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.
Probabilistic load forecasting is crucial for modern energy systems, providing point forecasts and uncertainty quantification to enhance decision-making and grid reliability. Achieving accurate forecasts often requires deeper hybrid ensemble models to capture power load complexities. However, increasing model depth can introduce challenges like vanishing gradients, hindering learning efficiency. To address this, we propose an ensemble framework integrating Bidirectional Temporal Convolutional Networks (BiTCN) to capture forward and backward temporal dependencies. An attention mechanism highlights key temporal features, improving accuracy, while skip connections mitigate vanishing gradients by preserving early outputs. A classical neural network models forecast distributions, ensuring a comprehensive uncertainty representation. Compared to state-of-the-art models like WaveNet and Transformers models, our framework achieves a 9.3\% improvement in mean standardized error on a public demand dataset of 500,000 data points. These results demonstrate the effectiveness and robustness of the proposed approach, offering a promising solution for probabilistic load forecasting in power systems.
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.
The intersection of computer vision and machine learning has emerged as a promising avenue for advancing historical research, facilitating a more profound exploration of our past. However, the application of machine learning approaches in historical palaeography is often met with criticism due to their perceived ``black box'' nature. In response to this challenge, we introduce NeuroPapyri, an innovative deep learning-based model specifically designed for the analysis of images containing ancient Greek papyri. To address concerns related to transparency and interpretability, the model incorporates an attention mechanism. This attention mechanism not only enhances the model's performance but also provides a visual representation of the image regions that significantly contribute to the decision-making process. Specifically calibrated for processing images of papyrus documents with lines of handwritten text, the model utilizes individual attention maps to inform the presence or absence of specific characters in the input image. This paper presents the NeuroPapyri model, including its architecture and training methodology. Results from the evaluation demonstrate NeuroPapyri's efficacy in document retrieval, showcasing its potential to advance the analysis of historical manuscripts.
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.
We propose to incorporate Riemannian geometric features from the geometry of aircraft wing surfaces in the prediction of coefficient of pressure (CP) on the aircraft wing. Contrary to existing approaches that treat the wing surface as a flat object, we represent the wing as a piecewise smooth manifold and calculate a set of Riemannian geometric features (Riemannian metric, connection, and curvature) over points of the wing. Combining these features in neighborhoods of points on the wing with coordinates and flight conditions gives inputs to a deep learning model that predicts CP distributions. Experimental results show that the method with incorporation of Riemannian geometric features, compared to state-of-the-art Deep Attention Network (DAN), reduces the predicted mean square error (MSE) of CP by an average of 15.00% for the DLR-F11 aircraft test set.
Missingness is ubiquitous in multivariate time series and poses an obstacle to reliable downstream analysis. Although recurrent network imputation achieved the SOTA, existing models do not scale to deep architectures that can potentially alleviate issues arising in complex data. Moreover, imputation carries the risk of biased estimations of the ground truth. Yet, confidence in the imputed values is always unmeasured or computed post hoc from model output. We propose DEep Attention Recurrent Imputation (DEARI), which jointly estimates missing values and their associated uncertainty in heterogeneous multivariate time series. By jointly representing feature-wise correlations and temporal dynamics, we adopt a self attention mechanism, along with an effective residual component, to achieve a deep recurrent neural network with good imputation performance and stable convergence. We also leverage self-supervised metric learning to boost performance by optimizing sample similarity. Finally, we transform DEARI into a Bayesian neural network through a novel Bayesian marginalization strategy to produce stochastic DEARI, which outperforms its deterministic equivalent. Experiments show that DEARI surpasses the SOTA in diverse imputation tasks using real-world datasets, namely air quality control, healthcare and traffic.
Images corrupted by snowy adverse weather can impose performance impediments to critical high-level vision-based applications. Restoring snow-degraded images is vital, but the task is ill-posed and very challenging due to the veiling effect, stochastic distribution, and multi-scale characteristics of snow in a scene. In this regard, many existing image denoising methods are often less successful with respect to snow removal, being that they mostly achieve success with one snow dataset and underperform in others, thus questioning their robustness in tackling real-world complex snowfall scenarios. In this paper, we propose the wavelet in transformer (WiT) network to address the image desnow inverse problem. Our model exploits the joint systemic capabilities of the vision transformer and the renowned discrete wavelet transform to achieve effective restoration of snow-degraded images. In our experiments, we evaluated the performance of our model on the popular SRRS, SNOW100K, and CSD datasets, respectively. The efficacy of our learning-based network is proven by our obtained numeric and qualitative result outcomes indicating significant performance gains compared to image desnow benchmark models and other state-of-the-art methods in the literature. The source code is available at https://github.com/WINS-lab/WiT.
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.
In deep learning theory, the covariance matrix of the representations serves as a proxy to examine the network's trainability. Motivated by the success of Transformers, we study the covariance matrix of a modified Softmax-based attention model with skip connections in the proportional limit of infinite-depth-and-width. We show that at initialization the limiting distribution can be described by a stochastic differential equation (SDE) indexed by the depth-to-width ratio. To achieve a well-defined stochastic limit, the Transformer's attention mechanism is modified by centering the Softmax output at identity, and scaling the Softmax logits by a width-dependent temperature parameter. We examine the stability of the network through the corresponding SDE, showing how the scale of both the drift and diffusion can be elegantly controlled with the aid of residual connections. The existence of a stable SDE implies that the covariance structure is well-behaved, even for very large depth and width, thus preventing the notorious issues of rank degeneracy in deep attention models. Finally, we show, through simulations, that the SDE provides a surprisingly good description of the corresponding finite-size model. We coin the name shaped Transformer for these architectural modifications.
