• Organizing Committee

General Chairs

Prof. Seong-Whan Lee

Korea University

Dr. Lidong Zhou

Corporate Vice President of Microsoft
and Managing Director of
Microsoft Research Asia

Program Chairs

Jinwoo Shin

KAIST

Alice Oh

KAIST

Miran Lee

Microsoft Research

Seungyong Lee

POSTECH

Meeyoung Cha

KAIST

Seung-won Hwang

SNU

Seungryong Kim

Korea University

PROGRAM SCHEDULE

Sessions INFORMATION

Panel Discussion

Theme

• Recent advances and challenges in AI and ML for healthcare and climate change

Description

Recent advances in artificial intelligence and machine learning have been applied to some of the most difficult problems our society is facing. We will look at two of those problems in this panel: healthcare and climate. We invite three experts who work at the intersection of AI/ML and healthcare/climate. Through this panel, we aim to understand the current state of the research, what challenges remain, and how the scientific community can work together to solve these most pressing problems.

Panel speakers (Moderator: Alice Oh, KAIST)

Prof.Finale Doshi-Velez

Affiliation Harvard University

Prof.Edward Choi

Affiliation KAIST

Prof.Yoo-Geun Ham

Affiliation Chonnam National University

Technology showcase

Presentation Lists

• Deep Learning for Structure-based Drug Design of A2A Adenosine Receptor
• Presenter: Sanghee Yoon, Sun Choi (Ewha Womans University) Improvements in biophysical methods further enhanced the appeal of reverse pharmacology along with structure-based drug design (SBDD), as the three-dimensional structures of proteins become more available, permitting the use of in silico tools to screen compounds for potential binding with the protein. Over the past few years, artificial intelligence (AI), particularly machine learning (ML), deep learning (DL), and reinforcement learning (RL) algorithms have been used to improve the drug discovery process. In G protein-coupled receptor (GPCR) which located on the cell surface that recognizing extracellular substances and transmit signals through cell membranes, the dynamics and functions of GPCRs are finely regulated depending on the type of bound ligands. Using A2A adenosine receptor (A2AAR) as a model system, we directly generate compounds from the given targets and analyze their drug-like properties.
• Investigating Gender Bias in Multilingual Neural Machine Translation Systems
• Presenter: Minwoo Lee, Kyomin Jung (Seoul National University) Recent works have shown that multilingual neural machine translation (MNMT) models are effective at improving translation performance of low-resource languages. This performance benefit has been mainly attributed to language-agnostic knowledge transfer during training. However, it is unclear whether transfer of gender bias, which bilingual NMT models have shown to exhibit, also occur. In this work, we find that contrary to previous claims, gender bias transfer does not necessarily occur for multilingual models, and multilingual models are capable of achieving higher gender accuracy than bilingual models.​
• Language-Driven Image-to-Video Generation with Generative Neural Radiance Fields
• Presenter: Sunghwan Hong, Seungryong Kim (Korea University) It has been a long-standing goal of Computer Vision, Multimedia, and Graphics to develop the ability to generate and manipulate videos. Recently, a new task called Text-Image-to-Video generation (TI2V) was introduced. Despite its exclusive performance, it does not properly reason about the 3D structure of the scene and may suffer from entangled representations, preventing from generating a diverse video with 3D awareness and explicitly manipulating attributes of a particular object of interest, e.g., object appearance, size or pose. ​
• Question Answering for Supporting Multi-modal Clinical Decision Making
• Presenter: Seongsu Bae, Yoonjae Choi (KAIST) There are a series of uni-modal EHR QA datasets such as emrQA (Pampari et al., 2018), MIMICSQL (Wang et al., 2020), and EHRSQL (Lee et al., 2022). However, none can handle access over the multi-modal data source. To make a breakthrough, we aim to build a new challenging benchmark and develop the model for multi-modal EHR QA.​
• Towards Battery-free System for Complex Context Sensing
• Presenter: Seungwoo Shim, Song Min Kim (KAIST) Remarkable advances in machine learning techniques has enabled lots of services improving the quality of everyday lives. For example, rich information in visual data provides various applications ranging from public safety to smart shopping and mobile healthcare. High-performance sensors such as LiDAR and high-resolution cameras are expensive, are difficult to attach to small objects, and have significant limitations in deployment due to privacy concerns. Instead, we propose battery-free sensor tags for understanding complex context. Low quality data due to limited energy budget can be compensated with multiple sensors.​
• Efficient Container Networking for Edge Computing
• Presenter: Yeonho Yoo, Chuck Yoo (Korea University) The edge devices adopt the internet of things (IoT) in various industrial fields and have become a new computing paradigm that tackles a set of technological challenges in industrial control, automation, and intelligence. As current edge device applications demand real-time control to manage heterogeneous devices properly, it is necessary to offer an efficient computing environment, which places computing resources near the edge to reduce network latency and data processing delays. Specifically, we focus on containers, which are a major building block in edge computing. We analyze the existing networking stack of containers in smart edge devices and then design and implement a streamlined container networking stack for edge computing called SCON. We observe that SCON provides both high network performance and low CPU usages. ​
