arXiv subjects

Fan Yin

Publications and source records attributed to Fan Yin.

At least 37 records · Page 2Linked to original sources

arXiv2023

Dynosaur: A Dynamic Growth Paradigm for Instruction-Tuning Data Curation

Instruction tuning has emerged to enhance the capabilities of large language models (LLMs) to comprehend instructions and generate appropriate responses. Existing methods either manually annotate or employ LLM (e.g., GPT-series) to generate data for instruction tuning. However, they often overlook associating instructions with existing annotated datasets. In this paper, we propose Dynosaur, a dynamic growth paradigm for the automatic curation of instruction-tuning data. Based on the metadata of existing datasets, we use LLMs to automatically construct instruction-tuning data by identifying relevant data fields and generating appropriate instructions. By leveraging the existing annotated datasets, Dynosaur offers several advantages: 1) it reduces the API cost for generating instructions (e.g., it costs less than $12 USD by calling GPT-3.5-turbo for generating 800K instruction tuning samples; 2) it provides high-quality data for instruction tuning (e.g., it performs better than Alpaca and Flan on Super-NI and Longform with comparable data sizes); and 3) it supports the continuous improvement of models by generating instruction-tuning data when a new annotated dataset becomes available. We further investigate a continual learning scheme for learning with the ever-growing instruction-tuning dataset, and demonstrate that replaying tasks with diverse instruction embeddings not only helps mitigate forgetting issues but generalizes to unseen tasks better. Code and data are available at https://github.com/WadeYin9712/Dynosaur.

cs.CL↗

arXiv2023

Red Teaming Language Model Detectors with Language Models

The prevalence and strong capability of large language models (LLMs) present significant safety and ethical risks if exploited by malicious users. To prevent the potentially deceptive usage of LLMs, recent works have proposed algorithms to detect LLM-generated text and protect LLMs. In this paper, we investigate the robustness and reliability of these LLM detectors under adversarial attacks. We study two types of attack strategies: 1) replacing certain words in an LLM's output with their synonyms given the context; 2) automatically searching for an instructional prompt to alter the writing style of the generation. In both strategies, we leverage an auxiliary LLM to generate the word replacements or the instructional prompt. Different from previous works, we consider a challenging setting where the auxiliary LLM can also be protected by a detector. Experiments reveal that our attacks effectively compromise the performance of all detectors in the study with plausible generations, underscoring the urgent need to improve the robustness of LLM-generated text detection systems.

cs.CL↗

arXiv2023

CleanCLIP: Mitigating Data Poisoning Attacks in Multimodal Contrastive Learning

Multimodal contrastive pretraining has been used to train multimodal representation models, such as CLIP, on large amounts of paired image-text data. However, previous studies have revealed that such models are vulnerable to backdoor attacks. Specifically, when trained on backdoored examples, CLIP learns spurious correlations between the embedded backdoor trigger and the target label, aligning their representations in the joint embedding space. Injecting even a small number of poisoned examples, such as 75 examples in 3 million pretraining data, can significantly manipulate the model's behavior, making it difficult to detect or unlearn such correlations. To address this issue, we propose CleanCLIP, a finetuning framework that weakens the learned spurious associations introduced by backdoor attacks by independently re-aligning the representations for individual modalities. We demonstrate that unsupervised finetuning using a combination of multimodal contrastive and unimodal self-supervised objectives for individual modalities can significantly reduce the impact of the backdoor attack. Additionally, we show that supervised finetuning on task-specific labeled image data removes the backdoor trigger from the CLIP vision encoder. We show empirically that CleanCLIP maintains model performance on benign examples while erasing a range of backdoor attacks on multimodal contrastive learning. The code and checkpoints are available at https://github.com/nishadsinghi/CleanCLIP.

cs.CV↗

arXiv2023

Did You Read the Instructions? Rethinking the Effectiveness of Task Definitions in Instruction Learning

