AI Innovation Institute

Abstract:

It is known that models like large language models (LLMs) can often suggest colloquial plans given verbal descriptions of tasks, yet they are unable to reliably provide executable and verifiable plans given formally specified environments. In this talk, I will discuss a strand of efforts to have LLMs generate accurate and explainable plans in textual simulations. Instead of directly generating the plan or actions, LLMs are prompted to generate Planning Domain Definition Language (PDDL) that specifies the environment (domain file) and the task (problem file), which can then be deterministically solved with an off-the-shelf planner. In a 3-phase study, my collaborators and I first observed that it is possible but very challenging for LLMs to generate long-form code such as PDDL domain and problem files given textual specifications. Next, we devise methodologies for LLMs to iteratively generate and refine problem files while exploring a partially-observed, simulated, textual environment. Finally, we show that domain files are even more difficult to generate correctly, even on well-established planning tasks such as BlocksWorld. Finally, I will discuss ongoing efforts to improve said ability of structured generation and promising frontiers to explore.

Bio:
Li Harry Zhang is an assistant professor at Drexel University, focusing on Natural Language Processing (NLP) and artificial intelligence (AI). He obtained his PhD degree from the University of Pennsylvania advised by Prof. Chris Callison-Burch. Prior, he obtained his Bachelor's degree at the University of Michigan mentored by Prof. Rada Mihalcea and Prof. Dragomir Radev. His current research uses large language models (LLMs) to reason and plan via symbolic and structured representations. He has published more than 20 peer-reviewed papers in NLP and AI conferences, such as ACL, EMNLP, and AACL, that have been cited more than 1,000 times. He also consistently serves as Area Chair, Session Chair, and reviewer in those venues. Being a musician, producer, and content creator having over 50,000 subscribers, he is also passionate in the research of AI music and creativity.

Zoom Link: https://stonybrook.zoom.us/j/98533029054?pwd=5FXO6lWGTJssCADEYkYbA7sjaacPRX.1

Meeting ID: 985 3302 9054

Passcode: 436997

Abstract:

Semantic segmentation, the task of assigning a semantic label to each pixel in an image, is a fundamental problem in the field of Computer Vision. with crucial applications in domains like autonomous driving, drone imagery and medical image analysis. Despite advancements in deep learning architectures, state-of-the-art models still heavily depend on large-scale pixel-level annotations, which are costly and time-consuming to acquire. To address this issue, Semi-Supervised Segmentation (SSS) has emerged as a promising solution, leveraging a small set of labeled images alongside a larger corpus of unlabeled data to reduce the annotation burden. In this proposal, I aim to investigate the challenges of SSS and propose approaches to address them. Existing SSS methods rely on a teacher-student framework to generate pseudo-labels for unlabeled images, which are then used for model training. However, this approach presents two major challenges. Pixel-level consistency fails to effectively capture contextual information, and pseudo-labels are noisy, especially in the early stages of training. To address the challenge of noisy pseudo-labels, existing methods rely on confidence-based thresholding to identify reliable pseudo-labels. However, during early training phases, when the model is poorly calibrated, this approach can select high-confidence but noisy pseudo-labels. To address this, we propose a novel approach that reduces reliance on model confidence to select reliable pseudo-labels. Our method employs an ensemble of a segmentation model and an object detection model to select more reliable pseudo-labels, which are then used to weight pseudo-labels using rank statistics, reducing the influence of noisy labels in training. Next, to address both the challenge of capturing contextual information and noisy pseudo-labels I introduce a novel Multi-scale Patch-based Multi-label Classifier (MPMC), which incorporates patch-level contextual information and reduces the impact of noisy pixel pseudo-labels by using the predictions of the patch-level Multi-label classifier to detect noisy labels, enhancing overall segmentation performance. While my work so far has focused on effectively utilizing unlabeled data to improve segmentation performance, as part of our future work, I will explore the use of textual information, such as category descriptions, for segmentation tasks. In limited labeled data scenarios it is more challenging to align visual features with textual features from large language models (LLMs).