PyTorchKR Papers

최신 논문

총 87편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2506.03370

Comparison of different Unique hard attention transformer models by the formal languages they can recognize

This note is a survey of various results on the capabilities of unique hard attention transformers encoders (UHATs) to recognize formal languages. We distinguish between masked vs. non-masked, finite vs. infinite image and general vs. bilinear attention score functions. We recall some relations between these models, as well as a lower bound in terms of first-order logic and an upper bound in terms of circuit complexity.

arXiv:2505.23623

Characterizing the Expressivity of Transformer Language Models

Transformer-based language models (LMs) have achieved widespread empirical success, but their theoretical expressive power remains only partially understood. Prior work often relies on idealized models with assumptions -- such as arbitrary numerical precision and hard attention -- that diverge from real-world transformers. In this work, we provide an exact characterization of fixed-precision transformers with strict future masking and soft attention, an idealization that more closely mirrors practical implementations. We show that these models are precisely as expressive as a specific fragment of linear temporal logic that includes only a single temporal operator: the past operator. We further relate this logic to established classes in formal language theory, automata theory, and algebra, yielding a rich and unified theoretical framework for understanding transformer expressivity. Finally, we present empirical results that align closely with our theory: transformers trained on languages within their theoretical capacity generalize perfectly over lengths, while they consistently fail to generalize on languages beyond it.

arXiv:2505.18948

Exact Expressive Power of Transformers with Padding

Chain of thought is a natural inference-time method for increasing the computational power of transformer-based large language models (LLMs), but comes at the cost of sequential decoding. Are there more efficient alternatives to expand a transformer's expressive power without adding parameters? We consider transformers with padding tokens as a form of parallelizable test-time compute. We show that averaging-hard-attention, masked-pre-norm transformers with polynomial padding converge to precisely the class $\mathsf{TC}^0$ of extremely parallelizable problems. While the $\mathsf{TC}^0$ upper bound was known, proving a matching lower bound had been elusive. Further, our novel analysis reveals the precise expanded power of padded transformers when coupled with another form of inference-time compute, namely dynamically increasing depth via looping. Our core technical contribution is to show how padding helps bring the notions of complete problems and reductions, which have been a cornerstone of classical complexity theory, to the formal study of transformers. Armed with this new tool, we prove that padded transformers with $O(\log^d n)$ looping on inputs of length $n$ recognize exactly the class $\mathsf{TC}^d$ of moderately parallelizable problems. Thus, padding and looping together systematically expand transformers' expressive power: with polylogarithmic looping, padded transformers converge to the class $\mathsf{NC}$, the best that could be expected without losing parallelism (unless $\mathsf{NC} = \mathsf{P}$). Our results thus motivate further exploration of padding and looping as parallelizable alternatives to chain of thought.

arXiv:2505.13191

Emergence of Fixational and Saccadic Movements in a Multi-Level Recurrent Attention Model for Vision

Inspired by foveal vision, hard attention models promise interpretability and parameter economy. However, existing models like the Recurrent Model of Visual Attention (RAM) and Deep Recurrent Attention Model (DRAM) failed to model the hierarchy of human vision system, that compromise on the visual exploration dynamics. As a result, they tend to produce attention that are either overly fixational or excessively saccadic, diverging from human eye movement behavior. In this paper, we propose a Multi-Level Recurrent Attention Model (MRAM), a novel hard attention framework that explicitly models the neural hierarchy of human visual processing. By decoupling the function of glimpse location generation and task execution in two recurrent layers, MRAM emergent a balanced behavior between fixation and saccadic movement. Our results show that MRAM not only achieves more human-like attention dynamics, but also consistently outperforms CNN, RAM and DRAM baselines on standard image classification benchmarks.

