Jekyll2023-10-27T02:11:39-05:00https://starling.utdallas.edu/feed.xmlStARLinGStARLinG Lab, an Artificial Intelligence Research Group at the University of Texas at Dallas.StARLinG LabSaurabh Mathur: “Knowledge Intensive Learning of Cutset Networks”2023-05-08T16:46:00-05:002023-05-08T16:46:00-05:00https://starling.utdallas.edu/research-highlights/saurabh-mathur<p>My recent work deals with learning probabilistic models using expert advice. Specifically, we focus on a class of probabilistic models called cutset networks which allow fast inference. Our algorithm learns concise and accurate cutset networks from data using expert advice in the form of monotonic influence statements. Since cutset networks are interpretable, the learned models can help the expert understand relations between risk factors for medical conditions like Gestation Diabetes.</p>Saurabh Mathursxm200015@utdallas.eduhttps://saurabhmathur96.github.io/My recent work deals with learning probabilistic models using expert advice. Specifically, we focus on a class of probabilistic models called cutset networks which allow fast inference. Our algorithm learns concise and accurate cutset networks from data using expert advice in the form of monotonic influence statements. Since cutset networks are interpretable, the learned models can help the expert understand relations between risk factors for medical conditions like Gestation Diabetes.Siwen Yan: “Efficient Learning of Fair Models using Human Guidance and Privileged Information”2023-02-20T15:46:00-06:002023-02-20T15:46:00-06:00https://starling.utdallas.edu/research-highlights/siwen-yan<p>My recent research is centered around developing tree-based ensemble models and utilizing large language models. To enhance the performance and fairness of our models at deployment, we leverage privileged information as guidance in XGBoost and incorporate human advice into relational regression trees using functional gradient boosting for more effective and efficient learning. Additionally, we employ LLMs to discover logical rules and generate missing data.
In my previous works, I have focused on various areas including Statistical Relational AI and Healthcare, Graph Neural Networks, and Reinforcement Learning.</p>Siwen Yansiwen.yan@utdallas.eduhttps://dtrycode.github.io/My recent research is centered around developing tree-based ensemble models and utilizing large language models. To enhance the performance and fairness of our models at deployment, we leverage privileged information as guidance in XGBoost and incorporate human advice into relational regression trees using functional gradient boosting for more effective and efficient learning. Additionally, we employ LLMs to discover logical rules and generate missing data. In my previous works, I have focused on various areas including Statistical Relational AI and Healthcare, Graph Neural Networks, and Reinforcement Learning.CS73012019-08-15T00:00:00-05:002019-08-15T00:00:00-05:00https://starling.utdallas.edu/courses/fall-19-cs7301<p>If you are having difficulty viewing this page on mobile, access the document <a href="https://docs.google.com/document/d/15APP4wbI9vfmvfxn73j3b58MYCDZSneNKx4Ghj9yJgc/edit?usp=sharing">here</a></p>
<link rel="stylesheet" type="text/css" href="/assets/css/courses.css">
<iframe src="https://docs.google.com/document/d/e/2PACX-1vTXEmLMZYdODMOt53CqmJUfsuCwmGhp6aj18AdcW4STI907SY09k6lWIkCTfjemBMj0j91fF-S5LwFu/pub" frameborder="0"
style="height: 2000px; width:100%; overflow: hidden" ></iframe>StARLinG LabSriraam.Natarajan@utdallas.eduIf you are having difficulty viewing this page on mobile, access the document hereAlexander L. Hayes: “Relation Extraction for Relational Learning”2018-02-20T15:46:00-06:002018-02-20T15:46:00-06:00https://starling.utdallas.edu/research-highlights/alexander-hayes<p>My work is mostly centered on relational approaches to natural language processing. Approaches such as word2vec have increased interest in the context that words appear in beyond the bag-of-words model. The approach we focus on models the relations between words, document structure, and word attributes like part-of-speech; which can provide powerful insights for classification tasks.</p>
<p>This work has been applied to SEC Form S-1 Documents and known drug side effect information from the openFDA database. Work for the latter can be found on <a href="https://github.com/batflyer/DrugInteractionDiscovery">GitHub</a>.</p>
