The swift advancement in the scales and capabilities of Large Language Models (LLMs) positions them as promising tools for a variety of downstream tasks. In addition to the pursuit of better performance and the avoidance of violent feedback on a certain prompt, to ensure the responsibility of the LLM, much attention is drawn to the robustness of LLMs. However, existing evaluation methods mostly rely on traditional question answering datasets with predefined supervised labels, which do not align with the superior generation capabilities of contemporary LLMs. To address this issue, we propose a novel rational evaluation approach that leverages pre-trained reward models as diagnostic tools to evaluate the longer conversation generated from more challenging open questions by LLMs, which we refer to as the Reward Model for Reasonable Robustness Evaluation (TREvaL). Longer conversations manifest the comprehensive grasp of language models in terms of their proficiency in understanding questions, a capability not entirely encompassed by individual words or letters, which may exhibit oversimplification and inherent biases. Our extensive empirical experiments demonstrate that TREvaL provides an innovative method for evaluating the robustness of an LLM. Furthermore, our results demonstrate that LLMs frequently exhibit vulnerability to word-level perturbations that are commonplace in daily language usage. Notably, we are surprised to discover that robustness tends to decrease as fine-tuning (SFT and RLHF) is conducted. The code of TREval is available in //github.com/Harry-mic/TREvaL.
相關內容
Differential Privacy (DP) was originally developed to protect privacy. However, it has recently been utilized to secure machine learning (ML) models from poisoning attacks, with DP-SGD receiving substantial attention. Nevertheless, a thorough investigation is required to assess the effectiveness of different DP techniques in preventing backdoor attacks in practice. In this paper, we investigate the effectiveness of DP-SGD and, for the first time in literature, examine PATE in the context of backdoor attacks. We also explore the role of different components of DP algorithms in defending against backdoor attacks and will show that PATE is effective against these attacks due to the bagging structure of the teacher models it employs. Our experiments reveal that hyperparameters and the number of backdoors in the training dataset impact the success of DP algorithms. Additionally, we propose Label-DP as a faster and more accurate alternative to DP-SGD and PATE. We conclude that while Label-DP algorithms generally offer weaker privacy protection, accurate hyper-parameter tuning can make them more effective than DP methods in defending against backdoor attacks while maintaining model accuracy.
Modern multi-centre randomized controlled trials (MCRCTs) collect massive amounts of tabular data, and are monitored intensively for irregularities by humans. We began by empirically evaluating 6 modern machine learning-based outlier detection algorithms on the task of identifying irregular data in 838 datasets from 7 real-world MCRCTs with a total of 77,001 patients from over 44 countries. Our results reinforce key findings from prior work in the outlier detection literature on data from other domains. Existing algorithms often succeed at identifying irregularities without any supervision, with at least one algorithm exhibiting positive performance 70.6% of the time. However, performance across datasets varies substantially with no single algorithm performing consistently well, motivating new techniques for unsupervised model selection or other means of aggregating potentially discordant predictions from multiple candidate models. We propose the Meta-learned Probabilistic Ensemble (MePE), a simple algorithm for aggregating the predictions of multiple unsupervised models, and show that it performs favourably compared to recent meta-learning approaches for outlier detection model selection. While meta-learning shows promise, small ensembles outperform all forms of meta-learning on average, a negative result that may guide the application of current outlier detection approaches in healthcare and other real-world domains.
As a promising distributed learning paradigm, federated learning (FL) involves training deep neural network (DNN) models at the network edge while protecting the privacy of the edge clients. To train a large-scale DNN model, batch normalization (BN) has been regarded as a simple and effective means to accelerate the training and improve the generalization capability. However, recent findings indicate that BN can significantly impair the performance of FL in the presence of non-i.i.d. data. While several FL algorithms have been proposed to address this issue, their performance still falls significantly when compared to the centralized scheme. Furthermore, none of them have provided a theoretical explanation of how the BN damages the FL convergence. In this paper, we present the first convergence analysis to show that under the non-i.i.d. data, the mismatch between the local and global statistical parameters in BN causes the gradient deviation between the local and global models, which, as a result, slows down and biases the FL convergence. In view of this, we develop a new FL algorithm that is tailored to BN, called FedTAN, which is capable of achieving robust FL performance under a variety of data distributions via iterative layer-wise parameter aggregation. Comprehensive experimental results demonstrate the superiority of the proposed FedTAN over existing baselines for training BN-based DNN models.
Weak-memory models are standard formal specifications of concurrency across hardware, programming languages, and distributed systems. A fundamental computational problem is consistency testing: is the observed execution of a concurrent program in alignment with the specification of the underlying system? The problem has been studied extensively across Sequential Consistency (SC) and weak memory, and proven to be NP-complete when some aspect of the input (e.g., number of threads/memory locations) is unbounded. This unboundedness has left a natural question open: are there efficient parameterized algorithms for testing? The main contribution of this paper is a deep hardness result for consistency testing under many popular weak-memory models: the problem remains NP-complete even in its bounded setting, where candidate executions contain a bounded number of threads, memory locations, and values. This hardness spreads across several Release-Acquire variants of C11, a popular variant of its Relaxed fragment, popular Causal Consistency models, and the POWER architecture. To our knowledge, this is the first result that fully exposes the hardness of weak-memory testing and proves that the problem admits no parameterization under standard input parameters. It also yields a computational separation of these models from SC, x86-TSO, PSO, and Relaxed, for which bounded consistency testing is either known (for SC), or shown here (for the rest), to be in polynomial time.
