Wei Du

Guo Yu, Lianbo Ma, Wei Du et al.

Recent years have seen the rapid development of fairness-aware machine learning in mitigating unfairness or discrimination in decision-making in a wide range of applications. However, much less attention has been paid to the fairness-aware multi-objective optimization, which is indeed commonly seen in real life, such as fair resource allocation problems and data driven multi-objective optimization problems. This paper aims to illuminate and broaden our understanding of multi-objective optimization from the perspective of fairness. To this end, we start with a discussion of user preferences in multi-objective optimization and then explore its relationship to fairness in machine learning and multi-objective optimization. Following the above discussions, representative cases of fairness-aware multiobjective optimization are presented, further elaborating the importance of fairness in traditional multi-objective optimization, data-driven optimization and federated optimization. Finally, challenges and opportunities in fairness-aware multi-objective optimization are addressed. We hope that this article makes a small step forward towards understanding fairness in the context of optimization and promote research interest in fairness-aware multi-objective optimization.

Haodong Zhao, Wei Du, Fangqi Li et al.

Federated learning (FL) has enabled global model training on decentralized data in a privacy-preserving way by aggregating model updates. However, for many natural language processing (NLP) tasks that utilize pre-trained language models (PLMs) with large numbers of parameters, there are considerable communication costs associated with FL. Recently, prompt tuning, which tunes some soft prompts without modifying PLMs, has achieved excellent performance as a new learning paradigm. Therefore we want to combine the two methods and explore the effect of prompt tuning under FL. In this paper, we propose "FedPrompt" to study prompt tuning in a model split aggregation way using FL, and prove that split aggregation greatly reduces the communication cost, only 0.01% of the PLMs' parameters, with little decrease on accuracy both on IID and Non-IID data distribution. This improves the efficiency of FL method while also protecting the data privacy in prompt tuning. In addition, like PLMs, prompts are uploaded and downloaded between public platforms and personal users, so we try to figure out whether there is still a backdoor threat using only soft prompts in FL scenarios. We further conduct backdoor attacks by data poisoning on FedPrompt. Our experiments show that normal backdoor attack can not achieve a high attack success rate, proving the robustness of FedPrompt. We hope this work can promote the application of prompt in FL and raise the awareness of the possible security threats.

Daniel Campos, Daniel Perry, Samir Joshi et al. · amazon-science

Experience management is an emerging business area where organizations focus on understanding the feedback of customers and employees in order to improve their end-to-end experiences. This results in a unique set of machine learning problems to help understand how people feel, discover issues they care about, and find which actions need to be taken on data that are different in content and distribution from traditional NLP domains. In this paper, we present a case study of building text analysis applications that perform multiple classification tasks efficiently in 12 languages in the nascent business area of experience management. In order to scale up modern ML methods on experience data, we leverage cross lingual and multi-task modeling techniques to consolidate our models into a single deployment to avoid overhead. We also make use of model compression and model distillation to reduce overall inference latency and hardware cost to the level acceptable for business needs while maintaining model prediction quality. Our findings show that multi-task modeling improves task performance for a subset of experience management tasks in both XLM-R and mBert architectures. Among the compressed architectures we explored, we found that MiniLM achieved the best compression/performance tradeoff. Our case study demonstrates a speedup of up to 15.61x with 2.60% average task degradation (or 3.29x speedup with 1.71% degradation) and estimated savings of 44% over using the original full-size model. These results demonstrate a successful scaling up of text classification for the challenging new area of ML for experience management.

Wei Du, Laksh Advani, Yashmeet Gambhir et al. · amazon-science

Large language models (LLMs) have demonstrated significant capability to generalize across a large number of NLP tasks. For industry applications, it is imperative to assess the performance of the LLM on unlabeled production data from time to time to validate for a real-world setting. Human labeling to assess model error requires considerable expense and time delay. Here we demonstrate that ensemble disagreement scores work well as a proxy for human labeling for language models in zero-shot, few-shot, and fine-tuned settings, per our evaluation on keyphrase extraction (KPE) task. We measure fidelity of the results by comparing to true error measured from human labeled ground truth. We contrast with the alternative of using another LLM as a source of machine labels, or silver labels. Results across various languages and domains show disagreement scores provide a better estimation of model performance with mean average error (MAE) as low as 0.4% and on average 13.8% better than using silver labels.

Tianjie Ju, Weiwei Sun, Wei Du et al.

