SEARCH · Search arXiv

Results for “transformer architecture”

Search indexed arXiv papers on artificial intelligence, large language models, computer vision and robotics. Read source abstracts and follow links to arXiv.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 883 records · Page 49Linked to original sources

arXiv2026

WISCA: A Lightweight Model Transition Method to Improve LLM Training via Weight Scaling

Transformer architecture gradually dominates the LLM field. Recent advances in training optimization for Transformer-based large language models (LLMs) primarily focus on architectural modifications or optimizer adjustments. However, these approaches lack systematic optimization of weight patterns during training. Weight pattern refers to the distribution and relative magnitudes of weight parameters in a neural network. To address this issue, we propose a Weight Scaling method called WISCA to enhance training efficiency and model quality by strategically improving neural network weight patterns without changing network structures. By rescaling weights while preserving model outputs, WISCA indirectly optimizes the model's training trajectory. Experiments demonstrate that WISCA significantly improves convergence quality (measured by generalization capability and loss reduction), particularly in LLMs with Grouped Query Attention (GQA) architectures and LoRA fine-tuning tasks. Empirical results show 5.6% average improvement on zero-shot validation tasks and 2.12% average reduction in training perplexity across multiple architectures.

cs.LG↗

arXiv2024

SmartMem: Layout Transformation Elimination and Adaptation for Efficient DNN Execution on Mobile

This work is motivated by recent developments in Deep Neural Networks, particularly the Transformer architectures underlying applications such as ChatGPT, and the need for performing inference on mobile devices. Focusing on emerging transformers (specifically the ones with computationally efficient Swin-like architectures) and large models (e.g., Stable Diffusion and LLMs) based on transformers, we observe that layout transformations between the computational operators cause a significant slowdown in these applications. This paper presents SmartMem, a comprehensive framework for eliminating most layout transformations, with the idea that multiple operators can use the same tensor layout through careful choice of layout and implementation of operations. Our approach is based on classifying the operators into four groups, and considering combinations of producer-consumer edges between the operators. We develop a set of methods for searching such layouts. Another component of our work is developing efficient memory layouts for 2.5 dimensional memory commonly seen in mobile devices. Our experimental results show that SmartMem outperforms 5 state-of-the-art DNN execution frameworks on mobile devices across 18 varied neural networks, including CNNs, Transformers with both local and global attention, as well as LLMs. In particular, compared to DNNFusion, SmartMem achieves an average speedup of 2.8$\times$, and outperforms TVM and MNN with speedups of 6.9$\times$ and 7.9$\times$, respectively, on average.

cs.LG↗

arXiv2025

CNN-TFT explained by SHAP with multi-head attention weights for time series forecasting

Convolutional neural networks (CNNs) and transformer architectures offer strengths for modeling temporal data: CNNs excel at capturing local patterns and translational invariances, while transformers effectively model long-range dependencies via self-attention. This paper proposes a hybrid architecture integrating convolutional feature extraction with a temporal fusion transformer (TFT) backbone to enhance multivariate time series forecasting. The CNN module first applies a hierarchy of one-dimensional convolutional layers to distill salient local patterns from raw input sequences, reducing noise and dimensionality. The resulting feature maps are then fed into the TFT, which applies multi-head attention to capture both short- and long-term dependencies and to weigh relevant covariates adaptively. We evaluate the CNN-TFT on a hydroelectric natural flow time series dataset. Experimental results demonstrate that CNN-TFT outperforms well-established deep learning models, with a mean absolute percentage error of up to 2.2%. The explainability of the model is obtained by a proposed Shapley additive explanations with multi-head attention weights (SHAP-MHAW). Our novel architecture, named CNN-TFT-SHAP-MHAW, is promising for applications requiring high-fidelity, multivariate time series forecasts, being available for future analysis at https://github.com/SFStefenon/CNN-TFT-SHAP-MHAW .