• Leveraging Large Language Models for Automatic and Traceable Debugging
• Presenter: Sungmin Kang, Shin Yoo (KAIST) Debugging is an important developer activity, and as such numerous techniques have been proposed to automated the task. Nonetheless, developer surveys consistently indicate the need for explainable techniques, which have been underexplored. In this work, we propose that large language models (LLMs) are an appropriate solution to this challenge: using their reasoning capability, we can automatically generate debugging artifacts, and ask the LLM to justify its suggestion in natural language. Our preliminary results indicate competitive performance with existing work and a capability of generating explanations.​
• Enhancing the Local Alignment in Chest X-ray Domain
• Presenter: Gangwoo Kim, Jaewoo Kang (Korea University) To improve the local alignment performance in the chest X-ray domain (i..e., phrase grounding, object detection), we propose a novel method to generate the synthetic datasets for the phrase grounding task by leveraging a recent model for aligning image-text pairs, CheXzero [1]. Our method can automatically generate datasets for the phrase grounding task by using bounding box labels from object detection dataset. Fine-tuned on them, CLIP-based backbone model provides pseudo-label on the unlabeled image-report pairs. We finally evaluate our models on a recent benchmark in the chest X-ray domain, MS-CXR [2]​.
• Audio Generation using HuBERT and EnCodec
• Presenter: Youngdo Ahn, Jong Won Shin (GIST) To generate speech that sounds natural is an interesting topic. Recently, audio generation has been studied to generate natural and continuous audio samples for movie production or the development of virtual reality. The recent audio generation model is constructed with the audio language model and tokenizers with pre-trained models. In this presentation, we introduce our audio generation model based on HuBERT and EnCodec as tokenizers. To verify the ability of the audio generation model, we evaluated the word error rates (WER) using an automatic speech recognition model. ​
• Revisiting and Reformulating Machine Learning-based Drug Repositioning
• Presenter: Jung-Hyun Won, Lee Hyeong Ki (Seoul National University) In this study, we aimed to point out problems in the current task formulation of machine learning (ML)-based drug repositioning (DR) studies. We identified the spurious correlation between data collection pattern of drug characteristics and approved indications of drug by performing unsupervised clustering analysis. In addition, we found that train-test split based on random sampling led to the overestimation of the performances of existing DR models. Now, we are planning to reformulate DR problems as predicting investigational disease and generating eligibility criteria of clinical trials for a drug-disease pair.​
• Accelerating Mixture of Experts Inference via DNN-Based Expert Activation Prediction
• Presenter: Ranggi Hwang, Minsoo Rhu (KAIST) In recent years, mixture of experts (MoE) models have been adopted in various machine learning (ML) applications with remarkable performance, including natural language processing (NLP), image classifications, and recognitions. Although MoE effectively extends model capacity with minimally increasing computation requirements, it causes compute under-utilization and huge memory overheads with increased latency for inference. To address these problems, we proposed a DNN-based prediction network for expert activation, which can effectively decrease total latency while maximizing resource utilization.​
• Towards accurate and efficient recommender systems
• Presenter: SeongKu Kang, Hwanjo Yu (POSTECH) Recent recommender systems tend to adopt increasingly complex and large models to better understand the complex nature of user-item interactions. Accordingly, the inference latency increases as well, which has become a major obstacle to deployment. We focus on knowledge distillation to generate a powerful but compact model. Despite its breakthrough in classification problems, knowledge distillation for recommendation models and ranking problems has not been studied well in the previous literature. Our current research is devoted to developing distillation methods tailored for recommender systems.​
• Panoramic Vision Transformer for Saliency Detection in 360 Videos
• Presenter: Heeseung Yun, Gunhee Kim (Seoul National University) 360° video saliency detection is one of the challenging benchmarks for 360° video understanding since non-negligible distortion and discontinuity occur, and capture-worthy viewpoint is ambiguous by nature. We present a new framework named Panoramic Vision Transformer (PAVER). We design the encoder using Vision Transformer with deformable convolution, which enables us not only to plug pretrained models from normal videos into our architecture without additional modules or finetuning but also to perform geometric approximation only once, unlike previous deep CNN-based approaches. Thanks to its powerful encoder, PAVER can learn the saliency from three simple relative relations among local patch features, outperforming state-of-the-art models for multiple benchmarks.
• Avatron: 3D Face Avatar and Voice Generation From Text