Previous work has showcased the intriguing capability of large language models (LLMs) in retrieving facts and processing context knowledge. However, only limited research exists on the layer-wise capability of LLMs to encode knowledge, which challenges our understanding of their internal mechanisms. In this paper, we devote the first attempt to investigate the layer-wise capability of LLMs through probing tasks. We leverage the powerful generative capability of ChatGPT to construct probing datasets, providing diverse and coherent evidence corresponding to various facts. We employ $\mathcal V$-usable information as the validation metric to better reflect the capability in encoding context knowledge across different layers. Our experiments on conflicting and newly acquired knowledge show that LLMs: (1) prefer to encode more context knowledge in the upper layers; (2) primarily encode context knowledge within knowledge-related entity tokens at lower layers while progressively expanding more knowledge within other tokens at upper layers; and (3) gradually forget the earlier context knowledge retained within the intermediate layers when provided with irrelevant evidence. Code is publicly available at https://github.com/Jometeorie/probing_llama.

Wei Du, Peixuan Li, Boqun Li et al.

Backdoors implanted in pre-trained language models (PLMs) can be transferred to various downstream tasks, which exposes a severe security threat. However, most existing backdoor attacks against PLMs are un-targeted and task-specific. Few targeted and task-agnostic methods use manually pre-defined triggers and output representations, which prevent the attacks from being more effective and general. In this paper, we first summarize the requirements that a more threatening backdoor attack against PLMs should satisfy, and then propose a new backdoor attack method called UOR, which breaks the bottleneck of the previous approach by turning manual selection into automatic optimization. Specifically, we define poisoned supervised contrastive learning which can automatically learn the more uniform and universal output representations of triggers for various PLMs. Moreover, we use gradient search to select appropriate trigger words which can be adaptive to different PLMs and vocabularies. Experiments show that our method can achieve better attack performance on various text classification tasks compared to manual methods. Further, we tested our method on PLMs with different architectures, different usage paradigms, and more difficult tasks, which demonstrated the universality of our method.

Minh-Hao Van, Wei Du, Xintao Wu et al.

Both fair machine learning and adversarial learning have been extensively studied. However, attacking fair machine learning models has received less attention. In this paper, we present a framework that seeks to effectively generate poisoning samples to attack both model accuracy and algorithmic fairness. Our attacking framework can target fair machine learning models trained with a variety of group based fairness notions such as demographic parity and equalized odds. We develop three online attacks, adversarial sampling , adversarial labeling, and adversarial feature modification. All three attacks effectively and efficiently produce poisoning samples via sampling, labeling, or modifying a fraction of training data in order to reduce the test accuracy. Our framework enables attackers to flexibly adjust the attack's focus on prediction accuracy or fairness and accurately quantify the impact of each candidate point to both accuracy loss and fairness violation, thus producing effective poisoning samples. Experiments on two real datasets demonstrate the effectiveness and efficiency of our framework.

Wei Du, Xintao Wu, Hanghang Tong

Recent research on fair regression focused on developing new fairness notions and approximation methods as target variables and even the sensitive attribute are continuous in the regression setting. However, all previous fair regression research assumed the training data and testing data are drawn from the same distributions. This assumption is often violated in real world due to the sample selection bias between the training and testing data. In this paper, we develop a framework for fair regression under sample selection bias when dependent variable values of a set of samples from the training data are missing as a result of another hidden process. Our framework adopts the classic Heckman model for bias correction and the Lagrange duality to achieve fairness in regression based on a variety of fairness notions. Heckman model describes the sample selection process and uses a derived variable called the Inverse Mills Ratio (IMR) to correct sample selection bias. We use fairness inequality and equality constraints to describe a variety of fairness notions and apply the Lagrange duality theory to transform the primal problem into the dual convex optimization. For the two popular fairness notions, mean difference and mean squared error difference, we derive explicit formulas without iterative optimization, and for Pearson correlation, we derive its conditions of achieving strong duality. We conduct experiments on three real-world datasets and the experimental results demonstrate the approach's effectiveness in terms of both utility and fairness metrics.

Wei Du, Fahad Islam, Maxim Likhachev

Heuristic search-based planning techniques are commonly used for motion planning on discretized spaces. The performance of these algorithms is heavily affected by the resolution at which the search space is discretized. Typically a fixed resolution is chosen for a given domain. While a finer resolution allows for better maneuverability, it significantly increases the size of the state space, and hence demands more search efforts. On the contrary, a coarser resolution gives a fast exploratory behavior but compromises on maneuverability and the completeness of the search. To effectively leverage the advantages of both high and low resolution discretizations, we propose Multi-Resolution A* (MRA*) algorithm, that runs multiple weighted-A*(WA*) searches having different resolution levels simultaneously and combines the strengths of all of them. In addition to these searches, MRA* uses one anchor search to control expansions from these searches. We show that MRA* is bounded suboptimal with respect to the anchor resolution search space and resolution complete. We performed experiments on several motion planning domains including 2D, 3D grid planning and 7 DOF manipulation planning and compared our approach with several search-based and sampling-based baselines.