⏰ Spring 2023 Projects
Discover more from AI@UIUC Newsletter
⏰ Spring 2023 Projects
🎶 Tats on my neck and my arm (Tats on my neck and my arm, yeah, what?) 🎶
Greetings, everyone! We are thrilled to unveil the Spring 2023 projects for AI @ UIUC. After an extensive application process, we are eager to dive in and get to work.
Projects
Research #1: Multilingual Natural Language Processing with Synthetic Data
Abstract from paper:
“Large language models (LLMs) have been leveraged to generate task-specific synthetic data on a variety of natural language processing (NLP) tasks, but their use in the cross-lingual setting has not been explored. We propose a novel pipeline for data augmentation for multilingual NLP tasks using LLMs. We will prompt and fine-tune pretrained LLMs on ground-truth or English examples from target languages and tasks from existing multilingual datasets like XNLI to generate high-quality samples of similar data. Then we will evaluate the performance of classifiers trained or models fine-tuned on this data and compare against models trained only on monolingual data, existing multilingual data, or both multilingual and augmented data. We will investigate the effect on downstream performance of the synthetic data based on the size of the data generator, choice of target task, quantity of augmented data, and choice of target language. On models fine-tuned with our method, we hope to see improvements in performance on a variety of new tasks, especially in specific tasks and low-resource languages that lack a large amount of easily available data.”
Research #2: High-resolution Video Object Segmentation
Abstract from paper:
“We propose a method for high-resolution video object segmentation (VOS) using memory networks and hierarchical matching. We extend recent work in semi-supervised, offline learning, and attention-based VOS methods into the high-resolution domain without significantly increasing memory consumption. Notably, our method does not require training on high-resolution videos.”
~ Powered by a 1.5 mil credit grant from NCSA’s Delta A100 Quad Cluster
Research #3: Personalized Federated Learning with Lottery Tickets
Abstract from paper:
“Federated learning is a common machine learning paradigm for training multiple models collectively. This allows for privacy while also offering stronger performance than what participants can achieve individually. One promising area of research is personalized federated learning, which aims to adapt models to individual clients without sacrificing global performance. We propose a novel. personalized federated learning framework that uses the Lottery Ticket hypothesis to customize client architectures during training, improving performance on both local and global distributions.”
~ Powered by a 750k credit grant from NCSA’s Delta A100 Quad Cluster
Updates:
Published: https://arxiv.org/abs/2306.13264 (FL-ICML 23)
Research #4: Applied Machine Learning in Earth Science
Abstract from paper:
“This research paper aims to examine various common approaches for climate forecasting and related applications such as pollution, wildfires, and extreme weather forecasting. We will explore diverse models, including auto-differentiable ensemble Kalman filters, KalmanNet, deep factors for forecasting, and sub-seasonal climate forecasting with machine learning models. Furthermore, we will evaluate the effectiveness of these models by implementing reanalysis datasets. In addition, we will delve into cloud removal from satellite imagery. Our goal is to provide valuable insights into the latest and most common approaches for climate forecasting and related applications.”
Technical Staff: Be+ [in partnership with DaxFlame]
This semester, our technical staff will be working with DaxFlame, a Youtuber/Actor along with industry partners from top tech companies to build and deploy Be+: a large language model (LLM) powered chatbot protocol that replicates the user’s mannerisms to function as a text-based “second you”. We are grateful to have such an awesome opportunity for collaboration and look forward to building in the LLM space.
New Member Event
AI-Generated Image of the Day (“i want bentley, i want money”)
