The advent of scalable deep models and large datasets has improved the performance of Neural Machine Translation. Knowledge Distillation (KD) enhances efficiency by transferring knowledge from a teacher model to a more compact student model. However, KD approaches to Transformer architecture often rely on heuristics, particularly when deciding which teacher layers to distill from. In this paper, we introduce the 'Align-to-Distill' (A2D) strategy, designed to address the feature mapping problem by adaptively aligning student attention heads with their teacher counterparts during training. The Attention Alignment Module in A2D performs a dense head-by-head comparison between student and teacher attention heads across layers, turning the combinatorial mapping heuristics into a learning problem. Our experiments show the efficacy of A2D, demonstrating gains of up to +3.61 and +0.63 BLEU points for WMT-2022 De->Dsb and WMT-2014 En->De, respectively, compared to Transformer baselines.
相關內容
通過學習、實踐或探索所獲得的認識、判斷或技能。
ZKP systems have surged attention and held a fundamental role in contemporary cryptography. Zk-SNARK protocols dominate the ZKP usage, often implemented through arithmetic circuit programming paradigm. However, underconstrained or overconstrained circuits may lead to bugs. Underconstrained circuits refer to circuits that lack the necessary constraints, resulting in unexpected solutions in the circuit and causing the verifier to accept a bogus witness. Overconstrained circuits refer to circuits that are constrained excessively, resulting in the circuit lacking necessary solutions and causing the verifier to accept no witness, rendering the circuit meaningless. This paper introduces a novel approach for pinpointing two distinct types of bugs in ZKP circuits. The method involves encoding the arithmetic circuit constraints to polynomial equation systems and solving polynomial equation systems over a finite field by algebraic computation. The classification of verification results is refined, greatly enhancing the expressive power of the system. We proposed a tool, AC4, to represent the implementation of this method. Experiments demonstrate that AC4 represents a substantial 29% increase in the checked ratio compared to prior work. Within a solvable range, the checking time of AC4 has also exhibited noticeable improvement, demonstrating a magnitude increase compared to previous efforts.
We introduce a new architecture for personalization of text-to-image diffusion models, coined Mixture-of-Attention (MoA). Inspired by the Mixture-of-Experts mechanism utilized in large language models (LLMs), MoA distributes the generation workload between two attention pathways: a personalized branch and a non-personalized prior branch. MoA is designed to retain the original model's prior by fixing its attention layers in the prior branch, while minimally intervening in the generation process with the personalized branch that learns to embed subjects in the layout and context generated by the prior branch. A novel routing mechanism manages the distribution of pixels in each layer across these branches to optimize the blend of personalized and generic content creation. Once trained, MoA facilitates the creation of high-quality, personalized images featuring multiple subjects with compositions and interactions as diverse as those generated by the original model. Crucially, MoA enhances the distinction between the model's pre-existing capability and the newly augmented personalized intervention, thereby offering a more disentangled subject-context control that was previously unattainable. Project page: //snap-research.github.io/mixture-of-attention
Graph Transformers (GTs) have significantly advanced the field of graph representation learning by overcoming the limitations of message-passing graph neural networks (GNNs) and demonstrating promising performance and expressive power. However, the quadratic complexity of self-attention mechanism in GTs has limited their scalability, and previous approaches to address this issue often suffer from expressiveness degradation or lack of versatility. To address this issue, we propose AnchorGT, a novel attention architecture for GTs with global receptive field and almost linear complexity, which serves as a flexible building block to improve the scalability of a wide range of GT models. Inspired by anchor-based GNNs, we employ structurally important $k$-dominating node set as anchors and design an attention mechanism that focuses on the relationship between individual nodes and anchors, while retaining the global receptive field for all nodes. With its intuitive design, AnchorGT can easily replace the attention module in various GT models with different network architectures and structural encodings, resulting in reduced computational overhead without sacrificing performance. In addition, we theoretically prove that AnchorGT attention can be strictly more expressive than Weisfeiler-Lehman test, showing its superiority in representing graph structures. Our experiments on three state-of-the-art GT models demonstrate that their AnchorGT variants can achieve better results while being faster and significantly more memory efficient.