arXiv:2505.11199

NoPE: The Counting Power of Transformers with No Positional Encodings

Positional Encodings (PEs) seem to be indispensable for ensuring expressiveness of transformers; without them attention transformers reduce to a bag-of-word model. NoPE-transformers (i.e. with No PEs) with unique hard attention mechanisms were very recently shown to only be able to express regular languages, i.e., with limited counting ability. This paper shows that, with average hard attention mechanisms, NoPE-transformers are still surprisingly expressive: they can express counting languages corresponding to nonnegative integer solutions to multivariate polynomial equations (i.e. Diophantine equations), reasoning about which is well-known to be undecidable. In fact, we provide a precise characterization of languages expressible by Average Hard Attention NoPE-Transformers (NoPE-AHATs): they correspond precisely to what we call \emph{semi-algebraic sets}, i.e., finite unions of sets of nonnegative integer solutions to systems of multivariate polynomial inequations. We obtain several interesting consequences of our characterization. Firstly, NoPE-transformers can express counting properties that are far more complex than established models like simplified counter machines and Petri nets, but cannot express a very simple counting property of PARITY. Secondly, the problem of analyzing NoPE-transformers is undecidable, e.g., whether a given NoPE transformer classifies all input strings in one class. To complement our results, we exhibit a counting language that is not expressible by average hard attention transformers even with arbitrary PEs but is expressible in the circuit complexity class TC$^0$, answering an open problem.

arXiv:2505.00186

Neuroevolution of Self-Attention Over Proto-Objects

Proto-objects - image regions that share common visual properties - offer a promising alternative to traditional attention mechanisms based on rectangular-shaped image patches in neural networks. Although previous work demonstrated that evolving a patch-based hard-attention module alongside a controller network could achieve state-of-the-art performance in visual reinforcement learning tasks, our approach leverages image segmentation to work with higher-level features. By operating on proto-objects rather than fixed patches, we significantly reduce the representational complexity: each image decomposes into fewer proto-objects than regular patches, and each proto-object can be efficiently encoded as a compact feature vector. This enables a substantially smaller self-attention module that processes richer semantic information. Our experiments demonstrate that this proto-object-based approach matches or exceeds the state-of-the-art performance of patch-based implementations with 62% less parameters and 2.6 times less training time.

arXiv:2504.12088

AttentionDrop: A Novel Regularization Method for Transformer Models

Transformer-based architectures achieve state-of-the-art performance across a wide range of tasks in natural language processing, computer vision, and speech. However, their immense capacity often leads to overfitting, especially when training data is limited or noisy. We propose AttentionDrop, a unified family of stochastic regularization techniques that operate directly on the self-attention distributions. We introduces three variants: 1. Hard Attention Masking: randomly zeroes out top-k attention logits per query to encourage diverse context utilization. 2. Blurred Attention Smoothing: applies a dynamic Gaussian convolution over attention logits to diffuse overly peaked distributions. 3. Consistency-Regularized AttentionDrop: enforces output stability under multiple independent AttentionDrop perturbations via a KL-based consistency loss.

arXiv:2503.16963공식 코드 · Papers with Code

Center-guided Classifier for Semantic Segmentation of Remote Sensing Images

Compared with natural images, remote sensing images (RSIs) have the unique characteristic. i.e., larger intraclass variance, which makes semantic segmentation for remote sensing images more challenging. Moreover, existing semantic segmentation models for remote sensing images usually employ a vanilla softmax classifier, which has three drawbacks: (1) non-direct supervision for the pixel representations during training; (2) inadequate modeling ability of parametric softmax classifiers under large intraclass variance; and (3) opaque process of classification decision. In this paper, we propose a novel classifier (called CenterSeg) customized for RSI semantic segmentation, which solves the abovementioned problems with multiple prototypes, direct supervision under Grassmann manifold, and interpretability strategy. Specifically, for each class, our CenterSeg obtains local class centers by aggregating corresponding pixel features based on ground-truth masks, and generates multiple prototypes through hard attention assignment and momentum updating. In addition, we introduce the Grassmann manifold and constrain the joint embedding space of pixel features and prototypes based on two additional regularization terms. Especially, during the inference, CenterSeg can further provide interpretability to the model by restricting the prototype as a sample of the training set. Experimental results on three remote sensing segmentation datasets validate the effectiveness of the model. Besides the superior performance, CenterSeg has the advantages of simplicity, lightweight, compatibility, and interpretability. Code is available at https://github.com/xwmaxwma/rssegmentation.

