Pairwise dot-product self-attention is key to the success of transformers that achieve state-of-the-art performance across a variety of applications in language and vision. This dot-product self-attention computes attention weights among the input tokens using Euclidean distance, which makes the model prone to representation collapse and vulnerable to contaminated samples. In this paper, we propose using a Mahalanobis distance metric for computing the attention weights to stretch the underlying feature space in directions of high contextual relevance. In particular, we define a hyper-ellipsoidal neighborhood around each query to increase the attention weights of the tokens lying in the contextually important directions. We term this novel class of attention Elliptical Attention. Our Elliptical Attention provides two benefits: 1) reducing representation collapse and 2) enhancing the model's robustness as Elliptical Attention pays more attention to contextually relevant information rather than focusing on some small subset of informative features. We empirically demonstrate the advantages of Elliptical Attention over the baseline dot-product attention and state-of-the-art attention m
During the training of large Transformer models, attention masks regulate the scope and direction of information flow across a sequence. Numerous mask variants exist, and operators such as FlexAttention already support arbitrary attention masks. Nevertheless, a systematic formal analysis of the information-flow structure induced by arbitrary masks has been missing. This paper develops a complete theoretical framework. We prove that, with sufficient depth, the information flow of a multi-layer Transformer converges to a Hasse diagram -- a directed acyclic graph representing a partial order. Building on this, we recast the design of parallel training tasks as the problem of finding a minimal common supergraph of Hasse diagrams, and we establish a criterion for the minimal common supergraph. This yields a constructive method to derive attention masks directly from a family of tasks. Applying the framework, we design two novel masks: a block-generation attention mask that ensures training-inference consistency (Block Two-Stream Attention), and a fully supervised bidirectional attention mask (Butterfly Attention). These results demonstrate the framework's capacity to discover new struct
Linear attention is an efficient attention mechanism that has recently emerged as a promising alternative to conventional softmax attention. With its ability to process tokens in linear computational complexities, linear attention, in theory, can handle sequences of unlimited length without sacrificing speed, i.e., maintaining a constant training speed for various sequence lengths with a fixed memory consumption. However, due to the issue with cumulative summation (cumsum), current linear attention algorithms cannot demonstrate their theoretical advantage in a causal setting. In this paper, we present Lightning Attention-2, the first linear attention implementation that enables linear attention to realize its theoretical computational benefits. To achieve this, we leverage the thought of tiling, separately handling the intra-block and inter-block components in linear attention calculation. Specifically, we utilize the conventional attention computation mechanism for the intra-blocks and apply linear attention kernel tricks for the inter-blocks. A tiling technique is adopted through both forward and backward procedures to take full advantage of the GPU hardware. We implement our alg
The computational burden of attention in long-context language models has motivated two largely independent lines of work: sparse attention mechanisms that reduce complexity by attending to selected tokens, and gated attention variants that improve training sta-bility while mitigating the attention sink phenomenon. We observe that these approaches address complementary weaknesses and propose Gated Sparse Attention (GSA), an architecture that realizes the benefits of both. GSA incorporates a gated lightning indexer with sigmoid activations that produce bounded, interpretable selection scores, an adaptive sparsity controller that modulates the number of attended tokens based on local uncertainty, and dual gating at the value and output stages. We establish theoretical foundations for the approach, including complexity analysis, expressiveness results, and convergence guarantees. In experiments with 1.7B parameter models trained on 400B tokens, GSA matches the efficiency of sparse-only baselines (12-16x speedup at 128K context) while achieving the quality gains associated with gated attention: perplexity improves from 6.03 to 5.70, RULER scores at 128K context nearly double, and atten
Recent advances in diffusion models have enabled high-quality video generation, but the additional temporal dimension significantly increases computational costs, making training and inference on long videos prohibitively expensive. In this paper, we identify a phenomenon we term Spatiotemporal Energy Decay in video diffusion models: post-softmax attention scores diminish as spatial and temporal distance between tokens increase, akin to the physical decay of signal or waves over space and time in nature. Motivated by this, we propose Radial Attention, a scalable sparse attention mechanism with $\mathcal{O}(n \log n)$ complexity that translates energy decay into exponentially decaying compute density, which is significantly more efficient than standard $\mathcal{O}(n^2)$ dense attention and more expressive than linear attention. Specifically, Radial Attention employs a simple, static attention mask where each token attends to spatially nearby tokens, with the attention window size shrinking with temporal distance. Moreover, it allows pre-trained video diffusion models to extend their generation length with efficient LoRA-based fine-tuning. Extensive experiments show that Radial Atte
