In October of final 12 months, Amazon and Virginia Tech introduced the inaugural class of fellowship and school award recipients as a part of the Amazon–Virginia Tech Initiative for Environment friendly and Strong Machine Studying.
The initiative, launched in March of 2022, is targeted on analysis pertaining to environment friendly and strong machine studying. It helps analysis efforts led by Virginia Tech college members and offers a chance for doctoral college students within the School of Engineering conducting artificial-intelligence (AI) and machine studying (ML) analysis to use for Amazon fellowships.
Amazon and Virginia Tech at the moment introduced the 2023–2024 class of educational fellows and school analysis award recipients as a part of the joint initiative.
“Our honest appreciation to the Virginia Tech workforce for his or her unwavering dedication to excellence in each analysis and training, as mirrored within the impactful analysis and vital progress made in the course of the first 12 months of our partnership in addition to the high-quality proposals and fellowship functions we now have acquired this 12 months,” stated Reza Ghanadan, a senior principal analysis scientist in Alexa AI who leads the initiative at Amazon. “I stay up for persevering with our collaborations with the esteemed college and college students at Virginia Tech to advance our shared purpose of guaranteeing the robustness of machine studying methods whereas creating impactful AI functions throughout numerous domains enriching our society.”
“We’re very happy to proceed our partnership with Amazon to encourage and assist our college and scholar researchers targeted on discovering options to necessary and worldwide industry-focused issues throughout a spread of machine studying functions,” stated Naren Ramakrishnan, the Thomas L. Phillips Professor of Engineering and director of the Amazon–Virginia Tech initiative. “As we transfer into our second 12 months, we’re increasing into extra areas of machine studying corresponding to strong large-language-model deployment, combining massive language fashions with reasoning capabilities and multimodal interfaces.”
The 2 fellows and 5 college members chosen for this 12 months will every obtain funding to conduct analysis tasks at Virginia Tech throughout a number of disciplines. What follows are the recipients and their areas of analysis.
Minsu Kim, left, is pursuing a PhD in electrical and pc engineering. Ying Shen, proper, is pursuing a PhD in pc science.
Tutorial fellows
Minsu Kim is learning beneath Walid Saad, professor {of electrical} and pc engineering, and pursuing a PhD in electrical and pc engineering. Kim’s present analysis focus is constructing inexperienced, sustainable, and strong federated-learning options with tangible advantages for all AI-embedded merchandise that use federated studying and wi-fi communications. Kim’s work requires a extra holistic view of the lifecycle of federated-learning algorithms, together with information acquisition, algorithm and mannequin design, coaching, and inference/retraining.
Ying Shen is pursuing a PhD in pc science and learning beneath Lifu Huang and Ismini Lourentzou, each assistant professors within the division of pc science. Shen’s analysis pursuits lie in natural-language processing (NLP) and multimodal messages. Shen is especially obsessed with constructing extra human-like interactive brokers that higher perceive, interpret, and cause concerning the world round us.
High row, left to proper, Lifu Huang, assistant professor, division of pc science; Ruoxi Jia, assistant professor, division {of electrical} and pc engineering; Ming Jin, assistant professor; and backside row, left to proper, Ismini Lourentzou, assistant professor, division of pc science; and Xuan Wang, assistant professor, division of pc science.
College analysis award recipients
Lifu Huang, assistant professor, division of pc science, “Semi-parametric open area dialog era and analysis with multidimensional judgements from instruction tuning”
“The purpose of this venture is two-fold. First, it should develop an progressive, semi-parametric conversational framework that augments a big parametric dialog era mannequin with a big assortment of knowledge sources in order that desired information is dynamically retrieved and built-in to the generative mannequin, thus enhancing the adaptivity and scalability of the conversational agent in direction of open area subjects. Secondly, it should simulate fine-grained human judgements on machine-generated responses in multi-dimensions by leveraging instruction tuning on large-scale, pre-trained fashions. The pseudo-human judgements could be additional used to coach a light-weight multi-dimensional dialog evaluator or present suggestions to dialog era.”
Ruoxi Jia, assistant professor, division {of electrical} and pc engineering, “Chopping to the chase: Strategic information acquisition and pruning for environment friendly and strong machine studying”
“This venture focuses on creating strategic information acquisition and pruning methods that improve coaching effectivity, whereas addressing robustness towards sub-optimal information high quality by creating focused information acquisition methods that optimize the gathering of essentially the most worthwhile and informative information for a particular process; designing information pruning strategies to remove redundant and irrelevant information factors; and assessing the impression of those approaches on computational prices, mannequin efficiency, and robustness. When efficiently accomplished, the brand new methods will optimize the data-for-AI pipeline by accelerating the event of correct and accountable machine studying fashions throughout numerous functions.”
Ming Jin, assistant professor, “Secure reinforcement studying for interactive methods with stakeholder alignment”
“This venture goals to use a novel strategy to addressing the challenges of designing secure and aligned interactive methods. The analysis goals to develop a novel framework for stakeholder alignment utilizing reinforcement studying and recreation concept, and its outcomes can have necessary implications for a spread of functions — significantly within the realm of recommender methods.”
Ismini Lourentzou, assistant professor, division of pc science, “Diffusion-based scene-graph enabled embodied AI brokers”
“The target of this analysis is to design embodied AI brokers able to monitoring long-term adjustments within the surroundings, modeling how bodily attributes of a number of objects rework in response to brokers’ actions. The venture will even assess how brokers adapt to human preferences and suggestions by studying multimodal reward features from sub-optimal demonstrations. The end result of the proposed work will likely be extra intuitive and attuned embodied process assistants, enhancing their means to work together with the world in a pure and responsive method.”
Xuan Wang, assistant professor, division of pc science, “Reality-checking in open-domain dialogue era by means of self-talk”
“There’s a rising concern about accuracy and truthfulness of knowledge offered by open-domain dialogue era methods, corresponding to chatbots and digital assistants — significantly in healthcare and finance the place incorrect info can have critical penalties. This venture proposes a brand new fact-checking strategy for open-domain dialogue era utilizing language-model-based self-talk, which mechanically validates the generated responses and offers supporting proof.”
