Predicting Subjective Attributes in Visual Data - Zijun Wei

ABSTRACT: Recent progress in deep neural networks has revolutionized many computer vision tasks such as image classification, detection and segmentation. However, in addition to excelling in tasks that predict well-defined objective information, human-centered artificial intelligence systems should also be able to model subjective attributes, as defined by human perceptual behavior, that goes beyond the pure physical content of visual data. Example subjective tasks are the prediction of spatial or temporal regions that are interesting to humans (e.g., attract attention or are visually pleasing) and the recognition of subjective attributes (e.g., visually elicited sentiments). Better models for these tasks will improve the human-computer interaction experience in various applications. This thesis investigates several approaches to address the challenges in predicting those subjective attributes in visual data over a diverse set of tasks. I first present a novel framework for real-time automatic photo composition. The framework consists of a cost-effective data collection workflow, an efficient model training pipeline and a lightweight module to account for personalized preferences. Then I develop a novel and general algorithm to detect interesting segments in sequential data, which can be naturally applied to video summarization tasks. Furthermore, I propose methods that learn to represent sentiments elicited by images, in an unsupervised manner, using linguistic features extracted from large scale Web data. To conclude this thesis, I introduce a human-vision-inspired image classification algorithm that also predicts spatial visual attention even though no attention data was used for training it.  
The overall purpose of this seminar is to bring together people with interests in Computer Vision theory and techniques and to examine current research issues. This course will be appropriate for people who already took a Computer Vision graduate course or already had research experience in Computer Vision. To enroll in this course, you must either: (1) be in the PhD program or (2) receive permission from the instructors.

Each seminar will consist of multiple short talks (around 10 minutes) by multiple people. Students can register for 1 credit for CSE 656. Registered students must attend and present a minimum of 2 or 3 talks. Everyone else is welcome to attend. Fill in https://forms.gle/pCVXovgfMfQwGqG38 to subscribe to our mailing list for further announcement.
Join Zoom Meeting
https://stonybrook.zoom.us/j/91945227869?pwd=emhoZDFWVTV0MVdPWW5uVk43MjQzUT09

Meeting ID: 919 4522 7869
Passcode: 452304
One tap mobile
+16468769923,,91945227869# US (New York)
+13126266799,,91945227869# US (Chicago)

Dial by your location
        +1 646 876 9923 US (New York)
        +1 312 626 6799 US (Chicago)
        +1 301 715 8592 US (Germantown)
        +1 669 900 6833 US (San Jose)
        +1 253 215 8782 US (Tacoma)
        +1 346 248 7799 US (Houston)
        +1 408 638 0968 US (San Jose)
Meeting ID: 919 4522 7869
Find your local number: https://stonybrook.zoom.us/u/aCvAYWkRg  
The New York Academy of Sciences Presents AI for Materials: From Discovery to Production - A Virtual Symposium

Event Description: This interdisciplinary symposium covers the application of artificial intelligence (AI) throughout the entire life cycle of new materials -- from materials simulations and synthesis to translating research into high-volume industrial production.

Event Link & Registration: nyas.org/AI4Materials2020
The overall purpose of this seminar is to bring together people with interests in Computer Vision theory and techniques and to examine current research issues. This course will be appropriate for people who already took a Computer Vision graduate course or already had research experience in Computer Vision. To enroll in this course, you must either: (1) be in the PhD program or (2) receive permission from the instructors.

Each seminar will consist of multiple short talks (around 10 minutes) by multiple people. Students can register for 1 credit for CSE 656. Registered students must attend and present a minimum of 2 or 3 talks. Everyone else is welcome to attend. Fill in https://forms.gle/pCVXovgfMfQwGqG38 to subscribe to our mailing list for further announcement.
Abstract: Large language models (LLMs) may exhibit unintended or undesirable behaviors. Recent works have concentrated on aligning LLMs to mitigate harmful outputs. Despite these efforts, some anomalies indicate that even a well-conducted alignment process can be easily circumvented, whether intentionally or accidentally. Does alignment fine-tuning yield have robust effects on models, or are its impacts merely superficial? In this work, we make the first exploration of this phenomenon from both theoretical and empirical perspectives. Empirically, we demonstrate the elasticity of post-alignment models, i.e., the tendency to revert to the behavior distribution formed during the pre-training phase upon further fine-tuning. Leveraging compression theory, we formally deduce that fine-tuning disproportionately undermines alignment relative to pre-training, potentially by orders of magnitude. We validate the presence of elasticity through experiments on models of varying types and scales. Specifically, we find that model performance declines rapidly before reverting to the pre-training distribution, after which the rate of decline drops significantly. Furthermore, we further reveal that elasticity positively correlates with the increased model size and the expansion of pre-training data. Our findings underscore the need to address the inherent elasticity of LLMs to mitigate their resistance to alignment.

Speaker: Huajian Zhang

Location: CS2311
Abstract: This dissertation addresses the methodological disconnect between Natural Language Processing (NLP) and human-centric analysis by shifting the unit of analysis from document to human behavior in two broad respects: (i) time-ordering: modeling documents as sequential person-indexed behavioral observations, and (ii) person-level semantics: evaluation and explainability of models by their latent structure of psychological constructs rather than just its predictive accuracy against narrow proxy measures. First, we consider the most basic implication of language as a person's behaviors when measuring their psychological constructs: relationship between language sample size and model's predictive performance. We empirically show that the state-of-the-art transformers are often over-parameterized for typical NLP dataset sizes and can be reduced in dimensionality without performance loss. Establishing the author as the unit of analysis naturally allows us to treat their behavior as a time-ordered sequence. Second, we introduce a longitudinal evaluation framework that establishes ecologically valid evaluation settings, namely, cross-sectional and prospective generalization, and separates error measurement of the model into within-person dynamics and between-person differences. We demonstrate that traditional NLP evaluations based on random document splits can yield reversed conclusions under ecologically valid generalization settings. To address this, we develop models that capture the trajectory of mental states (e.g., mood shifts) rather than static traits. Third, moving into person-level semantics, we evaluate the latent structure of large language models using a novel machine behavior analytic framework. We find that while GPT-4 achieves high predictive correlation with self-reports, its latent symptoms structure diverges from clinical understanding. Finally, we propose a method for modeling multidimensional behaviors, embedding concurrent behavioral signals alongside language to predict future states. Taken together, this work suggests that operationalizing language as behavior advances NLP methods into a rigorous instrument for valid psychological inquiry.

Speaker: Adithya Ganesan

Location: Join Zoom Meeting (ID: 99021939129, Passcode: 569493)