arXiv:2503.14615

Unique Hard Attention: A Tale of Two Sides

Understanding the expressive power of transformers has recently attracted attention, as it offers insights into their abilities and limitations. Many studies analyze unique hard attention transformers, where attention selects a single position that maximizes the attention scores. When multiple positions achieve the maximum score, either the rightmost or the leftmost of those is chosen. In this paper, we highlight the importance of this seeming triviality. Recently, finite-precision transformers with both leftmost- and rightmost-hard attention were shown to be equivalent to Linear Temporal Logic (LTL). We show that this no longer holds with only leftmost-hard attention -- in that case, they correspond to a \emph{strictly weaker} fragment of LTL. Furthermore, we show that models with leftmost-hard attention are equivalent to \emph{soft} attention, suggesting they may better approximate real-world transformers than right-attention models. These findings refine the landscape of transformer expressivity and underscore the role of attention directionality.

arXiv:2502.02393

Lower Bounds for Chain-of-Thought Reasoning in Hard-Attention Transformers

Chain-of-thought reasoning and scratchpads have emerged as critical tools for enhancing the computational capabilities of transformers. While theoretical results show that polynomial-length scratchpads can extend transformers' expressivity from $TC^0$ to $PTIME$, their required length remains poorly understood. Empirical evidence even suggests that transformers need scratchpads even for many problems in $TC^0$, such as Parity or Multiplication, challenging optimistic bounds derived from circuit complexity. In this work, we initiate the study of systematic lower bounds for the number of CoT steps across different algorithmic problems, in the hard-attention regime. We study a variety of algorithmic problems, and provide bounds that are tight up to logarithmic factors. Overall, these results contribute to emerging understanding of the power and limitations of chain-of-thought reasoning.

arXiv:2501.12997

Ehrenfeucht-Haussler Rank and Chain of Thought

The notion of rank of a Boolean function has been a cornerstone in the theory of PAC learning, enabling quasipolynomial-time learning algorithms for polynomial-size decision trees. We present a novel characterization of rank, grounded in the well-known Transformer architecture. We show that the rank of a function $f$ corresponds to the minimum number of Chain of Thought (CoT) steps required by a single-layer transformer decoder with hard attention to compute $f$. Based on this characterization we establish tight bounds on the number of CoT steps required for specific problems, showing that $\ell$-fold function composition necessitates exactly $\ell$ CoT steps. Furthermore, we analyze the problem of identifying the position of the $k$-th occurrence of 1 in a Boolean sequence, proving that it requires $k$ CoT steps.

arXiv:2412.09925

Simulating Hard Attention Using Soft Attention

We study conditions under which transformers using soft attention can simulate hard attention, that is, effectively focus all attention on a subset of positions. First, we examine several subclasses of languages recognized by hard-attention transformers, which can be defined in variants of linear temporal logic. We demonstrate how soft-attention transformers can compute formulas of these logics using unbounded positional embeddings or temperature scaling. Second, we demonstrate how temperature scaling allows softmax transformers to simulate general hard-attention transformers, using a temperature that depends on the minimum gap between the maximum attention scores and other attention scores.

arXiv:2409.13629

Transformers in Uniform TC$^0$

Previous work has shown that the languages recognized by average-hard attention transformers (AHATs) and softmax-attention transformers (SMATs) are within the circuit complexity class TC$^0$. However, these results assume limited-precision arithmetic: using floating-point numbers with O(log n) bits (where n is the length of the input string), Strobl showed that AHATs can be approximated in L-uniform TC$^0$, and Merrill and Sabharwal showed that SMATs can be approximated in DLOGTIME-uniform TC$^0$. Here, we improve these results, showing that AHATs with no approximation, SMATs with O(poly(n)) bits of floating-point precision, and SMATs with at most $2^{-O(poly(n))}$ absolute error are all in DLOGTIME-uniform TC$^0$.

arXiv:2408.03734

Soft-Hard Attention U-Net Model and Benchmark Dataset for Multiscale Image Shadow Removal