cs.LG↗

arXiv2025

Boosting Generative Adversarial Transferability with Self-supervised Vision Transformer Features

The ability of deep neural networks (DNNs) come from extracting and interpreting features from the data provided. By exploiting intermediate features in DNNs instead of relying on hard labels, we craft adversarial perturbation that generalize more effectively, boosting black-box transferability. These features ubiquitously come from supervised learning in previous work. Inspired by the exceptional synergy between self-supervised learning and the Transformer architecture, this paper explores whether exploiting self-supervised Vision Transformer (ViT) representations can improve adversarial transferability. We present dSVA -- a generative dual self-supervised ViT features attack, that exploits both global structural features from contrastive learning (CL) and local textural features from masked image modeling (MIM), the self-supervised learning paradigm duo for ViTs. We design a novel generative training framework that incorporates a generator to create black-box adversarial examples, and strategies to train the generator by exploiting joint features and the attention mechanism of self-supervised ViTs. Our findings show that CL and MIM enable ViTs to attend to distinct feature tendencies, which, when exploited in tandem, boast great adversarial generalizability. By disrupting dual deep features distilled by self-supervised ViTs, we are rewarded with remarkable black-box transferability to models of various architectures that outperform state-of-the-arts. Code available at https://github.com/spencerwooo/dSVA.

cs.CV↗

arXiv2021

PGT: Pseudo Relevance Feedback Using a Graph-Based Transformer

Most research on pseudo relevance feedback (PRF) has been done in vector space and probabilistic retrieval models. This paper shows that Transformer-based rerankers can also benefit from the extra context that PRF provides. It presents PGT, a graph-based Transformer that sparsifies attention between graph nodes to enable PRF while avoiding the high computational complexity of most Transformer architectures. Experiments show that PGT improves upon non-PRF Transformer reranker, and it is at least as accurate as Transformer PRF models that use full attention, but with lower computational costs.

cs.IR↗

arXiv2022

On the Surprising Effectiveness of Transformers in Low-Labeled Video Recognition

Recently vision transformers have been shown to be competitive with convolution-based methods (CNNs) broadly across multiple vision tasks. The less restrictive inductive bias of transformers endows greater representational capacity in comparison with CNNs. However, in the image classification setting this flexibility comes with a trade-off with respect to sample efficiency, where transformers require ImageNet-scale training. This notion has carried over to video where transformers have not yet been explored for video classification in the low-labeled or semi-supervised settings. Our work empirically explores the low data regime for video classification and discovers that, surprisingly, transformers perform extremely well in the low-labeled video setting compared to CNNs. We specifically evaluate video vision transformers across two contrasting video datasets (Kinetics-400 and SomethingSomething-V2) and perform thorough analysis and ablation studies to explain this observation using the predominant features of video transformer architectures. We even show that using just the labeled data, transformers significantly outperform complex semi-supervised CNN methods that leverage large-scale unlabeled data as well. Our experiments inform our recommendation that semi-supervised learning video work should consider the use of video transformers in the future.

cs.CV↗

arXiv2024

Gated Linear Attention Transformers with Hardware-Efficient Training

Transformers with linear attention allow for efficient parallel training but can simultaneously be formulated as an RNN with 2D (matrix-valued) hidden states, thus enjoying linear-time inference complexity. However, linear attention generally underperforms ordinary softmax attention. Moreover, current implementations of linear attention lack I/O-awareness and are thus slower than highly optimized implementations of softmax attention. This work describes a hardware-efficient algorithm for linear attention that trades off memory movement against parallelizability. The resulting implementation, dubbed FLASHLINEARATTENTION, is faster than FLASHATTENTION-2 (Dao, 2023) as a standalone layer even on short sequence lengths (e.g., 1K). We then generalize this algorithm to a more expressive variant of linear attention with data-dependent gates. When used as a replacement for the standard attention layer in Transformers, the resulting gated linear attention (GLA) Transformer is found to perform competitively against the LLaMA-architecture Transformer (Touvron et al., 2023) as well recent linear-time-inference baselines such as RetNet (Sun et al., 2023a) and Mamba (Gu & Dao, 2023) on moderate-scale language modeling experiments. GLA Transformer is especially effective at length generalization, enabling a model trained on 2K to generalize to sequences longer than 20K without significant perplexity degradations. For training speed, the GLA Transformer has higher throughput than a similarly-sized Mamba model.

cs.LG↗

arXiv2025

Is a Pure Transformer Effective for Separated and Online Multi-Object Tracking?