Large language models (LLMs) have shown impressive performance in following natural language instructions to solve unseen tasks. However, it remains unclear whether models truly understand task definitions and whether the human-written definitions are optimal. In this paper, we systematically study the role of task definitions in instruction learning. We first conduct an ablation analysis informed by human annotations to understand which parts of a task definition are most important, and find that model performance only drops substantially when removing contents describing the task output, in particular label information. Next, we propose an automatic algorithm to compress task definitions to a minimal supporting set of tokens, and find that 60\% of tokens can be removed while maintaining or even improving model performance. Based on these results, we propose two strategies to help models better leverage task instructions: (1) providing only key information for tasks in a common structured format, and (2) adding a meta-tuning stage to help the model better understand the definitions. With these two strategies, we achieve a 4.2 Rouge-L improvement over 119 unseen test tasks.

cs.CL↗

arXiv2023

Efficient Shapley Values Estimation by Amortization for Text Classification

Despite the popularity of Shapley Values in explaining neural text classification models, computing them is prohibitive for large pretrained models due to a large number of model evaluations. In practice, Shapley Values are often estimated with a small number of stochastic model evaluations. However, we show that the estimated Shapley Values are sensitive to random seed choices -- the top-ranked features often have little overlap across different seeds, especially on examples with longer input texts. This can only be mitigated by aggregating thousands of model evaluations, which on the other hand, induces substantial computational overheads. To mitigate the trade-off between stability and efficiency, we develop an amortized model that directly predicts each input feature's Shapley Value without additional model evaluations. It is trained on a set of examples whose Shapley Values are estimated from a large number of model evaluations to ensure stability. Experimental results on two text classification datasets demonstrate that our amortized model estimates Shapley Values accurately with up to 60 times speedup compared to traditional methods. Furthermore, the estimated values are stable as the inference is deterministic. We release our code at https://github.com/yangalan123/Amortized-Interpretability.

cs.CL↗

arXiv2022

A room-temperature electrical-field-enhanced ultrafast switch in organic microcavity polariton condensates

Integrated electro-optical switches are essential as one of the fundamental elements in the development of modern optoelectronics. As an architecture for photonic systems, exciton polaritons, that are hybrid bosonic quasiparticles that possess unique properties derived from both excitons and photons, have shown much promise. For this system, we demonstrate a significant improvement of emitted intensity and condensation threshold by applying an electric field to a microcavity filled with an organic microbelt. Our theoretical investigations indicate that the electric field makes the excitons dipolar and induces an enhancement of the exciton-polariton interaction and of the polariton lifetime. Based on these electric field induced changes, a sub-nanosecond electrical-field-enhanced polariton condensate switch is realized at room temperature, providing the basis for developing an on-chip integrated photonic device in the strong light-matter coupling regime.

physics.optics↗

arXiv2022

ADDMU: Detection of Far-Boundary Adversarial Examples with Data and Model Uncertainty Estimation

Adversarial Examples Detection (AED) is a crucial defense technique against adversarial attacks and has drawn increasing attention from the Natural Language Processing (NLP) community. Despite the surge of new AED methods, our studies show that existing methods heavily rely on a shortcut to achieve good performance. In other words, current search-based adversarial attacks in NLP stop once model predictions change, and thus most adversarial examples generated by those attacks are located near model decision boundaries. To surpass this shortcut and fairly evaluate AED methods, we propose to test AED methods with \textbf{F}ar \textbf{B}oundary (\textbf{FB}) adversarial examples. Existing methods show worse than random guess performance under this scenario. To overcome this limitation, we propose a new technique, \textbf{ADDMU}, \textbf{a}dversary \textbf{d}etection with \textbf{d}ata and \textbf{m}odel \textbf{u}ncertainty, which combines two types of uncertainty estimation for both regular and FB adversarial example detection. Our new method outperforms previous methods by 3.6 and 6.0 \emph{AUC} points under each scenario. Finally, our analysis shows that the two types of uncertainty provided by \textbf{ADDMU} can be leveraged to characterize adversarial examples and identify the ones that contribute most to model's robustness in adversarial training.