Background:Technical systems are growing in complexity with more components and functions across various disciplines. Model-Driven Engineering (MDE) helps manage this complexity by using models as key artifacts. Domain-Specific Languages (DSL) supported by MDE facilitate modeling. As data generation in product development increases, there's a growing demand for AI algorithms, which can be challenging to implement. Integrating AI algorithms with DSL and MDE can streamline this process. Objective:This study aims to investigate the existing model-driven approaches relying on DSL in support of the engineering of AI software systems to sharpen future research further and define the current state of the art. Method:We conducted a Systemic Literature Review (SLR), collecting papers from five major databases resulting in 1335 candidate studies, eventually retaining 18 primary studies. Each primary study will be evaluated and discussed with respect to the adoption of MDE principles and practices and the phases of AI development support aligned with the stages of the CRISP-DM methodology. Results:The study's findings show that language workbenches are of paramount importance in dealing with all aspects of modeling language development and are leveraged to define DSL explicitly addressing AI concerns. The most prominent AI-related concerns are training and modeling of the AI algorithm, while minor emphasis is given to the time-consuming preparation of the data. Early project phases that support interdisciplinary communication of requirements, e.g., CRISP-DM Business Understanding phase, are rarely reflected. Conclusion:The study found that the use of MDE for AI is still in its early stages, and there is no single tool or method that is widely used. Additionally, current approaches tend to focus on specific stages of development rather than providing support for the entire development process.
In the last three decades, the Steered Response Power (SRP) method has been widely used for the task of Sound Source Localization (SSL), due to its satisfactory localization performance on moderately reverberant and noisy scenarios. Many works have analyzed and extended the original SRP method to reduce its computational cost, to allow it to locate multiple sources, or to improve its performance in adverse environments. In this work, we review over 200 papers on the SRP method and its variants, with emphasis on the SRP-PHAT method. We also present eXtensible-SRP, or X-SRP, a generalized and modularized version of the SRP algorithm which allows the reviewed extensions to be implemented. We provide a Python implementation of the algorithm which includes selected extensions from the literature.
Facial Expression Recognition (FER) plays a pivotal role in understanding human emotional cues. However, traditional FER methods based on visual information have some limitations, such as preprocessing, feature extraction, and multi-stage classification procedures. These not only increase computational complexity but also require a significant amount of computing resources. Considering Convolutional Neural Network (CNN)-based FER schemes frequently prove inadequate in identifying the deep, long-distance dependencies embedded within facial expression images, and the Transformer's inherent quadratic computational complexity, this paper presents the FER-YOLO-Mamba model, which integrates the principles of Mamba and YOLO technologies to facilitate efficient coordination in facial expression image recognition and localization. Within the FER-YOLO-Mamba model, we further devise a FER-YOLO-VSS dual-branch module, which combines the inherent strengths of convolutional layers in local feature extraction with the exceptional capability of State Space Models (SSMs) in revealing long-distance dependencies. To the best of our knowledge, this is the first Vision Mamba model designed for facial expression detection and classification. To evaluate the performance of the proposed FER-YOLO-Mamba model, we conducted experiments on two benchmark datasets, RAF-DB and SFEW. The experimental results indicate that the FER-YOLO-Mamba model achieved better results compared to other models. The code is available from //github.com/SwjtuMa/FER-YOLO-Mamba.
Uncertainty in LiDAR measurements, stemming from factors such as range sensing, is crucial for LIO (LiDAR-Inertial Odometry) systems as it affects the accurate weighting in the loss function. While recent LIO systems address uncertainty related to range sensing, the impact of incident angle on uncertainty is often overlooked by the community. Moreover, the existing uncertainty propagation methods suffer from computational inefficiency. This paper proposes a comprehensive point uncertainty model that accounts for both the uncertainties from LiDAR measurements and surface characteristics, along with an efficient local uncertainty analytical method for LiDAR-based state estimation problem. We employ a projection operator that separates the uncertainty into the ray direction and its orthogonal plane. Then, we derive incremental Jacobian matrices of eigenvalues and eigenvectors w.r.t. points, which enables a fast approximation of uncertainty propagation. This approach eliminates the requirement for redundant traversal of points, significantly reducing the time complexity of uncertainty propagation from $\mathcal{O} (n)$ to $\mathcal{O} (1)$ when a new point is added. Simulations and experiments on public datasets are conducted to validate the accuracy and efficiency of our formulations. The proposed methods have been integrated into a LIO system, which is available at //github.com/tiev-tongji/LOG-LIO2.