Vision-language models (VLMs) typically encode substantially more visual tokens than text tokens, resulting in significant token redundancy. Pruning uninformative visual tokens is therefore crucial for improving computational efficiency, and language-to-vision attention has become a widely used importance criterion for this purpose. However, we find that attention in VLMs is systematically biased. It disproportionately favors tokens appearing later in the sequence, manifesting as over-attention to lower image regions, and assigns inflated scores to semantically empty padding tokens. These behaviors stem from intrinsic recency bias and attention sink effects inherited from large language models (LLMs), and they distort attention-based pruning by preserving irrelevant visual content. To derive a pruning criterion better aligned with semantic relevance, we introduce two lightweight yet effective debiasing techniques that restore the reliability of attention. The first compensates for positional distortions by removing recency-induced attention trends, producing a content-aware and position-agnostic importance measure. The second suppresses attention sink effects by eliminating spuriou
Standard transformer attention computes pairwise token similarity but treats all tokens as equally salient and all positions as equally local, regardless of the informational structure of the input. We identify two complementary inductive biases that standard attention lacks: energy salience (which tokens concentrate informational energy, learned end-to-end without explicit frequency decomposition) and scale-selective locality (how far positional influence extends at each frequency, implemented via Morlet wavelet encoding). We address both with two simple components. Energy-Gated Attention (EGA) gates value aggregation by a learned energy estimate of key token embeddings, computed via a single linear projection; it selects what to attend to. Morlet Positional Encoding (MoPE) replaces fixed sinusoidal encodings with learned Gaussian-windowed wavelets that adapt the joint position-frequency localization to the corpus; it specifies where attention operates at each scale. On TinyShakespeare, EGA alone achieves +0.092 validation loss improvement over standard attention (+0.103 over Phase 1-3 baseline); MoPE alone is -0.032 (below baseline as a standalone encoding); but their combination
Computer vision models have been shown to exhibit and amplify biases across a wide array of datasets and tasks. Existing methods for quantifying bias in classification models primarily focus on dataset distribution and model performance on subgroups, overlooking the internal workings of a model. We introduce the Attention-IoU (Attention Intersection over Union) metric and related scores, which use attention maps to reveal biases within a model's internal representations and identify image features potentially causing the biases. First, we validate Attention-IoU on the synthetic Waterbirds dataset, showing that the metric accurately measures model bias. We then analyze the CelebA dataset, finding that Attention-IoU uncovers correlations beyond accuracy disparities. Through an investigation of individual attributes through the protected attribute of Male, we examine the distinct ways biases are represented in CelebA. Lastly, by subsampling the training set to change attribute correlations, we demonstrate that Attention-IoU reveals potential confounding variables not present in dataset labels.
The core component of attention is the scoring function, which transforms the inputs into low-dimensional queries and keys and takes the dot product of each pair. While the low-dimensional projection improves efficiency, it causes information loss for certain tasks that have intrinsically high-dimensional inputs. Additionally, attention uses the same scoring function for all input pairs, without imposing a distance-dependent compute bias for neighboring tokens in the sequence. In this work, we address these shortcomings by proposing new scoring functions based on computationally efficient structured matrices with high ranks, including Block Tensor-Train (BTT) and contiguous Multi-Level Low Rank (MLR) matrices. On in-context regression tasks with high-dimensional inputs, our proposed scoring functions outperform standard attention for any fixed compute budget. On language modeling, a task that exhibits locality patterns, our MLR-based attention method achieves improved scaling laws compared to both standard attention and variants of sliding window attention. Additionally, we show that both BTT and MLR fall under a broader family of efficient structured matrices capable of encoding e
This paper focuses on the challenging crowd counting task. As large-scale variations often exist within crowd images, neither fixed-size convolution kernel of CNN nor fixed-size attention of recent vision transformers can well handle this kind of variation. To address this problem, we propose a Multifaceted Attention Network (MAN) to improve transformer models in local spatial relation encoding. MAN incorporates global attention from a vanilla transformer, learnable local attention, and instance attention into a counting model. Firstly, the local Learnable Region Attention (LRA) is proposed to assign attention exclusively for each feature location dynamically. Secondly, we design the Local Attention Regularization to supervise the training of LRA by minimizing the deviation among the attention for different feature locations. Finally, we provide an Instance Attention mechanism to focus on the most important instances dynamically during training. Extensive experiments on four challenging crowd counting datasets namely ShanghaiTech, UCF-QNRF, JHU++, and NWPU have validated the proposed method. Codes: https://github.com/LoraLinH/Boosting-Crowd-Counting-via-Multifaceted-Attention.