cs.CL↗

arXiv2022

On the Sensitivity and Stability of Model Interpretations in NLP

Recent years have witnessed the emergence of a variety of post-hoc interpretations that aim to uncover how natural language processing (NLP) models make predictions. Despite the surge of new interpretation methods, it remains an open problem how to define and quantitatively measure the faithfulness of interpretations, i.e., to what extent interpretations reflect the reasoning process by a model. We propose two new criteria, sensitivity and stability, that provide complementary notions of faithfulness to the existed removal-based criteria. Our results show that the conclusion for how faithful interpretations are could vary substantially based on different notions. Motivated by the desiderata of sensitivity and stability, we introduce a new class of interpretation methods that adopt techniques from adversarial robustness. Empirical results show that our proposed methods are effective under the new criteria and overcome limitations of gradient-based methods on removal-based criteria. Besides text classification, we also apply interpretation methods and metrics to dependency parsing. Our results shed light on understanding the diverse set of interpretations.

cs.CL↗

arXiv2021

Highly Scalable Maximum Likelihood and Conjugate Bayesian Inference for ERGMs on Graph Sets with Equivalent Vertices

The exponential family random graph modeling (ERGM) framework provides a flexible approach for the statistical analysis of networks. As ERGMs typically involve normalizing factors that are costly to compute, practical inference relies on a variety of approximations or other workarounds. Markov Chain Monte Carlo maximum likelihood (MCMC MLE) provides a powerful tool to approximate the MLE of ERGM parameters, and is feasible for typical models on single networks with as many as a few thousand nodes. MCMC-based algorithms for Bayesian analysis are more expensive, and high-quality answers are challenging to obtain on large graphs. For both strategies, extension to the pooled case - in which we observe multiple networks from a common generative process - adds further computational cost, with both time and memory scaling linearly in the number of graphs. This becomes prohibitive for large networks, or where large numbers of graph observations are available. Here, we exploit some basic properties of the discrete exponential families to develop an approach for ERGM inference in the pooled case that (where applicable) allows an arbitrarily large number of graph observations to be fit at no additional computational cost beyond preprocessing the data itself. Moreover, a variant of our approach can also be used to perform Bayesian inference under conjugate priors, again with no additional computational cost in the estimation phase. As we show, the conjugate prior is easily specified, and is well-suited to applications such as regularization. Simulation studies show that the pooled method leads to estimates with good frequentist properties, and posterior estimates under the conjugate prior are well-behaved. We demonstrate our approach with applications to pooled analysis of brain functional connectivity networks and to replicated x-ray crystal structures of hen egg-white lysozyme.

stat.ME↗

arXiv2021

Achieve Efficient Position-Heap-based Privacy-Preserving Substring-of-Keyword Query over Cloud

The cloud computing technique, which was initially used to mitigate the explosive growth of data, has been required to take both data privacy and users' query functionality into consideration. Symmetric searchable encryption (SSE) is a popular solution to supporting efficient keyword queries over encrypted data in the cloud. However, most of the existing SSE schemes focus on the exact keyword query and cannot work well when the user only remembers the substring of a keyword, i.e., substring-of-keyword query. This paper aims to investigate this issue by proposing an efficient and privacy-preserving substring-of-keyword query scheme over cloud. First, we employ the position heap technique to design a novel tree-based index to match substrings with corresponding keywords. Based on the tree-based index, we introduce our substring-of-keyword query scheme, which contains two consecutive phases. The first phase queries the keywords that match a given substring, and the second phase queries the files that match a keyword in which people are really interested. In addition, detailed security analysis and experimental results demonstrate the security and efficiency of our proposed scheme.

cs.CR↗

arXiv2020

Bayesian Nonparametric Estimation for Point Processes with Spatial Homogeneity: A Spatial Analysis of NBA Shot Locations

Basketball shot location data provide valuable summary information regarding players to coaches, sports analysts, fans, statisticians, as well as players themselves. Represented by spatial points, such data are naturally analyzed with spatial point process models. We present a novel nonparametric Bayesian method for learning the underlying intensity surface built upon a combination of Dirichlet process and Markov random field. Our method has the advantage of effectively encouraging local spatial homogeneity when estimating a globally heterogeneous intensity surface. Posterior inferences are performed with an efficient Markov chain Monte Carlo (MCMC) algorithm. Simulation studies show that the inferences are accurate and that the method is superior compared to the competing methods. Application to the shot location data of $20$ representative NBA players in the 2017-2018 regular season offers interesting insights about the shooting patterns of these players. A comparison against the competing method shows that the proposed method can effectively incorporate spatial contiguity into the estimation of intensity surfaces.

