We consider the problem of accurate sparse fine-tuning of large language models (LLMs), that is, fine-tuning pretrained LLMs on specialized tasks, while inducing sparsity in their weights. On the accuracy side, we observe that standard loss-based fine-tuning may fail to recover accuracy, especially at high sparsities. To address this, we perform a detailed study of distillation-type losses, determining an L2-based distillation approach we term SquareHead which enables accurate recovery even at higher sparsities, across all model types. On the practical efficiency side, we show that sparse LLMs can be executed with speedups by taking advantage of sparsity, for both CPU and GPU runtimes. While the standard approach is to leverage sparsity for computational reduction, we observe that in the case of memory-bound LLMs sparsity can also be leveraged for reducing memory bandwidth. We exhibit end-to-end results showing speedups due to sparsity, while recovering accuracy, on T5 (language translation), Whisper (speech translation), and open GPT-type (MPT for text generation). For MPT text generation, we show for the first time that sparse fine-tuning can reach 75% sparsity without accuracy drops, provide notable end-to-end speedups for both CPU and GPU inference, and highlight that sparsity is also compatible with quantization approaches. Models and software for reproducing our results are provided in Section 6.
相關內容
Transformer-based models, such as BERT and GPT, have been widely adopted in natural language processing (NLP) due to their exceptional performance. However, recent studies show their vulnerability to textual adversarial attacks where the model's output can be misled by intentionally manipulating the text inputs. Despite various methods that have been proposed to enhance the model's robustness and mitigate this vulnerability, many require heavy consumption resources (e.g., adversarial training) or only provide limited protection (e.g., defensive dropout). In this paper, we propose a novel method called dynamic attention, tailored for the transformer architecture, to enhance the inherent robustness of the model itself against various adversarial attacks. Our method requires no downstream task knowledge and does not incur additional costs. The proposed dynamic attention consists of two modules: (I) attention rectification, which masks or weakens the attention value of the chosen tokens, and (ii) dynamic modeling, which dynamically builds the set of candidate tokens. Extensive experiments demonstrate that dynamic attention significantly mitigates the impact of adversarial attacks, improving up to 33\% better performance than previous methods against widely-used adversarial attacks. The model-level design of dynamic attention enables it to be easily combined with other defense methods (e.g., adversarial training) to further enhance the model's robustness. Furthermore, we demonstrate that dynamic attention preserves the state-of-the-art robustness space of the original model compared to other dynamic modeling methods.
The successful integration of large language models (LLMs) into recommendation systems has proven to be a major breakthrough in recent studies, paving the way for more generic and transferable recommendations. However, LLMs struggle to effectively utilize user and item IDs, which are crucial identifiers for successful recommendations. This is mainly due to their distinct representation in a semantic space that is different from the natural language (NL) typically used to train LLMs. To tackle such issue, we introduce ControlRec, an innovative Contrastive prompt learning framework for Recommendation systems. ControlRec treats user IDs and NL as heterogeneous features and encodes them individually. To promote greater alignment and integration between them in the semantic space, we have devised two auxiliary contrastive objectives: (1) Heterogeneous Feature Matching (HFM) aligning item description with the corresponding ID or user's next preferred ID based on their interaction sequence, and (2) Instruction Contrastive Learning (ICL) effectively merging these two crucial data sources by contrasting probability distributions of output sequences generated by diverse tasks. Experimental results on four public real-world datasets demonstrate the effectiveness of the proposed method on improving model performance.
We are interested in assessing the order of a finite-state Hidden Markov Model (HMM) with the only two assumptions that the transition matrix of the latent Markov chain has full rank and that the density functions of the emission distributions are linearly independent. We introduce a new procedure for estimating this order by investigating the rank of some well-chosen integral operator which relies on the distribution of a pair of consecutive observations. This method circumvents the usual limits of the spectral method when it is used for estimating the order of an HMM: it avoids the choice of the basis functions; it does not require any knowledge of an upper-bound on the order of the HMM (for the spectral method, such an upper-bound is defined by the number of basis functions); it permits to easily handle different types of data (including continuous data, circular data or multivariate continuous data) with a suitable choice of kernel. The method relies on the fact that the order of the HMM can be identified from the distribution of a pair of consecutive observations and that this order is equal to the rank of some integral operator (\emph{i.e.} the number of its singular values that are non-zero). Since only the empirical counter-part of the singular values of the operator can be obtained, we propose a data-driven thresholding procedure. An upper-bound on the probability of overestimating the order of the HMM is established. Moreover, sufficient conditions on the bandwidth used for kernel density estimation and on the threshold are stated to obtain the consistency of the estimator of the order of the HMM. The procedure is easily implemented since the values of all the tuning parameters are determined by the sample size.
Recent vision-language models have achieved tremendous progress far beyond what we ever expected. However, their computational costs are also dramatically growing with rapid development, especially for the large models. It makes model acceleration exceedingly critical in a scenario of limited resources. Although extensively studied for unimodal models, the acceleration for multimodal models, especially the vision-language Transformers, is relatively under-explored. To pursue more efficient and accessible vision-language Transformers, this paper introduces \textbf{Cross}-\textbf{G}uided \textbf{E}nsemble of \textbf{T}okens (\textbf{\emph{CrossGET}}), a universal acceleration framework for vision-language Transformers. This framework adaptively combines tokens through real-time, cross-modal guidance, thereby achieving substantial acceleration while keeping high performance. \textit{CrossGET} has two key innovations: 1) \textit{Cross-Guided Matching and Ensemble}. \textit{CrossGET} incorporates cross-modal guided token matching and ensemble to exploit cross-modal information effectively, only introducing cross-modal tokens with negligible extra parameters. 2) \textit{Complete-Graph Soft Matching}. In contrast to the existing bipartite soft matching approach, \textit{CrossGET} introduces a complete-graph soft matching policy to achieve more reliable token-matching results while maintaining parallelizability and high efficiency. Extensive experiments are conducted on various vision-language tasks, including image-text retrieval, visual reasoning, image captioning, and visual question answering. Performance on both classic multimodal architectures and emerging multimodal LLMs demonstrate the effectiveness and versatility of the proposed \textit{CrossGET} framework. The code will be at \url{//github.com/sdc17/CrossGET}.