Effective shadow removal is pivotal in enhancing the visual quality of images in various applications, ranging from computer vision to digital photography. During the last decades physics and machine learning -based methodologies have been proposed; however, most of them have limited capacity in capturing complex shadow patterns due to restrictive model assumptions, neglecting the fact that shadows usually appear at different scales. Also, current datasets used for benchmarking shadow removal are composed of a limited number of images with simple scenes containing mainly uniform shadows cast by single objects, whereas only a few of them include both manual shadow annotations and paired shadow-free images. Aiming to address all these limitations in the context of natural scene imaging, including urban environments with complex scenes, the contribution of this study is twofold: a) it proposes a novel deep learning architecture, named Soft-Hard Attention U-net (SHAU), focusing on multiscale shadow removal; b) it provides a novel synthetic dataset, named Multiscale Shadow Removal Dataset (MSRD), containing complex shadow patterns of multiple scales, aiming to serve as a privacy-preserving dataset for a more comprehensive benchmarking of future shadow removal methodologies. Key architectural components of SHAU are the soft and hard attention modules, which along with multiscale feature extraction blocks enable effective shadow removal of different scales and intensities. The results demonstrate the effectiveness of SHAU over the relevant state-of-the-art shadow removal methods across various benchmark datasets, improving the Peak Signal-to-Noise Ratio and Root Mean Square Error for the shadow area by 25.1% and 61.3%, respectively.

arXiv:2407.04400공식 코드 · Papers with Code

Hard-Attention Gates with Gradient Routing for Endoscopic Image Computing

To address overfitting and enhance model generalization in gastroenterological polyp size assessment, our study introduces Feature-Selection Gates (FSG) or Hard-Attention Gates (HAG) alongside Gradient Routing (GR) for dynamic feature selection. This technique aims to boost Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) by promoting sparse connectivity, thereby reducing overfitting and enhancing generalization. HAG achieves this through sparsification with learnable weights, serving as a regularization strategy. GR further refines this process by optimizing HAG parameters via dual forward passes, independently from the main model, to improve feature re-weighting. Our evaluation spanned multiple datasets, including CIFAR-100 for a broad impact assessment and specialized endoscopic datasets (REAL-Colon, Misawa, and SUN) focusing on polyp size estimation, covering over 200 polyps in more than 370,000 frames. The findings indicate that our HAG-enhanced networks substantially enhance performance in both binary and triclass classification tasks related to polyp sizing. Specifically, CNNs experienced an F1 Score improvement to 87.8% in binary classification, while in triclass classification, the ViT-T model reached an F1 Score of 76.5%, outperforming traditional CNNs and ViT-T models. To facilitate further research, we are releasing our codebase, which includes implementations for CNNs, multistream CNNs, ViT, and HAG-augmented variants. This resource aims to standardize the use of endoscopic datasets, providing public training-validation-testing splits for reliable and comparable research in gastroenterological polyp size estimation. The codebase is available at github.com/cosmoimd/feature-selection-gates.

arXiv:2406.17483공식 코드 · Papers with Code

TRIP: Trainable Region-of-Interest Prediction for Hardware-Efficient Neuromorphic Processing on Event-based Vision

Neuromorphic processors are well-suited for efficiently handling sparse events from event-based cameras. However, they face significant challenges in the growth of computing demand and hardware costs as the input resolution increases. This paper proposes the Trainable Region-of-Interest Prediction (TRIP), the first hardware-efficient hard attention framework for event-based vision processing on a neuromorphic processor. Our TRIP framework actively produces low-resolution Region-of-Interest (ROIs) for efficient and accurate classification. The framework exploits sparse events' inherent low information density to reduce the overhead of ROI prediction. We introduced extensive hardware-aware optimizations for TRIP and implemented the hardware-optimized algorithm on the SENECA neuromorphic processor. We utilized multiple event-based classification datasets for evaluation. Our approach achieves state-of-the-art accuracies in all datasets and produces reasonable ROIs with varying locations and sizes. On the DvsGesture dataset, our solution requires 46x less computation than the state-of-the-art while achieving higher accuracy. Furthermore, TRIP enables more than 2x latency and energy improvements on the SENECA neuromorphic processor compared to the conventional solution.

arXiv:2404.02040

Transformers as Transducers

We study the sequence-to-sequence mapping capacity of transformers by relating them to finite transducers, and find that they can express surprisingly large classes of transductions. We do so using variants of RASP, a programming language designed to help people "think like transformers," as an intermediate representation. We extend the existing Boolean variant B-RASP to sequence-to-sequence functions and show that it computes exactly the first-order rational functions (such as string rotation). Then, we introduce two new extensions. B-RASP[pos] enables calculations on positions (such as copying the first half of a string) and contains all first-order regular functions. S-RASP adds prefix sum, which enables additional arithmetic operations (such as squaring a string) and contains all first-order polyregular functions. Finally, we show that masked average-hard attention transformers can simulate S-RASP.

