A5005779128- Topic Modeling
- Natural Language Processing Techniques
- Explainable Artificial Intelligence (XAI)
- Multimodal Machine Learning Applications
- Adversarial Robustness in Machine Learning
- Image Retrieval and Classification Techniques
- Advanced Image and Video Retrieval Techniques
- Medical Image Segmentation Techniques
- Neural Networks and Applications
- Machine Learning and Data Classification
- Advanced Graph Neural Networks
- Bayesian Modeling and Causal Inference
- Semantic Web and Ontologies
- Text Readability and Simplification
- Speech and dialogue systems
- Image Processing and 3D Reconstruction
- Video Analysis and Summarization
- Machine Learning and Algorithms
- Time Series Analysis and Forecasting
- AI in cancer detection
- Image and Object Detection Techniques
- Domain Adaptation and Few-Shot Learning
- Stock Market Forecasting Methods
- Handwritten Text Recognition Techniques
- Anomaly Detection Techniques and Applications
Dr. Hari Singh Gour University
2025
TIFR Centre for Interdisciplinary Sciences
2023-2025
Narendra Dev University of Agriculture and Technology
2024
Indian Institute of Technology Ropar
2019-2024
Tata Institute of Fundamental Research
2023-2024
Columbia University Irving Medical Center
2023-2024
Banaras Hindu University
2024
University of California, Irvine
2016-2023
Google (United States)
2009-2023
Harcourt Butler Technical University
2023
Despite widespread adoption, machine learning models remain mostly black boxes. Understanding the reasons behind predictions is, however, quite important in assessing trust, which is fundamental if one plans to take action based on a prediction, or when choosing whether deploy new model. Such understanding also provides insights into model, can be used transform an untrustworthy model prediction trustworthy one.
Despite widespread adoption in NLP, machine learning models remain mostly black boxes.Understanding the reasons behind predictions is, however, quite important assessing trust a model.Trust is fundamental if one plans to take action based on prediction, or when choosing whether not deploy new model.In this work, we describe LIME, novel explanation technique that explains of any classifier an interpretable and faithful manner.We further present method explain by presenting representative...
We introduce a novel model-agnostic system that explains the behavior of complex models with high-precision rules called anchors, representing local, "sufficient" conditions for predictions. propose an algorithm to efficiently compute these explanations any black-box model high-probability guarantees. demonstrate flexibility anchors by explaining myriad different domains and tasks. In user study, we show enable users predict how would behave on unseen instances less effort higher precision,...
The remarkable success of pretrained language models has motivated the study what kinds knowledge these learn during pretraining. Reformulating tasks as fill-in-the-blanks problems (e.g., cloze tests) is a natural approach for gauging such knowledge, however, its usage limited by manual effort and guesswork required to write suitable prompts. To address this, we develop AutoPrompt, an automated method create prompts diverse set tasks, based on gradient-guided search. Using show that masked...
Although measuring held-out accuracy has been the primary approach to evaluate generalization, it often overestimates performance of NLP models, while alternative approaches for evaluating models either focus on individual tasks or specific behaviors. Inspired by principles behavioral testing in software engineering, we introduce CheckList, a task-agnostic methodology models. CheckList includes matrix general linguistic capabilities and test types that facilitate comprehensive ideation, as...
Matthew E. Peters, Mark Neumann, Robert Logan, Roy Schwartz, Vidur Joshi, Sameer Singh, Noah A. Smith. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and 9th International Joint (EMNLP-IJCNLP). 2019.
Understanding why machine learning models behave the way they do empowers both system designers and end-users in many ways: model selection, feature engineering, order to trust act upon predictions, more intuitive user interfaces. Thus, interpretability has become a vital concern learning, work area of interpretable found renewed interest. In some applications, such are as accurate non-interpretable ones, thus preferred for their transparency. Even when not accurate, may still be is...
Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gardner, Sameer Singh. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and 9th International Joint (EMNLP-IJCNLP). 2019.
As machine learning black boxes are increasingly being deployed in domains such as healthcare and criminal justice, there is growing emphasis on building tools techniques for explaining these an interpretable manner. Such explanations leveraged by domain experts to diagnose systematic errors underlying biases of boxes. In this paper, we demonstrate that post hoc rely input perturbations, LIME SHAP, not reliable. Specifically, propose a novel scaffolding technique effectively hides the any...
