Quantization scale and bit-width are the most important parameters when considering how to quantize a neural network. Prior work focuses on optimizing quantization scales in a global manner through gradient methods (gradient descent \& Hessian analysis). Yet, when applying perturbations to quantization scales, we observe a very jagged, highly non-smooth test loss landscape. In fact, small perturbations in quantization scale can greatly affect accuracy, yielding a $0.5-0.8\%$ accuracy boost in 4-bit quantized vision transformers (ViTs). In this regime, gradient methods break down, since they cannot reliably reach local minima. In our work, dubbed Evol-Q, we use evolutionary search to effectively traverse the non-smooth landscape. Additionally, we propose using an infoNCE loss, which not only helps combat overfitting on the small calibration dataset ($1,000$ images) but also makes traversing such a highly non-smooth surface easier. Evol-Q improves the top-1 accuracy of a fully quantized ViT-Base by $10.30\%$, $0.78\%$, and $0.15\%$ for $3$-bit, $4$-bit, and $8$-bit weight quantization levels. Extensive experiments on a variety of CNN and ViT architectures further demonstrate its robustness in extreme quantization scenarios. Our code is available at //github.com/enyac-group/evol-q
相關內容
The Euler characteristic transform (ECT) is an integral transform used widely in topological data analysis. Previous efforts by Curry et al. and Ghrist et al. have independently shown that the ECT is injective on all compact definable sets. In this work, we study the invertibility of the ECT on definable sets that aren't necessarily compact, resulting in a complete classification of constructible functions that the Euler characteristic transform is not injective on. We then introduce the quadric Euler characteristic transform (QECT) as a natural generalization of the ECT by detecting definable shapes with quadric hypersurfaces rather than hyperplanes. We also discuss some criteria for the invertibility of QECT.
The attention module is the key component in Transformers. While the global attention mechanism offers high expressiveness, its excessive computational cost restricts its applicability in various scenarios. In this paper, we propose a novel attention paradigm, Agent Attention, to strike a favorable balance between computational efficiency and representation power. Specifically, the Agent Attention, denoted as a quadruple $(Q, A, K, V)$, introduces an additional set of agent tokens $A$ into the conventional attention module. The agent tokens first act as the agent for the query tokens $Q$ to aggregate information from $K$ and $V$, and then broadcast the information back to $Q$. Given the number of agent tokens can be designed to be much smaller than the number of query tokens, the agent attention is significantly more efficient than the widely adopted Softmax attention, while preserving global context modelling capability. Interestingly, we show that the proposed agent attention is equivalent to a generalized form of linear attention. Therefore, agent attention seamlessly integrates the powerful Softmax attention and the highly efficient linear attention. Extensive experiments demonstrate the effectiveness of agent attention with various vision Transformers and across diverse vision tasks, including image classification, object detection, semantic segmentation and image generation. Notably, agent attention has shown remarkable performance in high-resolution scenarios, owning to its linear attention nature. For instance, when applied to Stable Diffusion, our agent attention accelerates generation and substantially enhances image generation quality without any additional training. Code is available at //github.com/LeapLabTHU/Agent-Attention.
Macro-level modeling is still the dominant approach in many demographic applications because of its simplicity. Individual-level models, on the other hand, provide a more comprehensive understanding of observed patterns; however, their estimation using real data has remained a challenge. The approach we introduce in this article attempts to overcome this limitation. Using likelihood-free inference techniques, we show that it is possible to estimate the parameters of a simple but demographically interpretable individual-level model of the reproductive process from a set of aggregate fertility rates. By estimating individual-level quantities from widely available aggregate data, this approach can contribute to a better understanding of reproductive behavior and its driving mechanisms. It also allows for a more direct link between individual-level and population-level processes. We illustrate our approach using data from three natural fertility populations.
Over the last decade, applications of neural networks have spread to cover all aspects of life. A large number of companies base their businesses on building products that use neural networks for tasks such as face recognition, machine translation, and autonomous cars. They are being used in safety and security-critical applications like high definition maps and medical wristbands, or in globally used products like Google Translate and ChatGPT. Much of the intellectual property underpinning these products is encoded in the exact configuration of the neural networks. Consequently, protecting these is of utmost priority to businesses. At the same time, many of these products need to operate under a strong threat model, in which the adversary has unfettered physical control of the product. Past work has demonstrated that with physical access, attackers can reverse engineer neural networks that run on scalar microcontrollers, like ARM Cortex M3. However, for performance reasons, neural networks are often implemented on highly-parallel general purpose graphics processing units (GPGPUs), and so far, attacks on these have only recovered course-grained information on the structure of the neural network, but failed to retrieve the weights and biases. In this work, we present BarraCUDA, a novel attack on GPGPUs that can completely extract the parameters of neural networks. BarraCUDA uses correlation electromagnetic analysis to recover the weights and biases in the convolutional layers of neural networks. We use BarraCUDA to attack the popular NVIDIA Jetson Nano device, demonstrating successful parameter extraction of neural networks in a highly parallel and noisy environment.