• Presenter: Se-Yun Um, Hong-Goo Kang (Yonsei University) We propose a novel 3D facial animation model that does not require explicit temporal synchronization between talking faces and speech. Due to time domain differences between videos and speech, most approaches apply resampling to linguistic features from a pre-trained automatic speech recognition (ASR) model, resulting in unnatural facial movements. However, we efficiently utilize intermediate feature from a text-to-speech (TTS) model and feed it into an avatar model to be converted to the vertices of a 3D face avatar. Since our proposed model does not require an ASR model, we are able to reduce the number of parameters and computational complexity of the whole solution. ​
• Collective Relevance Labeling for Passage Retrieval
• Presenter: Jihyuk Kim, Seung-won Hwang (Seoul National University) Deep learning for Information Retrieval (IR) requires a large amount of high-quality query-document relevance labels, but such labels are inherently sparse. Label smoothing redistributes some observed probability mass over unobserved instances, often uniformly, uninformed of the true distribution. In contrast, we propose knowledge distillation for informed labeling, without incurring high computation overheads at evaluation time. Our contribution is designing a simple but efficient teacher model which utilizes collective knowledge, to outperform stateof-the-arts distilled from a more complex teacher model.
• Meta-Learning for Single Image Deblurring with Core-Model​
• Presenter: Hyunjin Son, Kyoung Mu Lee (Seoul National University) Previous end-to-end learning networks for deblurring are trained by large-scale external datasets, so it cannot exploit internal information within a given test image. In this work, we present a new approach for single image deblurring based on meta-learning using both external and internal data. First, we train the network via meta-learning with support set from pretrained network by external data. Then in the test time, this meta-trained model allows fast adaptation to target image with only a few parameter updates. Furthermore, we finetune the model and perform deblur in patches. The sub-network, called core-model learned relationship between patches in training time. To utilize the internal information to enhance the performance of deblurring, we can select important patches at test time.​
• GRIT-VLP: Grouped Mini-batch Sampling for ​Efficient Vision and Language Pre-training
• Presenter: Jaeseok Byun, Taesup Moon (Seoul National University) Most of the currently existing vision and language pre-training (VLP) methods have mainly focused on how to extract and align vision and text features. In contrast to the mainstream VLP methods, we highlight that two routinely applied steps during pre-training have crucial impact on the performance of the pre-trained model: in-batch hard negative sampling for image-text matching (ITM) and assigning the large masking probability for the masked language modeling (MLM). After empirically showing the unexpected effectiveness of above two steps, we systematically devise our GRIT-VLP, which adaptively samples mini-batches for more effective mining of hard negative samples for ITM while maintaining the computational cost for pre-training. Our method consists of three components: 1) GRouped mIni-baTch sampling (GRIT) strategy that collects similar examples in a mini-batch, 2) ITC consistency loss for improving the mining ability, and 3) enlarged masking probability for MLM. Consequently, we show our GRIT-VLP achieves a new state-of-the-art performance on various downstream tasks with much less computational cost. Furthermore, we demonstrate that our model is essentially in par with ALBEF, the previous state-of-the-art, only with one-third of training epochs on the same training data.
• FedX: Unsupervised Federated Learning with Cross Knowledge Distillation​
• Presenter: Sungwon Park, Meeyoung Cha (KAIST) We present FedX, an unsupervised federated learning framework. Our model learns unbiased representation from decentralized and heterogeneous local data. It employs a two-sided knowledge distillation with contrastive learning as a core component, allowing the federated system to function without requiring clients to share any data features. Furthermore, its adaptable architecture can be used as an add-on module for existing unsupervised algorithms in federated settings. Experiments show that our model improves performance significantly (1.58~5.52pp) on five unsupervised algorithms.​
• Advances in Polarimetric Appearance​ of 3D Geometry​
• Presenter: Shinyoung Yi, Min H. Kim (KAIST) Light is an electromagnetic wave which has a polarization information. In computer graphics and vision, polarization has played an important role to reconstruct 3D geometry and material appearance from captured images, and investigating polarimetric appearance itself is also considered as an important problem. In this context, our research group has proposed novel methods to deal with polarimetric appearance. Additionally, we introduces key challenges in polarimetric forward and inverse rendering problems, which have been being solved by our research group.​
• Unsupervised Learning Framework for Fairness-aware Classification
• Presenter: Sungho Park, HYERAN BYUN (Yonsei University) Previous studies for fairness has a limitation that they necessarily require additional sensitive attribute labels. In this paper, we propose a classification method that ensures fairness with respect to all potential sensitive attributes. The core idea is to dynamically cluster data samples based on whether they are being correctly classified or not. By giving more weight to the group with misclassified samples, we encourage a classification model to be trained fairly with respect to all the unknown sensitive attributes on benchmark datasets.

AI INNOVATION HUB

Research Unit