a graph convolutional network in each layer and then uses a scale-wise fusion module to dynamically concatenate the multi-scale features in different layers. Such modules are capable of learning a discriminative feature embedding via an attention-based mechanism. In addition, we design a distribution-wise fusion module that leverages cluster assignments to acquire clustering results directly. To better explore the discriminative information from the cluster assignments, we develop a dual self-supervision solution consisting of a soft self-supervision strategy with a triplet Kullback-Leibler divergence loss and a hard self-supervision strategy with a pseudo supervision loss. Extensive experiments validate that our method consistently outperforms state-of-the-art methods on six benchmark datasets. Especially, our method improves the ARI by more than 18.14% over the best baseline.
Annotating cancerous regions in whole-slide images (WSIs) of pathology samples plays a critical role in clinical diagnosis, biomedical research, and machine learning algorithms development. However, generating exhaustive and accurate annotations is labor-intensive, challenging, and costly. Drawing only coarse and approximate annotations is a much easier task, less costly, and it alleviates pathologists' workload. In this paper, we study the problem of refining these approximate annotations in digital pathology to obtain more accurate ones. Some previous works have explored obtaining machine learning models from these inaccurate annotations, but few of them tackle the refinement problem where the mislabeled regions should be explicitly identified and corrected, and all of them require a -- often very large -- number of training samples. We present a method, named Label Cleaning Multiple Instance Learning (LC-MIL), to refine coarse annotations on a single WSI without the need of external training data. Patches cropped from a WSI with inaccurate labels are processed jointly within a multiple instance learning framework, mitigating their impact on the predictive model and refining the segmentation. Our experiments on a heterogeneous WSI set with breast cancer lymph node metastasis, liver cancer, and colorectal cancer samples show that LC-MIL significantly refines the coarse annotations, outperforming state-of-the-art alternatives, even while learning from a single slide. Moreover, we demonstrate how real annotations drawn by pathologists can be efficiently refined and improved by the proposed approach. All these results demonstrate that LC-MIL is a promising, light-weight tool to provide fine-grained annotations from coarsely annotated pathology sets.
Thermal images model the long-infrared range of the electromagnetic spectrum and provide meaningful information even when there is no visible illumination. Yet, unlike imagery that represents radiation from the visible continuum, infrared images are inherently low-resolution due to hardware constraints. The restoration of thermal images is critical for applications that involve safety, search and rescue, and military operations. In this paper, we introduce a system to efficiently reconstruct thermal images. Specifically, we explore how to effectively attend to contrasting receptive fields (RFs) where increasing the RFs of a network can be computationally expensive. For this purpose, we introduce a deep attention to varying receptive fields network (AVRFN). We supply a gated convolutional layer with higher-order information extracted from disparate RFs, whereby an RF is parameterized by a dilation rate. In this way, the dilation rate can be tuned to use fewer parameters thus increasing the efficacy of AVRFN. Our experimental results show an improvement over the state of the art when compared against competing thermal image super-resolution methods.
Attention is a general reasoning mechanism than can flexibly deal with image information, but its memory requirements had made it so far impractical for high resolution image generation. We present Grid Partitioned Attention (GPA), a new approximate attention algorithm that leverages a sparse inductive bias for higher computational and memory efficiency in image domains: queries attend only to few keys, spatially close queries attend to close keys due to correlations. Our paper introduces the new attention layer, analyzes its complexity and how the trade-off between memory usage and model power can be tuned by the hyper-parameters.We will show how such attention enables novel deep learning architectures with copying modules that are especially useful for conditional image generation tasks like pose morphing. Our contributions are (i) algorithm and code1of the novel GPA layer, (ii) a novel deep attention-copying architecture, and (iii) new state-of-the art experimental results in human pose morphing generation benchmarks.
Understanding linguistics and morphology of resource-scarce code-mixed texts remains a key challenge in text processing. Although word embedding comes in handy to support downstream tasks for low-resource languages, there are plenty of scopes in improving the quality of language representation particularly for code-mixed languages. In this paper, we propose HIT, a robust representation learning method for code-mixed texts. HIT is a hierarchical transformer-based framework that captures the semantic relationship among words and hierarchically learns the sentence-level semantics using a fused attention mechanism. HIT incorporates two attention modules, a multi-headed self-attention and an outer product attention module, and computes their weighted sum to obtain the attention weights. Our evaluation of HIT on one European (Spanish) and five Indic (Hindi, Bengali, Tamil, Telugu, and Malayalam) languages across four NLP tasks on eleven datasets suggests significant performance improvement against various state-of-the-art systems. We further show the adaptability of learned representation across tasks in a transfer learning setup (with and without fine-tuning).
Attention based neural networks are state of the art in a large range of applications. However, their performance tends to degrade when the number of layers increases. In this work, we show that enforcing Lipschitz continuity by normalizing the attention scores can significantly improve the performance of deep attention models. First, we show that, for deep graph attention networks (GAT), gradient explosion appears during training, leading to poor performance of gradient-based training algorithms. To address this issue, we derive a theoretical analysis of the Lipschitz continuity of attention modules and introduce LipschitzNorm, a simple and parameter-free normalization for self-attention mechanisms that enforces the model to be Lipschitz continuous. We then apply LipschitzNorm to GAT and Graph Transformers and show that their performance is substantially improved in the deep setting (10 to 30 layers). More specifically, we show that a deep GAT model with LipschitzNorm achieves state of the art results for node label prediction tasks that exhibit long-range dependencies, while showing consistent improvements over their unnormalized counterparts in benchmark node classification tasks.