<p>I am involved in the <a href="/projects/nlp">Natural Language Processing</a> project.</p>Alexander L. Hayeshayesall@iu.eduhttps://hayesall.com/My work is mostly centered on relational approaches to natural language processing. Approaches such as word2vec have increased interest in the context that words appear in beyond the bag-of-words model. The approach we focus on models the relations between words, document structure, and word attributes like part-of-speech; which can provide powerful insights for classification tasks.Devendra Singh Dhami: “Structure-based Discovery of Drug-Drug Interactions”2018-02-20T15:46:00-06:002018-02-20T15:46:00-06:00https://starling.utdallas.edu/research-highlights/devendra-dhami<p>I am currently working on the problem of drug-drug interactions using structural data from the molecules and chemical reaction pathways which exist between interacting medications. Most of the work in this area has existed in the area of natural language processing where interactions are inferred through mining the medical literature. We take a non-conventional approach by considering the structure of the drugs themselves.</p>
<p>This method not only predicts interactions between different drugs but discovers new potential interactions, thereby exploiting deeper structures and drug features.</p>
<p>Relevant publications:</p>
<ol>
<li>Devendra Singh Dhami, Gautam Kunapuli, Mayukh Das, David Page, and Sriraam Natarajan, <em>Drug-Drug Interaction Discovery: Kernel Learning from Heterogeneous Similarities (Under review in IEEE Conference on Connected Health: Applications, Systems, and Engineering Technologies (CHASE), 2018)</em></li>
</ol>Devendra Singh Dhamihttps://sites.google.com/view/devendradhamiI am currently working on the problem of drug-drug interactions using structural data from the molecules and chemical reaction pathways which exist between interacting medications. Most of the work in this area has existed in the area of natural language processing where interactions are inferred through mining the medical literature. We take a non-conventional approach by considering the structure of the drugs themselves.Mayukh Das: “Human-Allied Problem Solving and Planning”2018-02-20T15:46:00-06:002018-02-20T15:46:00-06:00https://starling.utdallas.edu/research-highlights/mayukh-das<p>Knowledge-augmented approaches to sequential decision-making try to alleviate the limitations of data-driven techniques caused by noise, stochasticity and asymmetry of knowledge. They leverage rich knowledge assimilated by domain experts through years of experience to learn better behaviour. Recent years have witnessed a major research thrust in this direction and our group is an active contributor to this cause; focused on building human-in-the-loop frameworks for representation and elicitation of knowledge from, potentially multiple, expert(s) for sequential decision-making (RL), planning and prediction.</p>
<p>My research focuses on real-time/active elicitation of human knowledge, at varying levels of generality, for guided planning and problem-solving. Specifically, we address ‘asymmetry of knowledge’ where an AI planning system may have access to certain resources and vast computational power but may lack the necessary knowledge to prioritize among critical tasks. Human experts understand such priorities implicitly and we leverage that to generate better plans. Our DARPA-funded project “Communicating with Computers” has motivated research towards human-AI collaborative planning systems where both humans and AI agents solve problems together and learn from each other. Humans will teach new concepts, the agents will seek guidance when uncertain, and both will grow to augment one another. Part of my research also involves scaling Probabilistic Logic Models via approximation.</p>
<p>I am involved in the <a href="/projects/communicating-with-computers/">Communicating with Computers</a> project.</p>Mayukh Dashttps://sites.google.com/site/mayukhdas3986/Knowledge-augmented approaches to sequential decision-making try to alleviate the limitations of data-driven techniques caused by noise, stochasticity and asymmetry of knowledge. They leverage rich knowledge assimilated by domain experts through years of experience to learn better behaviour. Recent years have witnessed a major research thrust in this direction and our group is an active contributor to this cause; focused on building human-in-the-loop frameworks for representation and elicitation of knowledge from, potentially multiple, expert(s) for sequential decision-making (RL), planning and prediction. My research focuses on real-time/active elicitation of human knowledge, at varying levels of generality, for guided planning and problem-solving. Specifically, we address ‘asymmetry of knowledge’ where an AI planning system may have access to certain resources and vast computational power but may lack the necessary knowledge to prioritize among critical tasks. Human experts understand such priorities implicitly and we leverage that to generate better plans. Our DARPA-funded project “Communicating with Computers” has motivated research towards human-AI collaborative planning systems where both humans and AI agents solve problems together and learn from each other. Humans will teach