The Streetlight Effect represents an observation bias that occurs when individuals search for something only where it is easiest to look. Despite the significant development of Post-Publication Peer Review (PPPR) in recent years, facilitated in part by platforms such as PubPeer, existing literature has not examined whether PPPR is affected by this type of bias. In other words, if the PPPR mainly concerns publications to which researchers have direct access (eg to analyze image duplications, etc.). In this study, we compare the Open Access (OA) structures of publishers and journals among 51,882 publications commented on PubPeer to those indexed in OpenAlex database (\#156,700,177). Our findings indicate that OA journals are 33% more prevalent in PubPeer than in the global total (52% for the most commented journals). This result can be attributed to disciplinary bias in PubPeer, with overrepresentation of medical and biological research (which exhibits higher levels of openness). However, after normalization, the results reveal that PPPR does not exhibit a Streetlight Effect, as OA publications, within the same discipline, are on average 16% less prevalent in PubPeer than in the global total. These results suggest that the process of scientific self-correction operates independently of publication access status.
Feature attribution methods are popular in interpretable machine learning. These methods compute the attribution of each input feature to represent its importance, but there is no consensus on the definition of "attribution", leading to many competing methods with little systematic evaluation, complicated in particular by the lack of ground truth attribution. To address this, we propose a dataset modification procedure to induce such ground truth. Using this procedure, we evaluate three common methods: saliency maps, rationales, and attentions. We identify several deficiencies and add new perspectives to the growing body of evidence questioning the correctness and reliability of these methods applied on datasets in the wild. We further discuss possible avenues for remedy and recommend new attribution methods to be tested against ground truth before deployment. The code is available at \url{//github.com/YilunZhou/feature-attribution-evaluation}.
Deep learning has become the dominant approach in coping with various tasks in Natural LanguageProcessing (NLP). Although text inputs are typically represented as a sequence of tokens, there isa rich variety of NLP problems that can be best expressed with a graph structure. As a result, thereis a surge of interests in developing new deep learning techniques on graphs for a large numberof NLP tasks. In this survey, we present a comprehensive overview onGraph Neural Networks(GNNs) for Natural Language Processing. We propose a new taxonomy of GNNs for NLP, whichsystematically organizes existing research of GNNs for NLP along three axes: graph construction,graph representation learning, and graph based encoder-decoder models. We further introducea large number of NLP applications that are exploiting the power of GNNs and summarize thecorresponding benchmark datasets, evaluation metrics, and open-source codes. Finally, we discussvarious outstanding challenges for making the full use of GNNs for NLP as well as future researchdirections. To the best of our knowledge, this is the first comprehensive overview of Graph NeuralNetworks for Natural Language Processing.
Knowledge graphs (KGs) serve as useful resources for various natural language processing applications. Previous KG completion approaches require a large number of training instances (i.e., head-tail entity pairs) for every relation. The real case is that for most of the relations, very few entity pairs are available. Existing work of one-shot learning limits method generalizability for few-shot scenarios and does not fully use the supervisory information; however, few-shot KG completion has not been well studied yet. In this work, we propose a novel few-shot relation learning model (FSRL) that aims at discovering facts of new relations with few-shot references. FSRL can effectively capture knowledge from heterogeneous graph structure, aggregate representations of few-shot references, and match similar entity pairs of reference set for every relation. Extensive experiments on two public datasets demonstrate that FSRL outperforms the state-of-the-art.
Language model pre-training has proven to be useful in learning universal language representations. As a state-of-the-art language model pre-training model, BERT (Bidirectional Encoder Representations from Transformers) has achieved amazing results in many language understanding tasks. In this paper, we conduct exhaustive experiments to investigate different fine-tuning methods of BERT on text classification task and provide a general solution for BERT fine-tuning. Finally, the proposed solution obtains new state-of-the-art results on eight widely-studied text classification datasets.
Graph Neural Networks (GNNs) for representation learning of graphs broadly follow a neighborhood aggregation framework, where the representation vector of a node is computed by recursively aggregating and transforming feature vectors of its neighboring nodes. Many GNN variants have been proposed and have achieved state-of-the-art results on both node and graph classification tasks. However, despite GNNs revolutionizing graph representation learning, there is limited understanding of their representational properties and limitations. Here, we present a theoretical framework for analyzing the expressive power of GNNs in capturing different graph structures. Our results characterize the discriminative power of popular GNN variants, such as Graph Convolutional Networks and GraphSAGE, and show that they cannot learn to distinguish certain simple graph structures. We then develop a simple architecture that is provably the most expressive among the class of GNNs and is as powerful as the Weisfeiler-Lehman graph isomorphism test. We empirically validate our theoretical findings on a number of graph classification benchmarks, and demonstrate that our model achieves state-of-the-art performance.
小貼士
登錄享
相關主題
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191