Transformers are quickly becoming one of the most heavily applied deep learning architectures across modalities, domains, and tasks. In vision, on top of ongoing efforts into plain transformers, hierarchical transformers have also gained significant attention, thanks to their performance and easy integration into existing frameworks. These models typically employ localized attention mechanisms, such as the sliding-window Neighborhood Attention (NA) or Swin Transformer's Shifted Window Self Attention. While effective at reducing self attention's quadratic complexity, local attention weakens two of the most desirable properties of self attention: long range inter-dependency modeling, and global receptive field. In this paper, we introduce Dilated Neighborhood Attention (DiNA), a natural, flexible and efficient extension to NA that can capture more global context and expand receptive fields exponentially at no additional cost. NA's local attention and DiNA's sparse global attention complement each other, and therefore we introduce Dilated Neighborhood Attention Transformer (DiNAT), a new hierarchical vision transformer built upon both. DiNAT variants enjoy significant improvements ove
Recently, a new line of works has emerged to understand and improve self-attention in Transformers by treating it as a kernel machine. However, existing works apply the methods for symmetric kernels to the asymmetric self-attention, resulting in a nontrivial gap between the analytical understanding and numerical implementation. In this paper, we provide a new perspective to represent and optimize self-attention through asymmetric Kernel Singular Value Decomposition (KSVD), which is also motivated by the low-rank property of self-attention normally observed in deep layers. Through asymmetric KSVD, $i$) a primal-dual representation of self-attention is formulated, where the optimization objective is cast to maximize the projection variances in the attention outputs; $ii$) a novel attention mechanism, i.e., Primal-Attention, is proposed via the primal representation of KSVD, avoiding explicit computation of the kernel matrix in the dual; $iii$) with KKT conditions, we prove that the stationary solution to the KSVD optimization in Primal-Attention yields a zero-value objective. In this manner, KSVD optimization can be implemented by simply minimizing a regularization loss, so that low-
Conventional wisdom suggests that pre-training Vision Transformers (ViT) improves downstream performance by learning useful representations. Is this actually true? We investigate this question and find that the features and representations learned during pre-training are not essential. Surprisingly, using only the attention patterns from pre-training (i.e., guiding how information flows between tokens) is sufficient for models to learn high quality features from scratch and achieve comparable downstream performance. We show this by introducing a simple method called attention transfer, where only the attention patterns from a pre-trained teacher ViT are transferred to a student, either by copying or distilling the attention maps. Since attention transfer lets the student learn its own features, ensembling it with a fine-tuned teacher also further improves accuracy on ImageNet. We systematically study various aspects of our findings on the sufficiency of attention maps, including distribution shift settings where they underperform fine-tuning. We hope our exploration provides a better understanding of what pre-training accomplishes and leads to a useful alternative to the standard p
The task of driver attention prediction has drawn considerable interest among researchers in robotics and the autonomous vehicle industry. Driver attention prediction can play an instrumental role in mitigating and preventing high-risk events, like collisions and casualties. However, existing driver attention prediction models neglect the distraction state and intention of the driver, which can significantly influence how they observe their surroundings. To address these issues, we present a new driver attention dataset, CoCAtt (Cognitive-Conditioned Attention). Unlike previous driver attention datasets, CoCAtt includes per-frame annotations that describe the distraction state and intention of the driver. In addition, the attention data in our dataset is captured in both manual and autopilot modes using eye-tracking devices of different resolutions. Our results demonstrate that incorporating the above two driver states into attention modeling can improve the performance of driver attention prediction. To the best of our knowledge, this work is the first to provide autopilot attention data. Furthermore, CoCAtt is currently the largest and the most diverse driver attention dataset in
Attention is typically used to select informative sub-phrases that are used for prediction. This paper investigates the novel use of attention as a form of feature augmentation, i.e, casted attention. We propose Multi-Cast Attention Networks (MCAN), a new attention mechanism and general model architecture for a potpourri of ranking tasks in the conversational modeling and question answering domains. Our approach performs a series of soft attention operations, each time casting a scalar feature upon the inner word embeddings. The key idea is to provide a real-valued hint (feature) to a subsequent encoder layer and is targeted at improving the representation learning process. There are several advantages to this design, e.g., it allows an arbitrary number of attention mechanisms to be casted, allowing for multiple attention types (e.g., co-attention, intra-attention) and attention variants (e.g., alignment-pooling, max-pooling, mean-pooling) to be executed simultaneously. This not only eliminates the costly need to tune the nature of the co-attention layer, but also provides greater extents of explainability to practitioners. Via extensive experiments on four well-known benchmark dat