The subrank of tensors is a measure of how much a tensor can be ''diagonalized''. This parameter was introduced by Strassen to study fast matrix multiplication algorithms in algebraic complexity theory and is closely related to many central tensor parameters (e.g. slice rank, partition rank, analytic rank, geometric rank, G-stable rank) and problems in combinatorics, computer science and quantum information theory. Strassen (J. Reine Angew. Math., 1988) proved that there is a gap in the subrank when taking large powers under the tensor product: either the subrank of all powers is at most one, or it grows as a power of a constant strictly larger than one. In this paper, we precisely determine this constant for tensors of any order. Additionally, for tensors of order three, we prove that there is a second gap in the possible rates of growth. Our results strengthen the recent work of Costa and Dalai (J. Comb. Theory, Ser. A, 2021), who proved a similar gap for the slice rank. Our theorem on the subrank has wider applications by implying such gaps not only for the slice rank, but for any ``normalized monotone''. In order to prove the main result, we characterize when a tensor has a very structured tensor (the W-tensor) in its orbit closure. Our methods include degenerations in Grassmanians, which may be of independent interest.
The advent of large language models marks a revolutionary breakthrough in artificial intelligence. With the unprecedented scale of training and model parameters, the capability of large language models has been dramatically improved, leading to human-like performances in understanding, language synthesizing, and common-sense reasoning, etc. Such a major leap-forward in general AI capacity will change the pattern of how personalization is conducted. For one thing, it will reform the way of interaction between humans and personalization systems. Instead of being a passive medium of information filtering, large language models present the foundation for active user engagement. On top of such a new foundation, user requests can be proactively explored, and user's required information can be delivered in a natural and explainable way. For another thing, it will also considerably expand the scope of personalization, making it grow from the sole function of collecting personalized information to the compound function of providing personalized services. By leveraging large language models as general-purpose interface, the personalization systems may compile user requests into plans, calls the functions of external tools to execute the plans, and integrate the tools' outputs to complete the end-to-end personalization tasks. Today, large language models are still being developed, whereas the application in personalization is largely unexplored. Therefore, we consider it to be the right time to review the challenges in personalization and the opportunities to address them with LLMs. In particular, we dedicate this perspective paper to the discussion of the following aspects: the development and challenges for the existing personalization system, the newly emerged capabilities of large language models, and the potential ways of making use of large language models for personalization.
Pre-trained Language Models (PLMs) which are trained on large text corpus via self-supervised learning method, have yielded promising performance on various tasks in Natural Language Processing (NLP). However, though PLMs with huge parameters can effectively possess rich knowledge learned from massive training text and benefit downstream tasks at the fine-tuning stage, they still have some limitations such as poor reasoning ability due to the lack of external knowledge. Research has been dedicated to incorporating knowledge into PLMs to tackle these issues. In this paper, we present a comprehensive review of Knowledge-Enhanced Pre-trained Language Models (KE-PLMs) to provide a clear insight into this thriving field. We introduce appropriate taxonomies respectively for Natural Language Understanding (NLU) and Natural Language Generation (NLG) to highlight these two main tasks of NLP. For NLU, we divide the types of knowledge into four categories: linguistic knowledge, text knowledge, knowledge graph (KG), and rule knowledge. The KE-PLMs for NLG are categorized into KG-based and retrieval-based methods. Finally, we point out some promising future directions of KE-PLMs.
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.
Deep convolutional neural networks (CNNs) have recently achieved great success in many visual recognition tasks. However, existing deep neural network models are computationally expensive and memory intensive, hindering their deployment in devices with low memory resources or in applications with strict latency requirements. Therefore, a natural thought is to perform model compression and acceleration in deep networks without significantly decreasing the model performance. During the past few years, tremendous progress has been made in this area. In this paper, we survey the recent advanced techniques for compacting and accelerating CNNs model developed. These techniques are roughly categorized into four schemes: parameter pruning and sharing, low-rank factorization, transferred/compact convolutional filters, and knowledge distillation. Methods of parameter pruning and sharing will be described at the beginning, after that the other techniques will be introduced. For each scheme, we provide insightful analysis regarding the performance, related applications, advantages, and drawbacks etc. Then we will go through a few very recent additional successful methods, for example, dynamic capacity networks and stochastic depths networks. After that, we survey the evaluation matrix, the main datasets used for evaluating the model performance and recent benchmarking efforts. Finally, we conclude this paper, discuss remaining challenges and possible directions on this topic.
Within the rapidly developing Internet of Things (IoT), numerous and diverse physical devices, Edge devices, Cloud infrastructure, and their quality of service requirements (QoS), need to be represented within a unified specification in order to enable rapid IoT application development, monitoring, and dynamic reconfiguration. But heterogeneities among different configuration knowledge representation models pose limitations for acquisition, discovery and curation of configuration knowledge for coordinated IoT applications. This paper proposes a unified data model to represent IoT resource configuration knowledge artifacts. It also proposes IoT-CANE (Context-Aware recommendatioN systEm) to facilitate incremental knowledge acquisition and declarative context driven knowledge recommendation.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191