Recent advances in Multi-Object Tracking (MOT) have demonstrated significant success in short-term association within the separated tracking-by-detection online paradigm. However, long-term tracking remains challenging. While graph-based approaches address this by modeling trajectories as global graphs, these methods are unsuitable for real-time applications due to their non-online nature. In this paper, we review the concept of trajectory graphs and propose a novel perspective by representing them as directed acyclic graphs. This representation can be described using frame-ordered object sequences and binary adjacency matrices. We observe that this structure naturally aligns with Transformer attention mechanisms, enabling us to model the association problem using a classic Transformer architecture. Based on this insight, we introduce a concise Pure Transformer (PuTR) to validate the effectiveness of Transformer in unifying short- and long-term tracking for separated online MOT. Extensive experiments on four diverse datasets (SportsMOT, DanceTrack, MOT17, and MOT20) demonstrate that PuTR effectively establishes a solid baseline compared to existing foundational online methods while exhibiting superior domain adaptation capabilities. Furthermore, the separated nature enables efficient training and inference, making it suitable for practical applications. Implementation code and trained models are available at https://github.com/chongweiliu/PuTR .

cs.CV↗

arXiv2025

Theory, Analysis, and Best Practices for Sigmoid Self-Attention

Attention is a key part of the transformer architecture. It is a sequence-to-sequence mapping that transforms each sequence element into a weighted sum of values. The weights are typically obtained as the softmax of dot products between keys and queries. Recent work has explored alternatives to softmax attention in transformers, such as ReLU and sigmoid activations. In this work, we revisit sigmoid attention and conduct an in-depth theoretical and empirical analysis. Theoretically, we prove that transformers with sigmoid attention are universal function approximators and benefit from improved regularity compared to softmax attention. Through detailed empirical analysis, we identify stabilization of large initial attention norms during the early stages of training as a crucial factor for the successful training of models with sigmoid attention, outperforming prior attempts. We also introduce FLASHSIGMOID, a hardware-aware and memory-efficient implementation of sigmoid attention yielding a 17% inference kernel speed-up over FLASHATTENTION2 on H100 GPUs. Experiments across language, vision, and speech show that properly normalized sigmoid attention matches the strong performance of softmax attention on a wide range of domains and scales, which previous attempts at sigmoid attention were unable to fully achieve. Our work unifies prior art and establishes best practices for sigmoid attention as a drop-in softmax replacement in transformers.

cs.LG↗

arXiv2024

SDformer: Efficient End-to-End Transformer for Depth Completion

Depth completion aims to predict dense depth maps with sparse depth measurements from a depth sensor. Currently, Convolutional Neural Network (CNN) based models are the most popular methods applied to depth completion tasks. However, despite the excellent high-end performance, they suffer from a limited representation area. To overcome the drawbacks of CNNs, a more effective and powerful method has been presented: the Transformer, which is an adaptive self-attention setting sequence-to-sequence model. While the standard Transformer quadratically increases the computational cost from the key-query dot-product of input resolution which improperly employs depth completion tasks. In this work, we propose a different window-based Transformer architecture for depth completion tasks named Sparse-to-Dense Transformer (SDformer). The network consists of an input module for the depth map and RGB image features extraction and concatenation, a U-shaped encoder-decoder Transformer for extracting deep features, and a refinement module. Specifically, we first concatenate the depth map features with the RGB image features through the input model. Then, instead of calculating self-attention with the whole feature maps, we apply different window sizes to extract the long-range depth dependencies. Finally, we refine the predicted features from the input module and the U-shaped encoder-decoder Transformer module to get the enriching depth features and employ a convolution layer to obtain the dense depth map. In practice, the SDformer obtains state-of-the-art results against the CNN-based depth completion models with lower computing loads and parameters on the NYU Depth V2 and KITTI DC datasets.