An essential component of modern recurrent sequence models is the forget gate. While Transformers do not have an explicit recurrent form, we show that a forget gate can be naturally incorporated into Transformers by down-weighting the unnormalized attention scores in a data-dependent way. We name this attention mechanism Forgetting Attention and the resulting model the Forgetting Transformer (FoX). We show that FoX outperforms the Transformer on long-context language modeling, length extrapolation, and short-context downstream tasks, while performing on par with the Transformer on long-context downstream tasks. Moreover, it is compatible with the FlashAttention algorithm and does not require any positional embeddings. Several analyses, including the needle-in-the-haystack test, show that FoX also retains the Transformer's superior long-context capabilities over recurrent sequence models such as Mamba-2, HGRN2, and DeltaNet. We also introduce a "Pro" block design that incorporates some common architectural components in recurrent sequence models and find it significantly improves the performance of both FoX and the Transformer. Our code is available at https://github.com/zhixuan-lin
While the great capability of Transformers significantly boosts prediction accuracy, it could also yield overconfident predictions and require calibrated uncertainty estimation, which can be commonly tackled by Gaussian processes (GPs). Existing works apply GPs with symmetric kernels under variational inference to the attention kernel; however, omitting the fact that attention kernels are in essence asymmetric. Moreover, the complexity of deriving the GP posteriors remains high for large-scale data. In this work, we propose Kernel-Eigen Pair Sparse Variational Gaussian Processes (KEP-SVGP) for building uncertainty-aware self-attention where the asymmetry of attention kernels is tackled by Kernel SVD (KSVD) and a reduced complexity is acquired. Through KEP-SVGP, i) the SVGP pair induced by the two sets of singular vectors from KSVD w.r.t. the attention kernel fully characterizes the asymmetry; ii) using only a small set of adjoint eigenfunctions from KSVD, the derivation of SVGP posteriors can be based on the inversion of a diagonal matrix containing singular values, contributing to a reduction in time complexity; iii) an evidence lower bound is derived so that variational parameter
Modern large language models increasingly require long contexts for reasoning and multi-document tasks, but attention's quadratic complexity creates a severe computational bottleneck. We present Block-Sparse FlashAttention (BSFA), a drop-in replacement that accelerates long-context inference while preserving model quality. Unlike methods that predict importance before computing scores, BSFA computes exact query-key similarities to select the top-k most important value blocks for each query. By comparing per-block maximum scores against calibrated thresholds, we skip approximately 50% of the computation and memory transfers for pruned blocks. Our training-free approach requires only a one-time threshold calibration on a small dataset to learn the per-layer and per-head attention score distributions. We provide a CUDA kernel implementation that can be used as a drop-in replacement for FlashAttention. On Llama-3.1-8B, BSFA achieves up to 1.10x speedup on real-world reasoning benchmarks and up to 1.24x for needle-in-a-haystack retrieval tasks while maintaining above 99% baseline accuracy, with certain configurations even improving accuracy by focusing on the most relevant content, subs
Geometric Attention (GA) specifies an attention layer by four independent inputs: a finite carrier (what indices are addressable), an evidence-kernel rule (how masked proto-scores and a link induce nonnegative weights), a probe family (which observables are treated as admissible), and an anchor/update rule (which representative kernel is selected and how it is applied). Probe families induce an operational equivalence relation on kernels and therefore a gauge; anchors select representatives relative to that probe. Under a scalar relational-work representation and a multiplicative compositionality law for evidence, the admissible link family is exponential, yielding Gibbs weights; with row anchoring this includes the softmax kernel family as a subregime. After quotienting unary row/column score fields, the remaining interaction component admits a canonical rank-r normal form (Eckart-Young/SVD); dot-product score charts implement the corresponding low-rank interaction regime. Fixing the carrier and extensionalizing the update yields the standard fixed-token Transformer attention operator; allowing carrier updates yields adaptive-carrier and staged-depth regimes. The operator language
AI-driven pulse thermography (PT) has become a crucial tool in non-destructive testing (NDT), enabling automatic detection of hidden anomalies in various industrial components. Current state-of-the-art techniques feed segmentation and depth estimation networks compressed PT sequences using either Principal Component Analysis (PCA) or Thermographic Signal Reconstruction (TSR). However, treating these two modalities independently constrains the performance of PT inspection models as these representations possess complementary semantic features. To address this limitation, this work proposes PT-Fusion, a multi-modal attention-based fusion network that fuses both PCA and TSR modalities for defect segmentation and depth estimation of subsurface defects in PT setups. PT-Fusion introduces novel feature fusion modules, Encoder Attention Fusion Gate (EAFG) and Attention Enhanced Decoding Block (AEDB), to fuse PCA and TSR features for enhanced segmentation and depth estimation of subsurface defects. In addition, a novel data augmentation technique is proposed based on random data sampling from thermographic sequences to alleviate the scarcity of PT datasets. The proposed method is benchmarke