Traditional machine learning assumes samples in tabular data to be independent and identically distributed (i.i.d). This assumption may miss useful information within and between sample relationships in representation learning. This paper relaxes the i.i.d assumption to learn tabular data representations by incorporating between-sample relationships for the first time using graph neural networks (GNN). We investigate our hypothesis using several GNNs and state-of-the-art (SOTA) deep attention models to learn the between-sample relationship on ten tabular data sets by comparing them to traditional machine learning methods. GNN methods show the best performance on tabular data with large feature-to-sample ratios. Our results reveal that attention-based GNN methods outperform traditional machine learning on five data sets and SOTA deep tabular learning methods on three data sets. Between-sample learning via GNN and deep attention methods yield the best classification accuracy on seven of the ten data sets. This suggests that the i.i.d assumption may not always hold for most tabular data sets.
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.
The learned denoising-based approximate message passing (LDAMP) algorithm has attracted great attention for image compressed sensing (CS) tasks. However, it has two issues: first, its global measurement model severely restricts its applicability to high-dimensional images, and its block-based measurement method exhibits obvious block artifacts; second, the denoiser in the LDAMP is too simple, and existing denoisers have limited ability in detail recovery. In this paper, to overcome the issues and develop a high-performance LDAMP method for image block compressed sensing (BCS), we propose a novel sparsity and coefficient permutation-based AMP (SCP-AMP) method consisting of the block-based sampling and the two-domain reconstruction modules. In the sampling module, SCP-AMP adopts a discrete cosine transform (DCT) based sparsity strategy to reduce the impact of the high-frequency coefficient on the reconstruction, followed by a coefficient permutation strategy to avoid block artifacts. In the reconstruction module, a two-domain AMP method with DCT domain noise correction and pixel domain denoising is proposed for iterative reconstruction. Regarding the denoiser, we proposed a multi-level deep attention network (MDANet) to enhance the texture details by employing multi-level features and multiple attention mechanisms. Extensive experiments demonstrated that the proposed SCP-AMP method achieved better reconstruction accuracy than other state-of-the-art BCS algorithms in terms of both visual perception and objective metrics.
Speaker protection algorithm is to leverage the playback signal properties to prevent over excursion while maintaining maximum loudness, especially for the mobile phone with tiny loudspeakers. This paper proposes efficient DL solutions to accurately model and predict the nonlinear excursion, which is challenging for conventional solutions. Firstly, we build the experiment and pre-processing pipeline, where the feedback current and voltage are sampled as input, and laser is employed to measure the excursion as ground truth. Secondly, one FFTNet model is proposed to explore the dominant low-frequency and other unknown harmonics, and compares to a baseline ConvNet model. In addition, BN re-estimation is designed to explore the online adaptation; and INT8 quantization based on AI Model efficiency toolkit (AIMET\footnote{AIMET is a product of Qualcomm Innovation Center, Inc.}) is applied to further reduce the complexity. The proposed algorithm is verified in two speakers and 3 typical deployment scenarios, and $>$99\% residual DC is less than 0.1 mm, much better than traditional solutions.
Remote sensing image segmentation is a specific task of remote sensing image interpretation. A good remote sensing image segmentation algorithm can provide guidance for environmental protection, agricultural production, and urban construction. This paper proposes a new type of UNet image segmentation algorithm based on channel self attention mechanism and residual connection called . In my experiment, the new network model improved mIOU by 2.48% compared to traditional UNet on the FoodNet dataset. The image segmentation algorithm proposed in this article enhances the internal connections between different items in the image, thus achieving better segmentation results for remote sensing images with occlusion.
Despite the robust security features inherent in the 5G framework, attackers will still discover ways to disrupt 5G unmanned aerial vehicle (UAV) operations and decrease UAV control communication performance in Air-to-Ground (A2G) links. Operating under the assumption that the 5G UAV communications infrastructure will never be entirely secure, we propose Deep Attention Recognition (DAtR) as a solution to identify attacks based on a small deep network embedded in authenticated UAVs. Our proposed solution uses two observable parameters: the Signal-to-Interference-plus-Noise Ratio (SINR) and the Reference Signal Received Power (RSSI) to recognize attacks under Line-of-Sight (LoS), Non-Line-of-Sight (NLoS), and a probabilistic combination of the two conditions. In the tested scenarios, a number of attackers are located in random positions, while their power is varied in each simulation. Moreover, terrestrial users are included in the network to impose additional complexity on attack detection. To improve the systems overall performance in the attack scenarios, we propose complementing the deep network decision with two mechanisms based on data manipulation and majority voting techniques. We compare several performance parameters in our proposed Deep Network. For example, the impact of Long Short-Term-Memory (LSTM) and Attention layers in terms of their overall accuracy, the window size effect, and test the accuracy when only partial data is available in the training process. Finally, we benchmark our deep network with six widely used classifiers regarding classification accuracy. Our algorithms accuracy exceeds 4% compared with the eXtreme Gradient Boosting (XGB) classifier in LoS condition and around 3% in the short distance NLoS condition. Considering the proposed deep network, all other classifiers present lower accuracy than XGB.