new concepts, the agents will seek guidance when uncertain, and both will grow to augment one another. Part of my research also involves scaling Probabilistic Logic Models via approximation. I am involved in the Communicating with Computers project.Nandini Ramanan: “Precision Health”2018-02-20T15:46:00-06:002018-02-20T15:46:00-06:00https://starling.utdallas.edu/research-highlights/nandini-ramanan<p>We aim to bridge the gap between the machine learning community and the existing applications to healthcare. Our work involves developing efficient algorithms and probabilistic models using real-world data and expert knowledge. We employ state-of-the-art optimization techniques to understand the progression of disease symptoms and comorbidities over time. So far we have focused on customized probabilistic models in the context of Postpartum depression (PPD), cardiovascular disease, Alzheimer’s disease, and Parkinson’s disease; but our method is generalizable to summarizing patients with greater exactness to allow us to move toward personalized disease management strategies.</p>Nandini Ramanannxr173030@utdallas.eduhttps://www.nandhiniramanan5.com/We aim to bridge the gap between the machine learning community and the existing applications to healthcare. Our work involves developing efficient algorithms and probabilistic models using real-world data and expert knowledge. We employ state-of-the-art optimization techniques to understand the progression of disease symptoms and comorbidities over time. So far we have focused on customized probabilistic models in the context of Postpartum depression (PPD), cardiovascular disease, Alzheimer’s disease, and Parkinson’s disease; but our method is generalizable to summarizing patients with greater exactness to allow us to move toward personalized disease management strategies.Navdeep Kaur: “Relational Connectionist Models”2018-02-20T15:46:00-06:002018-02-20T15:46:00-06:00https://starling.utdallas.edu/research-highlights/navdeep-kaur<p><strong>Relational Connectionist Models</strong> bring together the complementary strengths of scalability and interpretability. We have considered learning Boltzmann machines for relational data; specifically generating features from lifted random walks that form the observed features of Boltzmann machines and produce efficient models when tested on six relational datasets.</p>
<p>Furthermore, we look at lifted relational connectionist models learned with the help of parameter tying and combining rules. This work is ongoing.</p>
<p>References:</p>
<ol>
<li>Navdeep Kaur, Gautam Kunapuli, Tushar Khot, William Cohen, Kristian Kersting and Sriraam Natarajan, “Relational Restricted Boltzmann Machines: A Probabilistic Logic Learning Approach”, ILP 2017.</li>
</ol>Navdeep KaurRelational Connectionist Models bring together the complementary strengths of scalability and interpretability. We have considered learning Boltzmann machines for relational data; specifically generating features from lifted random walks that form the observed features of Boltzmann machines and produce efficient models when tested on six relational datasets.Srijita Das: “Active Learning from Minimum Information”2018-02-20T15:46:00-06:002018-02-20T15:46:00-06:00https://starling.utdallas.edu/research-highlights/srijita-das<p>My research focus has been on learning robust machine learning models from small pool of complete instances.
A practical application of this has been in finding potential recruits given the availability of a small number of patients with complete information from a clinical study or survey data. This has been applied to real world medical problems like Parkinson’s, Alzheimer’s disease,rare disease and Post partum depression where a small number of patients with the complete feature set remains available and the goal is to identify potential recruits having easily available partial feature set for better disease prediction. My big goal is to build an intelligent decision support agent that can assist physicians in making better healthcare decisions.</p>Srijita Dashttps://sites.google.com/view/srijitadas/My research focus has been on learning robust machine learning models from small pool of complete instances. A practical application of this has been in finding potential recruits given the availability of a small number of patients with complete information from a clinical study or survey data. This has been applied to real world medical problems like Parkinson’s, Alzheimer’s disease,rare disease and Post partum depression where a small number of patients with the complete feature set remains available and the goal is to identify potential recruits having easily available partial feature set for better disease prediction. My big goal is to build an intelligent decision support agent that can assist physicians in making better healthcare decisions.