Meta tags:
author=Jiyeon Han;

Headings (most frequently used words):

research, miscellanea, invited, talks, teaching,

Text of the page (most frequently used words):
the (21), choi (10), artifact (9), jiyeon (9), han (9), this (8), models (8), and (7), for (7), jaesik (7), generative (7), internal (6), with (6), arxiv (6), based (5), that (5), generation (5), creative (5), feature (4), maps (4), training (4), pretrained (4), github (4), fall (3), 2021 (3), spring (3), 2025 (3), research (3), equally (3), contributed (3), work (3), method (3), defective (3), generations (3), associated (3), regions (3), building (3), correction (3), related (3), lee (3), neural (3), networks (3), neurons (3), score (3), sample (3), data (3), kim (3), enhancing (3), deep (2), learning (2), analysis (2), 2020 (2), 2018 (2), proposes (2), automatically (2), identifying (2), responsible (2), external (2), classifier (2), applying (2), explainable (2), technique (2), grad (2), cam (2), highlight (2), most (2), strongly (2), introduce (2), sequential (2), algorithm (2), progressively (2), refines (2), suppressing (2), activations (2), cvpr (2), hwanil (2), haedong (2), jeong (2), units (2), local (2), activation (2), aaai (2), generating (2), present (2), new (2), metric (2), rarity (2), density (2), junho (2), diverse (2), rare (2), samples (2), end (2), stable (2), diffusion (2), achieve (2), interested (2), creativity (2), prof (2), website, built, upon, source, code, jon, barron, time, series, interpretability, interactivity, star, mooc, principles, programming, languages, teaching, seminar, ai4, summit, 2024, invited, talks, miscellanea, 2019, project, page, ijcai, anh, tong, kyowoon, confirmatory, bayesian, online, change, point, detection, covariance, structure, gaussian, processes, ali, tousi, automatic, 2022, investigate, are, closely, our, reveals, corresponding, often, exhibit, abrupt, changes, observation, propose, identify, unsupervised, way, understand, 2023, nbsp, assess, rareness, individual, generated, from, computes, inverse, estimated, manifold, real, oral, presentation, iclr, jung, woo, yunjey, evaluate, uncommonness, synthesized, images, explores, remain, faithful, original, outputs, gan, leveraging, normalizing, flows, estimation, enable, optimization, produce, lower, likelihood, soyeon, subeen, gans, free, approach, enhance, amplifying, low, frequency, features, shallow, blocks, gayoung, dahee, kwon, through, broader, focus, pushing, boundaries, what, can, also, deeply, understanding, analyzing, mechanisms, gain, insights, into, how, they, learn, generate, reason, linkedin, scholar, email, researcher, currently, beginning, postdoctoral, position, school, computing, science, under, supervision, recently, completed, korea, advised, dissertation, computational, assessing, output, reflects, core, interests, developing, reliable, methods, evaluating, capabilities, kaist, hao, richard, zhang, simon, fraser, university, gruvi, lab,

Text of the page (random words):
jiyeon han jiyeon han i m jiyeon han a researcher in ai currently beginning my postdoctoral position in the gruvi lab at the school of computing science simon fraser university under the supervision of prof hao richard zhang i recently completed my ph d at kaist korea advised by prof jaesik choi my dissertation computational creativity in ai assessing and enhancing the creative output of generative models reflects my core research interests in developing reliable methods for evaluating creativity and enhancing the creative capabilities of generative models email cv scholar linkedin github research i m interested in building creative ai through generative models with a broader focus on pushing the boundaries of what ai can achieve i m also deeply interested in understanding and analyzing the internal mechanisms of ai models to gain insights into how they learn generate and reason enhancing creative generation on stable diffusion based models jiyeon han dahee kwon gayoung lee junho kim jaesik choi contributed equally cvpr 2025 github arxiv we present a training free approach to enhance creative generation on stable diffusion based models we achieve this by amplifying low frequency features in the shallow blocks in the pretrained models diverse rare sample generation with pretrained gans subeen lee jiyeon han soyeon kim jaesik choi aaai 2025 github arxiv this work explores a method for generating diverse rare samples that remain faithful to the original outputs of a pretrained gan by leveraging normalizing flows for density estimation we enable end to end optimization to produce samples with lower data likelihood rarity score a new metric to evaluate the uncommonness of synthesized images jiyeon han hwanil choi yunjey choi junho kim jung woo ha jaesik choi iclr 2023 nbsp oral presentation github arxiv we present a new metric to assess rareness of an individual generated sample from pretrained generative models rarity score computes the inverse density of the sample estimated by k nn based data manifold of real training data an unsupervised way to understand artifact generating internal units in generative neural networks haedong jeong jiyeon han jaesik choi aaai 2022 arxiv we investigate internal neurons that are closely associated with defective generations in generative neural networks our analysis reveals that neurons corresponding to artifact regions often exhibit abrupt local activation changes based on this observation we propose the local activation score to identify artifact related neurons automatic correction of internal units in generative neural networks ali tousi haedong jeong jiyeon han hwanil choi jaesik choi equally contributed cvpr 2021 arxiv this work proposes a method for automatically identifying internal feature maps responsible for defective generations by training an external artifact classifier and applying the explainable ai technique grad cam we highlight the feature maps most strongly associated with artifact regions building on this we introduce a sequential correction algorithm that progressively refines the generation by suppressing artifact related activations confirmatory bayesian online change point detection in the covariance structure of gaussian processes jiyeon han kyowoon lee anh tong jaesik choi equally contributed ijcai 2019 project page arxiv this work proposes a method for automatically identifying internal feature maps responsible for defective generations by training an external artifact classifier and applying the explainable ai technique grad cam we highlight the feature maps most strongly associated with artifact regions building on this we introduce a sequential correction algorithm that progressively refines the generation by suppressing artifact related activations miscellanea invited talks lg ai seminar 2025 ai4 research summit 2024 teaching ta deep learning fall 2021 ta ai based time series analysis spring 2021 ta deep learning fall 2020 ta interpretability and interactivity in ai spring 2020 ta star mooc fall 2018 ta principles of programming languages spring 2018 this website is built upon jon barron s source code