cs.CV↗

arXiv2025

SageAttention: Accurate 8-Bit Attention for Plug-and-play Inference Acceleration

The transformer architecture predominates across various models. As the heart of the transformer, attention has a computational complexity of $O(N^2)$, compared to $O(N)$ for linear transformations. When handling large sequence lengths, attention becomes the primary time-consuming component. Although quantization has proven to be an effective method for accelerating model inference, existing quantization methods primarily focus on optimizing the linear layer. In response, we first analyze the feasibility of quantization in attention detailedly. Following that, we propose SageAttention, a highly efficient and accurate quantization method for attention. The OPS (operations per second) of our approach outperforms FlashAttention2 and xformers by about 2.1 times and 2.7 times, respectively. SageAttention also achieves superior accuracy performance over FlashAttention3. Comprehensive experiments confirm that our approach incurs almost no end-to-end metrics loss across diverse models, including those for large language processing, image generation, and video generation. The codes are available at https://github.com/thu-ml/SageAttention.

cs.LG↗

arXiv2024

FastPCI: Motion-Structure Guided Fast Point Cloud Frame Interpolation

Point cloud frame interpolation is a challenging task that involves accurate scene flow estimation across frames and maintaining the geometry structure. Prevailing techniques often rely on pre-trained motion estimators or intensive testing-time optimization, resulting in compromised interpolation accuracy or prolonged inference. This work presents FastPCI that introduces Pyramid Convolution-Transformer architecture for point cloud frame interpolation. Our hybrid Convolution-Transformer improves the local and long-range feature learning, while the pyramid network offers multilevel features and reduces the computation. In addition, FastPCI proposes a unique Dual-Direction Motion-Structure block for more accurate scene flow estimation. Our design is motivated by two facts: (1) accurate scene flow preserves 3D structure, and (2) point cloud at the previous timestep should be reconstructable using reverse motion from future timestep. Extensive experiments show that FastPCI significantly outperforms the state-of-the-art PointINet and NeuralPCI with notable gains (e.g. 26.6% and 18.3% reduction in Chamfer Distance in KITTI), while being more than 10x and 600x faster, respectively. Code is available at https://github.com/genuszty/FastPCI

cs.CV↗

arXiv2025

E2CB2former: Effecitve and Explainable Transformer for CB2 Receptor Ligand Activity Prediction

Accurate prediction of CB2 receptor ligand activity is pivotal for advancing drug discovery targeting this receptor, which is implicated in inflammation, pain management, and neurodegenerative conditions. Although conventional machine learning and deep learning techniques have shown promise, their limited interpretability remains a significant barrier to rational drug design. In this work, we introduce CB2former, a framework that combines a Graph Convolutional Network with a Transformer architecture to predict CB2 receptor ligand activity. By leveraging the Transformer's self attention mechanism alongside the GCN's structural learning capability, CB2former not only enhances predictive performance but also offers insights into the molecular features underlying receptor activity. We benchmark CB2former against diverse baseline models including Random Forest, Support Vector Machine, K Nearest Neighbors, Gradient Boosting, Extreme Gradient Boosting, Multilayer Perceptron, Convolutional Neural Network, and Recurrent Neural Network and demonstrate its superior performance with an R squared of 0.685, an RMSE of 0.675, and an AUC of 0.940. Moreover, attention weight analysis reveals key molecular substructures influencing CB2 receptor activity, underscoring the model's potential as an interpretable AI tool for drug discovery. This ability to pinpoint critical molecular motifs can streamline virtual screening, guide lead optimization, and expedite therapeutic development. Overall, our results showcase the transformative potential of advanced AI approaches exemplified by CB2former in delivering both accurate predictions and actionable molecular insights, thus fostering interdisciplinary collaboration and innovation in drug discovery.