stat.ME↗

arXiv2020

Bayesian Analysis of Static Light Scattering Data for Globular Proteins

Static light scattering is a popular physical chemistry technique that enables calculation of physical attributes such as the radius of gyration and the second virial coefficient for a macromolecule (e.g., a polymer or a protein) in solution. The second virial coefficient is a physical quantity that characterizes the magnitude and sign of pairwise interactions between particles, and hence is related to aggregation propensity, a property of considerable scientific and practical interest. Estimating the second virial coefficient from experimental data is challenging due both to the degree of precision required and the complexity of the error structure involved. In contrast to conventional approaches based on heuristic OLS estimates, Bayesian inference for the second virial coefficient allows explicit modeling of error processes, incorporation of prior information, and the ability to directly test competing physical models. Here, we introduce a fully Bayesian model for static light scattering experiments on small-particle systems, with joint inference for concentration, index of refraction, oligomer size, and the second virial coefficient. We apply our proposed model to study the aggregation behavior of hen egg-white lysozyme and human gammaS-crystallin using in-house experimental data. Based on these observations, we also perform a simulation study on the primary drivers of uncertainty in this family of experiments, showing in particular the potential for improved monitoring and control of concentration to aid inference.

stat.AP↗

arXiv2020

Analysis of professional basketball field goal attempts via a Bayesian matrix clustering approach

We propose a Bayesian nonparametric matrix clustering approach to analyze the latent heterogeneity structure in the shot selection data collected from professional basketball players in the National Basketball Association (NBA). The proposed method adopts a mixture of finite mixtures framework and fully utilizes the spatial information via a mixture of matrix normal distribution representation. We propose an efficient Markov chain Monte Carlo algorithm for posterior sampling that allows simultaneous inference on both the number of clusters and the cluster configurations. We also establish large sample convergence properties for the posterior distribution. The excellent empirical performance of the proposed method is demonstrated via simulation studies and an application to shot chart data from selected players in the 2017 18 NBA regular season.

stat.ME↗

arXiv2020

Kernel-based Approximate Bayesian Inference for Exponential Family Random Graph Models

Bayesian inference for exponential family random graph models (ERGMs) is a doubly-intractable problem because of the intractability of both the likelihood and posterior normalizing factor. Auxiliary variable based Markov Chain Monte Carlo (MCMC) methods for this problem are asymptotically exact but computationally demanding, and are difficult to extend to modified ERGM families. In this work, we propose a kernel-based approximate Bayesian computation algorithm for fitting ERGMs. By employing an adaptive importance sampling technique, we greatly improve the efficiency of the sampling step. Though approximate, our easily parallelizable approach is yields comparable accuracy to state-of-the-art methods with substantial improvements in compute time on multi-core hardware. Our approach also flexibly accommodates both algorithmic enhancements (including improved learning algorithms for estimating conditional expectations) and extensions to non-standard cases such as inference from non-sufficient statistics. We demonstrate the performance of this approach on two well-known network data sets, comparing its accuracy and efficiency with results obtained using the approximate exchange algorithm. Our tests show a wallclock time advantage of up to 50% with five cores, and the ability to fit models in 1/5th the time at 30 cores; further speed enhancements are possible when more cores are available.

stat.CO↗

arXiv2020

Spatial Heterogeneity Can Lead to Substantial Local Variations in COVID-19 Timing and Severity