Can we leverage high-resolution information without the unsustainable quadratic complexity to input scale? We propose Traversal Network (TNet), a novel multi-scale hard-attention architecture, which traverses image scale-space in a top-down fashion, visiting only the most informative image regions along the way. TNet offers an adjustable trade-off between accuracy and complexity, by changing the number of attended image locations. We compare our model against hard-attention baselines on ImageNet, achieving higher accuracy with less resources (FLOPs, processing time and memory). We further test our model on fMoW dataset, where we process satellite images of size up to $896 \times 896$ px, getting up to $2.5$x faster processing compared to baselines operating on the same resolution, while achieving higher accuracy as well. TNet is modular, meaning that most classification models could be adopted as its backbone for feature extraction, making the reported performance gains orthogonal to benefits offered by existing optimized deep models. Finally, hard-attention guarantees a degree of interpretability to our model's predictions, without any extra cost beyond inference. Code is available at $\href{https://github.com/Tpap/TNet}{github.com/Tpap/TNet}$.
We present a multi-stage 3D computer-aided detection and diagnosis (CAD) model for automated localization of clinically significant prostate cancer (csPCa) in bi-parametric MR imaging (bpMRI). Deep attention mechanisms drive its detection network, targeting salient structures and highly discriminative feature dimensions across multiple resolutions. Its goal is to accurately identify csPCa lesions from indolent cancer and the wide range of benign pathology that can afflict the prostate gland. Simultaneously, a decoupled residual classifier is used to achieve consistent false positive reduction, without sacrificing high sensitivity or computational efficiency. In order to guide model generalization with domain-specific clinical knowledge, a probabilistic anatomical prior is used to encode the spatial prevalence and zonal distinction of csPCa. Using a large dataset of 1950 prostate bpMRI paired with radiologically-estimated annotations, we hypothesize that such CNN-based models can be trained to detect biopsy-confirmed malignancies in an independent cohort. For 486 institutional testing scans, the 3D CAD system achieves 83.69$\pm$5.22% and 93.19$\pm$2.96% detection sensitivity at 0.50 and 1.46 false positive(s) per patient, respectively, with 0.882$\pm$0.030 AUROC in patient-based diagnosis $-$significantly outperforming four state-of-the-art baseline architectures (U-SEResNet, UNet++, nnU-Net, Attention U-Net) from recent literature. For 296 external biopsy-confirmed testing scans, the ensembled CAD system shares moderate agreement with a consensus of expert radiologists (76.69%; $kappa$ $=$ 0.51$\pm$0.04) and independent pathologists (81.08%; $kappa$ $=$ 0.56$\pm$0.06); demonstrating strong generalization to histologically-confirmed csPCa diagnosis.
We present BRIDGE, a powerful sequential architecture for modeling dependencies between natural language questions and relational databases in cross-DB semantic parsing. BRIDGE represents the question and DB schema in a tagged sequence where a subset of the fields are augmented with cell values mentioned in the question. The hybrid sequence is encoded by BERT with minimal subsequent layers and the text-DB contextualization is realized via the fine-tuned deep attention in BERT. Combined with a pointer-generator decoder with schema-consistency driven search space pruning, BRIDGE attained state-of-the-art performance on popular cross-DB text-to-SQL benchmarks, Spider (71.1\% dev, 67.5\% test with ensemble model) and WikiSQL (92.6\% dev, 91.9\% test). Our analysis shows that BRIDGE effectively captures the desired cross-modal dependencies and has the potential to generalize to more text-DB related tasks. Our implementation is available at \url{https://github.com/salesforce/TabularSemanticParsing}.
We introduce PREDATOR, a model for pairwise point-cloud registration with deep attention to the overlap region. Different from previous work, our model is specifically designed to handle (also) point-cloud pairs with low overlap. Its key novelty is an overlap-attention block for early information exchange between the latent encodings of the two point clouds. In this way the subsequent decoding of the latent representations into per-point features is conditioned on the respective other point cloud, and thus can predict which points are not only salient, but also lie in the overlap region between the two point clouds. The ability to focus on points that are relevant for matching greatly improves performance: PREDATOR raises the rate of successful registrations by more than 20% in the low-overlap scenario, and also sets a new state of the art for the 3DMatch benchmark with 89% registration recall.
Deep learning applied to weather forecasting has started gaining popularity because of the progress achieved by data-driven models. The present paper compares two different deep learning architectures to perform weather prediction on daily data gathered from 18 cities across Europe and spanned over a period of 15 years. We propose the Deep Attention Unistream Multistream (DAUM) networks that investigate different types of input representations (i.e. tensorial unistream vs. multistream ) as well as the incorporation of the attention mechanism. In particular, we show that adding a self-attention block within the models increases the overall forecasting performance. Furthermore, visualization techniques such as occlusion analysis and score maximization are used to give an additional insight on the most important features and cities for predicting a particular target feature of target cities.