cs.LG↗

arXiv2025

LLM Inference Acceleration via Efficient Operation Fusion

The rapid development of the Transformer-based Large Language Models (LLMs) in recent years has been closely linked to their ever-growing and already enormous sizes. Many LLMs contain hundreds of billions of parameters and require dedicated hardware resources for training and inference. One of the key challenges inherent to the Transformer architecture is the requirement to support numerous non-linear transformations that involves normalization. For instance, each decoder block typically contains at least one Softmax operation and two Layernorms. The computation of the corresponding normalization scaling factors becomes a major bottleneck as it requires spatial collective operations. In other words, when it comes to the computation of denominators for Softmax and Layernorm, all vector elements must be aggregated into a single location, requiring significant communication. These collective operations slow down inference on Transformers by approximately 20%, defeating the whole purpose of distributed in-memory compute. In this work, we propose an extremely efficient technique that can completely hide the overhead caused by such collective operations. Note that each Softmax and Layernorm operation is typically followed by a linear layer. Since non-linear and linear operations are performed on different hardware engines, they can be easily parallelized once the algebra allows such commutation. By leveraging the inherent properties of linear operations, we can defer the normalization of the preceding Softmax and Layernorm until after the linear layer is computed. Now we can compute the collective scaling factors concurrently with the matrix multiplication and completely hide the latency of the former behind the latter. Such parallelization preserves the numerical accuracy while significantly improving the hardware utilization and reducing the overall latency.

cs.CL↗

arXiv2026

Enhancing Alzheimer's Diagnosis: Leveraging Anatomical Landmarks in Graph Convolutional Neural Networks on Tetrahedral Meshes

Alzheimer's disease (AD) is a major neurodegenerative condition that affects millions around the world. As one of the main biomarkers in the AD diagnosis procedure, brain amyloid positivity is typically identified by positron emission tomography (PET), which is costly and invasive. Brain structural magnetic resonance imaging (sMRI) may provide a safer and more convenient solution for the AD diagnosis. Recent advances in geometric deep learning have facilitated sMRI analysis and early diagnosis of AD. However, determining AD pathology, such as brain amyloid deposition, in preclinical stage remains challenging, as less significant morphological changes can be observed. As a result, few AD classification models are generalizable to the brain amyloid positivity classification task. Blood-based biomarkers (BBBMs), on the other hand, have recently achieved remarkable success in predicting brain amyloid positivity and identifying individuals with high risk of being brain amyloid positive. However, individuals in medium risk group still require gold standard tests such as Amyloid PET for further evaluation. Inspired by the recent success of transformer architectures, we propose a geometric deep learning model based on transformer that is both scalable and robust to variations in input volumetric mesh size. Our work introduced a novel tokenization scheme for tetrahedral meshes, incorporating anatomical landmarks generated by a pre-trained Gaussian process model. Our model achieved superior classification performance in AD classification task. In addition, we showed that the model was also generalizable to the brain amyloid positivity prediction with individuals in the medium risk class, where BM alone cannot achieve a clear classification. Our work may enrich geometric deep learning research and improve AD diagnosis accuracy without using expensive and invasive PET scans.

eess.IV↗

arXiv2025

BlockDance: Reuse Structurally Similar Spatio-Temporal Features to Accelerate Diffusion Transformers