Standard epidemiological models for COVID-19 employ variants of compartment (SIR) models at local scales, implicitly assuming spatially uniform local mixing. Here, we examine the effect of employing more geographically detailed diffusion models based on known spatial features of interpersonal networks, most particularly the presence of a long-tailed but monotone decline in the probability of interaction with distance, on disease diffusion. Based on simulations of unrestricted COVID-19 diffusion in 19 U.S cities, we conclude that heterogeneity in population distribution can have large impacts on local pandemic timing and severity, even when aggregate behavior at larger scales mirrors a classic SIR-like pattern. Impacts observed include severe local outbreaks with long lag time relative to the aggregate infection curve, and the presence of numerous areas whose disease trajectories correlate poorly with those of neighboring areas. A simple catchment model for hospital demand illustrates potential implications for health care utilization, with substantial disparities in the timing and extremity of impacts even without distancing interventions. Likewise, analysis of social exposure to others who are morbid or deceased shows considerable variation in how the epidemic can appear to individuals on the ground, potentially affecting risk assessment and compliance with mitigation measures. These results demonstrate the potential for spatial network structure to generate highly non-uniform diffusion behavior even at the scale of cities, and suggest the importance of incorporating such structure when designing models to inform healthcare planning, predict community outcomes, or identify potential disparities.

physics.soc-ph↗

arXiv2020

On the Robustness of Language Encoders against Grammatical Errors

We conduct a thorough study to diagnose the behaviors of pre-trained language encoders (ELMo, BERT, and RoBERTa) when confronted with natural grammatical errors. Specifically, we collect real grammatical errors from non-native speakers and conduct adversarial attacks to simulate these errors on clean text data. We use this approach to facilitate debugging models on downstream applications. Results confirm that the performance of all tested models is affected but the degree of impact varies. To interpret model behaviors, we further design a linguistic acceptability task to reveal their abilities in identifying ungrammatical sentences and the position of errors. We find that fixed contextual encoders with a simple classifier trained on the prediction of sentence correctness are able to locate error positions. We also design a cloze test for BERT and discover that BERT captures the interaction between errors and specific tokens in context. Our results shed light on understanding the robustness and behaviors of language encoders against grammatical errors.

cs.CL↗

arXiv2020

Finite Mixtures of ERGMs for Modeling Ensembles of Networks

Ensembles of networks arise in many scientific fields, but there are few statistical tools for inferring their generative processes, particularly in the presence of both dyadic dependence and cross-graph heterogeneity. To fill in this gap, we propose characterizing network ensembles via finite mixtures of exponential family random graph models, a framework for parametric statistical modeling of graphs that has been successful in explicitly modeling the complex stochastic processes that govern the structure of edges in a network. Our proposed modeling framework can also be used for applications such as model-based clustering of ensembles of networks and density estimation for complex graph distributions. We develop a Metropolis-within-Gibbs algorithm to conduct fully Bayesian inference and adapt a version of deviance information criterion for missing data models to choose the number of latent heterogeneous generative mechanisms. Simulation studies show that the proposed procedure can recover the true number of latent heterogeneous generative processes and corresponding parameters. We demonstrate the utility of the proposed approach using an ensemble of political co-voting networks among U.S. Senators.

stat.ME↗

arXiv2019

Selection of Exponential-Family Random Graph Models via Held-Out Predictive Evaluation (HOPE)

Statistical models for networks with complex dependencies pose particular challenges for model selection and evaluation. In particular, many well-established statistical tools for selecting between models assume conditional independence of observations and/or conventional asymptotics, and their theoretical foundations are not always applicable in a network modeling context. While simulation-based approaches to model adequacy assessment are now widely used, there remains a need for procedures that quantify a model's performance in a manner suitable for selecting among competing models. Here, we propose to address this issue by developing a predictive evaluation strategy for exponential family random graph models that is analogous to cross-validation. Our approach builds on the held-out predictive evaluation (HOPE) scheme introduced by Wang et al. (2016) to assess imputation performance. We systematically hold out parts of the observed network to: evaluate how well the model is able to predict the held-out data; identify where the model performs poorly based on which data are held-out, indicating e.g. potential weaknesses; and calculate general summaries of predictive performance that can be used for model selection. As such, HOPE can assist researchers in improving models by indicating where a model performs poorly, and by quantitatively comparing predictive performance across competing models. The proposed method is applied to model selection problem of two well-known data sets, and the results are compared to those obtained via nominal AIC and BIC scores.

stat.ME↗