By interacting, synchronizing, and cooperating with its physical counterpart in real time, digital twin is promised to promote an intelligent, predictive, and optimized modern city. Via interconnecting massive physical entities and their virtual twins with inter-twin and intra-twin communications, the Internet of digital twins (IoDT) enables free data exchange, dynamic mission cooperation, and efficient information aggregation for composite insights across vast physical/virtual entities. However, as IoDT incorporates various cutting-edge technologies to spawn the new ecology, severe known/unknown security flaws and privacy invasions of IoDT hinders its wide deployment. Besides, the intrinsic characteristics of IoDT such as \emph{decentralized structure}, \emph{information-centric routing} and \emph{semantic communications} entail critical challenges for security service provisioning in IoDT. To this end, this paper presents an in-depth review of the IoDT with respect to system architecture, enabling technologies, and security/privacy issues. Specifically, we first explore a novel distributed IoDT architecture with cyber-physical interactions and discuss its key characteristics and communication modes. Afterward, we investigate the taxonomy of security and privacy threats in IoDT, discuss the key research challenges, and review the state-of-the-art defense approaches. Finally, we point out the new trends and open research directions related to IoDT.
Graph neural networks (GNNs) have demonstrated a significant boost in prediction performance on graph data. At the same time, the predictions made by these models are often hard to interpret. In that regard, many efforts have been made to explain the prediction mechanisms of these models from perspectives such as GNNExplainer, XGNN and PGExplainer. Although such works present systematic frameworks to interpret GNNs, a holistic review for explainable GNNs is unavailable. In this survey, we present a comprehensive review of explainability techniques developed for GNNs. We focus on explainable graph neural networks and categorize them based on the use of explainable methods. We further provide the common performance metrics for GNNs explanations and point out several future research directions.
Graph Neural Networks (GNNs) have been studied from the lens of expressive power and generalization. However, their optimization properties are less well understood. We take the first step towards analyzing GNN training by studying the gradient dynamics of GNNs. First, we analyze linearized GNNs and prove that despite the non-convexity of training, convergence to a global minimum at a linear rate is guaranteed under mild assumptions that we validate on real-world graphs. Second, we study what may affect the GNNs' training speed. Our results show that the training of GNNs is implicitly accelerated by skip connections, more depth, and/or a good label distribution. Empirical results confirm that our theoretical results for linearized GNNs align with the training behavior of nonlinear GNNs. Our results provide the first theoretical support for the success of GNNs with skip connections in terms of optimization, and suggest that deep GNNs with skip connections would be promising in practice.
Since real-world objects and their interactions are often multi-modal and multi-typed, heterogeneous networks have been widely used as a more powerful, realistic, and generic superclass of traditional homogeneous networks (graphs). Meanwhile, representation learning (\aka~embedding) has recently been intensively studied and shown effective for various network mining and analytical tasks. In this work, we aim to provide a unified framework to deeply summarize and evaluate existing research on heterogeneous network embedding (HNE), which includes but goes beyond a normal survey. Since there has already been a broad body of HNE algorithms, as the first contribution of this work, we provide a generic paradigm for the systematic categorization and analysis over the merits of various existing HNE algorithms. Moreover, existing HNE algorithms, though mostly claimed generic, are often evaluated on different datasets. Understandable due to the application favor of HNE, such indirect comparisons largely hinder the proper attribution of improved task performance towards effective data preprocessing and novel technical design, especially considering the various ways possible to construct a heterogeneous network from real-world application data. Therefore, as the second contribution, we create four benchmark datasets with various properties regarding scale, structure, attribute/label availability, and \etc.~from different sources, towards handy and fair evaluations of HNE algorithms. As the third contribution, we carefully refactor and amend the implementations and create friendly interfaces for 13 popular HNE algorithms, and provide all-around comparisons among them over multiple tasks and experimental settings.
Deep neural networks (DNNs) are successful in many computer vision tasks. However, the most accurate DNNs require millions of parameters and operations, making them energy, computation and memory intensive. This impedes the deployment of large DNNs in low-power devices with limited compute resources. Recent research improves DNN models by reducing the memory requirement, energy consumption, and number of operations without significantly decreasing the accuracy. This paper surveys the progress of low-power deep learning and computer vision, specifically in regards to inference, and discusses the methods for compacting and accelerating DNN models. The techniques can be divided into four major categories: (1) parameter quantization and pruning, (2) compressed convolutional filters and matrix factorization, (3) network architecture search, and (4) knowledge distillation. We analyze the accuracy, advantages, disadvantages, and potential solutions to the problems with the techniques in each category. We also discuss new evaluation metrics as a guideline for future research.
Distant supervision can effectively label data for relation extraction, but suffers from the noise labeling problem. Recent works mainly perform soft bag-level noise reduction strategies to find the relatively better samples in a sentence bag, which is suboptimal compared with making a hard decision of false positive samples in sentence level. In this paper, we introduce an adversarial learning framework, which we named DSGAN, to learn a sentence-level true-positive generator. Inspired by Generative Adversarial Networks, we regard the positive samples generated by the generator as the negative samples to train the discriminator. The optimal generator is obtained until the discrimination ability of the discriminator has the greatest decline. We adopt the generator to filter distant supervision training dataset and redistribute the false positive instances into the negative set, in which way to provide a cleaned dataset for relation classification. The experimental results show that the proposed strategy significantly improves the performance of distant supervision relation extraction comparing to state-of-the-art systems.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191