Diffusion models have demonstrated impressive generation capabilities, particularly with recent advancements leveraging transformer architectures to improve both visual and artistic quality. However, Diffusion Transformers (DiTs) continue to encounter challenges related to low inference speed, primarily due to the iterative denoising process. To address this issue, we propose BlockDance, a training-free approach that explores feature similarities at adjacent time steps to accelerate DiTs. Unlike previous feature-reuse methods that lack tailored reuse strategies for features at different scales, BlockDance prioritizes the identification of the most structurally similar features, referred to as Structurally Similar Spatio-Temporal (STSS) features. These features are primarily located within the structure-focused blocks of the transformer during the later stages of denoising. BlockDance caches and reuses these highly similar features to mitigate redundant computation, thereby accelerating DiTs while maximizing consistency with the generated results of the original model. Furthermore, considering the diversity of generated content and the varying distributions of redundant features, we introduce BlockDance-Ada, a lightweight decision-making network tailored for instance-specific acceleration. BlockDance-Ada dynamically allocates resources and provides superior content quality. Both BlockDance and BlockDance-Ada have proven effective across various generation tasks and models, achieving accelerations between 25% and 50% while maintaining generation quality.

cs.CV↗

arXiv2025

Making Acoustic Side-Channel Attacks on Noisy Keyboards Viable with LLM-Assisted Spectrograms' "Typo" Correction

The large integration of microphones into devices increases the opportunities for Acoustic Side-Channel Attacks (ASCAs), as these can be used to capture keystrokes' audio signals that might reveal sensitive information. However, the current State-Of-The-Art (SOTA) models for ASCAs, including Convolutional Neural Networks (CNNs) and hybrid models, such as CoAtNet, still exhibit limited robustness under realistic noisy conditions. Solving this problem requires either: (i) an increased model's capacity to infer contextual information from longer sequences, allowing the model to learn that an initially noisily typed word is the same as a futurely collected non-noisy word, or (ii) an approach to fix misidentified information from the contexts, as one does not type random words, but the ones that best fit the conversation context. In this paper, we demonstrate that both strategies are viable and complementary solutions for making ASCAs practical. We observed that no existing solution leverages advanced transformer architectures' power for these tasks and propose that: (i) Visual Transformers (VTs) are the candidate solutions for capturing long-term contextual information and (ii) transformer-powered Large Language Models (LLMs) are the candidate solutions to fix the ``typos'' (mispredictions) the model might make. Thus, we here present the first-of-its-kind approach that integrates VTs and LLMs for ASCAs. We first show that VTs achieve SOTA performance in classifying keystrokes when compared to the previous CNN benchmark. Second, we demonstrate that LLMs can mitigate the impact of real-world noise. Evaluations on the natural sentences revealed that: (i) incorporating LLMs (e.g., GPT-4o) in our ASCA pipeline boosts the performance of error-correction tasks; and (ii) the comparable performance can be attained by a lightweight, fine-tuned smaller LLM (67 times smaller than GPT-4o), using...

cs.CR↗

arXiv2026

MesaNet: Sequence Modeling by Locally Optimal Test-Time Training

Sequence modeling is currently dominated by causal transformer architectures that use softmax self-attention. Although widely adopted, transformers require scaling memory and compute linearly during inference. A recent stream of work linearized the softmax operation, resulting in powerful recurrent neural network (RNN) models with constant memory and compute costs such as DeltaNet, Mamba or xLSTM. These models can be unified by noting that their recurrent layer dynamics can all be derived from an in-context regression objective, approximately optimized through an online learning rule. Here, we join this line of work and introduce a numerically stable, chunkwise parallelizable version of the recently proposed Mesa layer (von Oswald et al., 2024), which could only run sequentially in time and was therefore not scalable. This layer again stems from an in-context loss, but which is now minimized to optimality at every time point using a fast conjugate gradient solver. Through an extensive suite of experiments study up to the billion-parameter scale, we show that optimal test-time training enables reaching lower language modeling perplexity and higher downstream benchmark performance than previous RNNs, especially on tasks requiring long context understanding. This performance gain comes at the cost of additional flops spent during inference time. Our results are therefore intriguingly related to recent trends of increasing test-time compute to improve performance -- here by spending compute to solve sequential optimization problems within the neural network itself.

cs.LG↗