arXiv:2402.08874공식 코드 · Papers with Code

Recurrent Alignment with Hard Attention for Hierarchical Text Rating

While large language models (LLMs) excel at understanding and generating plain text, they are not tailored to handle hierarchical text structures or directly predict task-specific properties such as text rating. In fact, selectively and repeatedly grasping the hierarchical structure of large-scale text is pivotal for deciphering its essence. To this end, we propose a novel framework for hierarchical text rating utilizing LLMs, which incorporates Recurrent Alignment with Hard Attention (RAHA). Particularly, hard attention mechanism prompts a frozen LLM to selectively focus on pertinent leaf texts associated with the root text and generate symbolic representations of their relationships. Inspired by the gradual stabilization of the Markov Chain, recurrent alignment strategy involves feeding predicted ratings iteratively back into the prompts of another trainable LLM, aligning it to progressively approximate the desired target. Experimental results demonstrate that RAHA outperforms existing state-of-the-art methods on three hierarchical text rating datasets. Theoretical and empirical analysis confirms RAHA's ability to gradually converge towards the underlying target through multiple inferences. Additional experiments on plain text rating datasets verify the effectiveness of this Markov-like alignment. Our data and code can be available in https://github.com/ECNU-Text-Computing/Markov-LLM.

arXiv:2401.14089

GQHAN: A Grover-inspired Quantum Hard Attention Network

Numerous current Quantum Machine Learning (QML) models exhibit an inadequacy in discerning the significance of quantum data, resulting in diminished efficacy when handling extensive quantum datasets. Hard Attention Mechanism (HAM), anticipated to efficiently tackle the above QML bottlenecks, encounters the substantial challenge of non-differentiability, consequently constraining its extensive applicability. In response to the dilemma of HAM and QML, a Grover-inspired Quantum Hard Attention Mechanism (GQHAM) consisting of a Flexible Oracle (FO) and an Adaptive Diffusion Operator (ADO) is proposed. Notably, the FO is designed to surmount the non-differentiable issue by executing the activation or masking of Discrete Primitives (DPs) with Flexible Control (FC) to weave various discrete destinies. Based on this, such discrete choice can be visualized with a specially defined Quantum Hard Attention Score (QHAS). Furthermore, a trainable ADO is devised to boost the generality and flexibility of GQHAM. At last, a Grover-inspired Quantum Hard Attention Network (GQHAN) based on QGHAM is constructed on PennyLane platform for Fashion MNIST binary classification. Experimental findings demonstrate that GQHAN adeptly surmounts the non-differentiability hurdle, surpassing the efficacy of extant quantum soft self-attention mechanisms in accuracies and learning ability. In noise experiments, GQHAN is robuster to bit-flip noise in accuracy and amplitude damping noise in learning performance. Predictably, the proposal of GQHAN enriches the Quantum Attention Mechanism (QAM), lays the foundation for future quantum computers to process large-scale data, and promotes the development of quantum computer vision.

arXiv:2401.06430공식 코드 · Papers with Code

Mutual Distillation Learning For Person Re-Identification

With the rapid advancements in deep learning technologies, person re-identification (ReID) has witnessed remarkable performance improvements. However, the majority of prior works have traditionally focused on solving the problem via extracting features solely from a single perspective, such as uniform partitioning, hard attention mechanisms, or semantic masks. While these approaches have demonstrated efficacy within specific contexts, they fall short in diverse situations. In this paper, we propose a novel approach, Mutual Distillation Learning For Person Re-identification (termed as MDPR), which addresses the challenging problem from multiple perspectives within a single unified model, leveraging the power of mutual distillation to enhance the feature representations collectively. Specifically, our approach encompasses two branches: a hard content branch to extract local features via a uniform horizontal partitioning strategy and a Soft Content Branch to dynamically distinguish between foreground and background and facilitate the extraction of multi-granularity features via a carefully designed attention mechanism. To facilitate knowledge exchange between these two branches, a mutual distillation and fusion process is employed, promoting the capability of the outputs of each branch. Extensive experiments are conducted on widely used person ReID datasets to validate the effectiveness and superiority of our approach. Notably, our method achieves an impressive $88.7\%/94.4\%$ in mAP/Rank-1 on the DukeMTMC-reID dataset, surpassing the current state-of-the-art results. Our source code is available at https://github.com/KuilongCui/MDPR.