In the context of machine learning for graphs, many researchers have empirically observed that Deep Graph Networks (DGNs) perform favourably on node classification tasks when the graph structure is homophilic (\ie adjacent nodes are similar). In this paper, we introduce Lying-GCN, a new DGN inspired by opinion dynamics that can adaptively work in both the heterophilic and the homophilic setting. At each layer, each agent (node) shares its own opinions (node embeddings) with its neighbours. Instead of sharing its opinion directly as in GCN, we introduce a mechanism which allows agents to lie. Such a mechanism is adaptive, thus the agents learn how and when to lie according to the task that should be solved. We provide a characterisation of our proposal in terms of dynamical systems, by studying the spectral property of the coefficient matrix of the system. While the steady state of the system collapses to zero, we believe the lying mechanism is still usable to solve node classification tasks. We empirically prove our belief on both synthetic and real-world datasets, by showing that the lying mechanism allows to increase the performances in the heterophilic setting without harming the results in the homophilic one.
Knowledge Graph Embedding (KGE) aims to learn representations for entities and relations. Most KGE models have gained great success, especially on extrapolation scenarios. Specifically, given an unseen triple (h, r, t), a trained model can still correctly predict t from (h, r, ?), or h from (?, r, t), such extrapolation ability is impressive. However, most existing KGE works focus on the design of delicate triple modeling function, which mainly tells us how to measure the plausibility of observed triples, but offers limited explanation of why the methods can extrapolate to unseen data, and what are the important factors to help KGE extrapolate. Therefore in this work, we attempt to study the KGE extrapolation of two problems: 1. How does KGE extrapolate to unseen data? 2. How to design the KGE model with better extrapolation ability? For the problem 1, we first discuss the impact factors for extrapolation and from relation, entity and triple level respectively, propose three Semantic Evidences (SEs), which can be observed from train set and provide important semantic information for extrapolation. Then we verify the effectiveness of SEs through extensive experiments on several typical KGE methods. For the problem 2, to make better use of the three levels of SE, we propose a novel GNN-based KGE model, called Semantic Evidence aware Graph Neural Network (SE-GNN). In SE-GNN, each level of SE is modeled explicitly by the corresponding neighbor pattern, and merged sufficiently by the multi-layer aggregation, which contributes to obtaining more extrapolative knowledge representation. Finally, through extensive experiments on FB15k-237 and WN18RR datasets, we show that SE-GNN achieves state-of-the-art performance on Knowledge Graph Completion task and performs a better extrapolation ability.
Collecting supporting evidence from large corpora of text (e.g., Wikipedia) is of great challenge for open-domain Question Answering (QA). Especially, for multi-hop open-domain QA, scattered evidence pieces are required to be gathered together to support the answer extraction. In this paper, we propose a new retrieval target, hop, to collect the hidden reasoning evidence from Wikipedia for complex question answering. Specifically, the hop in this paper is defined as the combination of a hyperlink and the corresponding outbound link document. The hyperlink is encoded as the mention embedding which models the structured knowledge of how the outbound link entity is mentioned in the textual context, and the corresponding outbound link document is encoded as the document embedding representing the unstructured knowledge within it. Accordingly, we build HopRetriever which retrieves hops over Wikipedia to answer complex questions. Experiments on the HotpotQA dataset demonstrate that HopRetriever outperforms previously published evidence retrieval methods by large margins. Moreover, our approach also yields quantifiable interpretations of the evidence collection process.
小貼士
登錄享
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191