Complex machine learning models for NLP are often brittle, making different predictions input instances that extremely similar semantically. To automatically detect this behavior individual instances, we present semantically equivalent adversaries (SEAs) – semantic-preserving perturbations induce changes in the model’s predictions. We generalize these into adversarial rules (SEARs) simple, universal replacement on many instances. demonstrate usefulness and flexibility of SEAs SEARs by...
Dheeru Dua, Yizhong Wang, Pradeep Dasigi, Gabriel Stanovsky, Sameer Singh, Matt Gardner. Proceedings of the 2019 Conference North American Chapter Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers). 2019.
Due to their complex nature, it is hard characterize the ways in which machine learning models can misbehave or be exploited when deployed. Recent work on adversarial examples, i.e. inputs with minor perturbations that result substantially different model predictions, helpful evaluating robustness of these by exposing scenarios where they fail. However, malicious are often unnatural, not semantically meaningful, and applicable complicated domains such as language. In this paper, we propose a...
GPT-3 can perform numerous tasks when provided a natural language prompt that contains few training examples. We show this type of few-shot learning be unstable: the choice format, examples, and even order examples cause accuracy to vary from near chance state-of-the-art. demonstrate instability arises bias models towards predicting certain answers, e.g., those are placed end or common in pre-training data. To mitigate this, we first estimate model's each answer by asking for its prediction...
Matt Gardner, Yoav Artzi, Victoria Basmov, Jonathan Berant, Ben Bogin, Sihao Chen, Pradeep Dasigi, Dheeru Dua, Yanai Elazar, Ananth Gottumukkala, Nitish Gupta, Hannaneh Hajishirzi, Gabriel Ilharco, Daniel Khashabi, Kevin Lin, Jiangming Liu, Nelson F. Phoebe Mulcaire, Qiang Ning, Sameer Singh, Noah A. Smith, Sanjay Subramanian, Reut Tsarfaty, Eric Wallace, Ally Zhang, Zhou. Findings of the Association for Computational Linguistics: EMNLP 2020.
Matrix factorization approaches to relation extraction provide several attractive features: they support distant supervision, handle open schemas, and leverage unlabeled data.Unfortunately, these methods share a shortcoming with all other distantly supervised approaches: cannot learn extract target relations without existing data in the knowledge base, likewise, models are inaccurate for sparse data.Rule-based extractors, on hand, can be easily extended novel improved but relations, through...
For accurate entity linking, we need to capture various information aspects of an entity, such as its description in a KB, contexts which it is mentioned, and structured knowledge. Additionally, linking system should work on texts from different domains without requiring domain-specific training data or hand-engineered features. In this present neural, modular that learns unified dense representation for each using multiple sources information, description, around mentions, fine-grained...
Eric Wallace, Yizhong Wang, Sujian Li, Sameer Singh, Matt Gardner. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and 9th International Joint (EMNLP-IJCNLP). 2019.
Recent research on entity linking (EL) has introduced a plethora of promising techniques, ranging from deep neural networks to joint inference. But despite numerous papers there is surprisingly little understanding the state art in EL. We attack this confusion by analyzing differences between several versions EL problem and presenting simple yet effective, modular, unsupervised system, called Vinculum, for linking. conduct an extensive evaluation nine data sets, comparing Vinculum with two...
Reading comprehension has recently seen rapid progress, with systems matching humans on the most popular datasets for task. However, a large body of work highlighted brittleness these systems, showing that there is much left to be done. We introduce new English reading benchmark, DROP, which requires Discrete Reasoning Over content Paragraphs. In this crowdsourced, adversarially-created, 96k-question system must resolve references in question, perhaps multiple input positions, and perform...
The ongoing pandemic has heightened the need for developing tools to flag COVID-19-related misinformation on internet, specifically social media such as Twitter. However, due novel language and rapid change of information, existing detection datasets are not effective evaluating systems designed detect this topic. Misinformation can be divided into two sub-tasks: (i) retrieval misconceptions relevant posts being checked veracity, (ii) stance identify whether Agree, Disagree, or express No...
Modeling human language requires the ability to not only generate fluent text but also encode factual knowledge. However, traditional models are capable of remembering facts seen at training time, and often have difficulty recalling them. To address this, we introduce knowledge graph model (KGLM), a neural with mechanisms for selecting copying from that relevant context. These enable render information it has never before, as well out-of-vocabulary tokens. We Linked WikiText-2 dataset,...