Huiqi Deng

Huiqi Deng

Auto-generated profile, based on publicly available information

Huiqi Deng

Shanghai Jiao Tong University, Shanghai, China

Research Focus & Metrics

Academic Career Length

7Years (2020-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
22 February 2024 in ACM Transactions on Intelligent Systems and Technology
Verified
Journal Article
Open access
HZHaiyan Zhao
Haiyan Zhao
HCHanjie Chen
Hanjie Chen
FYFan Yang
Fan Yang
NLNinghao Liu
Ninghao Liu
HDHuiqi Deng
Huiqi Deng
MDMengnan Du
Mengnan Du
Large language models (LLMs) have demonstrated impressive capabilities in natural language processing. However, their internal mechanisms are still unclear and this lack of transparency poses unwanted risks for downstream applications. Therefore, understanding and explaining these models is crucial for elucidating their behaviors, limitations, and social impacts. In this article, we introduce a taxonomy of explainability techniques and provide a structured overview of methods for explaining Transformer-based language models. We categorize techniques based on the training paradigms of LLMs: traditional fine-tuning-based paradigm and prompting-based paradigm. For each paradigm, we summarize the goals and dominant approaches for generating local explanations of individual predictions and global explanations of overall model knowledge. We also discuss metrics for evaluating generated explanations and discuss how explanations can be leveraged to debug models and improve performance. Lastly, we examine key challenges and emerging opportunities for explanation techniques in the era of LLMs in comparison to conventional deep learning models.

25 January 2024 in IEEE Transactions on Pattern Analysis and Machine Intelligence
Verified
Journal Article
Closed access
HDHuiqi Deng
Huiqi Deng
NZNa Zou
Na Zou
MDMengnan Du
Mengnan Du
WCWeifu Chen
Weifu Chen
GFGuocan Feng
Guocan Feng
QZQuanshi Zhang
Quanshi Zhang
Various attribution methods have been developed to explain deep neural networks (DNNs) by inferring the attribution/importance/contribution score of each input variable to the final output. However, existing attribution methods are often built upon different heuristics. There remains a lack of a unified theoretical understanding of why these methods are effective and how they are related. Furthermore, there is still no universally accepted criterion to compare whether one attribution method is preferable over another. In this paper, we resort to Taylor interactions and for the first time, we discover that fourteen existing attribution methods, which define attributions based on fully different heuristics, actually share the same core mechanism. Specifically, we prove that attribution scores of input variables estimated by the fourteen attribution methods can all be mathematically reformulated as a weighted allocation of two typical types of effects, i.e. , independent effects of each input variable and interaction effects between input variables. The essential difference among these attribution methods lies in the weights of allocating different effects. Inspired by these insights, we propose three principles for fairly allocating the effects, which serve as new criteria to evaluate the faithfulness of attribution methods. In summary, this study can be considered as a new unified perspective to revisit fourteen attribution methods, which theoretically clarifies essential similarities and differences among these methods. Besides, the proposed new principles enable people to make a direct and fair comparison among different methods under the unified perspective.

1 June 2020 in Pattern Recognition
Verified
Journal Article
Closed access
HDHuiqi Deng
Huiqi Deng
WCWeifu Chen
Weifu Chen
QSQi Shen
Qi Shen
AMAndy J. Ma
Andy J. Ma
PYPong C. Yuen
Pong C. Yuen
GFGuocan Feng
Guocan Feng

28 November 2023 in arXiv
Verified
Preprint
Open access
PDF file
HZHaiyan Zhao
Haiyan Zhao
HCHanjie Chen
Hanjie Chen
FYFan Yang
Fan Yang
NLNinghao Liu
Ninghao Liu
HDHuiqi Deng
Huiqi Deng
MDMengnan Du
Mengnan Du
Large language models (LLMs) have demonstrated impressive capabilities in natural language processing. However, their internal mechanisms are still unclear and this lack of transparency poses unwanted risks for downstream applications. Therefore, understanding and explaining these models is crucial for elucidating their behaviors, limitations, and social impacts. In this paper, we introduce a taxonomy of explainability techniques and provide a structured overview of methods for explaining Transformer-based language models. We categorize techniques based on the training paradigms of LLMs: traditional fine-tuning-based paradigm and prompting-based paradigm. For each paradigm, we summarize the goals and dominant approaches for generating local explanations of individual predictions and global explanations of overall model knowledge. We also discuss metrics for evaluating generated explanations, and discuss how explanations can be leveraged to debug models and improve performance. Lastly, we examine key challenges and emerging opportunities for explanation techniques in the era of LLMs in comparison to conventional machine learning models.

13 May 2024
Verified
Conference Paper
Open access
HPHongbin Pei
Hongbin Pei
YXYuheng Xiong
Yuheng Xiong
PWPinghui Wang
Pinghui Wang
JTJing Tao
Jing Tao
JLJialun Liu
Jialun Liu
XGXiaohong Guan
Xiaohong Guan
In the realm of semi-supervised graph learning, pseudo-labeling is a pivotal strategy to utilize both labeled and unlabeled nodes for model training. Currently, confidence score is the most frequently used pseudo-labeling measure, however, it suffers from poor calibration and issues in out-of-distribution data. In this paper, we propose memory disagreement (MoDis for short), a novel uncertainty measure for pseudo-labeling. We uncover that training dynamics offer significant insights into prediction uncertainty --- if a graph model makes consistent predictions for an unlabeled node throughout training, the corresponding predicted label is likely to be correct. Thus, the node should be suitable for pseudo-labeling. The basic idea is supported by recent studies on training dynamics. We implement MoDis as the entropy of an accumulated distribution that summarizes the disagreement of the model's predictions throughout training. We further enhance and analyze MoDis in case studies, which show nodes with low MoDis are suitable for pseudo-labeling as these nodes tend to be distant from boundaries in both graph and representation space. We design MoDis based pseudo-label selection algorithm and corresponding pseudo-labeling algorithm, which are applicable to various graph neural networks. We empirically validate MoDis on eight benchmark graph datasets. The experimental results show that pseudo labels given by MoDis have better quality in correctness and information gain, and the algorithm benefits various graph neural networks, achieving an average relative improvement of 3.11% and reaching up to 30.24% when compared to the wildly-used uncertainty measure, confidence score. Moreover, we demonstrate the efficacy of MoDis on out-of-distribution nodes.

12 January 2024 in Machine Intelligence Research
Verified
Journal Article
Closed access
PDF file
HZHuilin Zhou
Huilin Zhou
JRJie Ren
Jie Ren
HDHuiqi Deng
Huiqi Deng
XCXu Cheng
Xu Cheng
JZJinpeng Zhang
Jinpeng Zhang
QZQuanshi Zhang
Quanshi Zhang