Jieun Choi

Jaesung Yoo, Sunghyuk Choi, Ye Seul Yang et al.

When a deep learning model is trained sequentially on different datasets, it often forgets the knowledge learned from previous data, a problem known as catastrophic forgetting. This damages the model's performance on diverse datasets, which is critical in privacy-preserving deep learning (PPDL) applications based on transfer learning (TL). To overcome this, we introduce "review learning" (RevL), a low cost continual learning algorithm for diagnosis prediction using electronic health records (EHR) within a PPDL framework. RevL generates data samples from the model which are used to review knowledge from previous datasets. Six simulated institutional experiments and one real-world experiment involving three medical institutions were conducted to validate RevL, using three binary classification EHR data. In the real-world experiment with data from 106,508 patients, the mean global area under the receiver operating curve was 0.710 for RevL and 0.655 for TL. These results demonstrate RevL's ability to retain previously learned knowledge and its effectiveness in real-world PPDL scenarios. Our work establishes a realistic pipeline for PPDL research based on model transfers across institutions and highlights the practicality of continual learning in real-world medical settings using private EHR data.

Kyungmin Bin, Seungbeom Choi, Jimyoung Son et al.

Recent advances in Post-Training Quantization (PTQ) techniques have significantly increased demand for serving quantized large language models (LLMs), enabling higher throughput and substantially reduced memory usage with minimal accuracy loss. Quantized models address memory constraints in LLMs and enhance GPU resource utilization through efficient GPU sharing. However, quantized models have smaller KV block sizes than non-quantized models, causing limited memory efficiency due to memory fragmentation. Also, distinct resource usage patterns between quantized and non-quantized models require efficient scheduling to maximize throughput. To address these challenges, we propose FineServe, an inference serving framework for mixed-precision LLMs. FineServe's key contributions include: (1) KV Slab, a precision-aware adaptive memory management technique dynamically allocating KV cache based on model quantization characteristics, significantly reducing GPU memory fragmentation, and (2) a two-level scheduling framework comprising a global scheduler that places models to GPUs based on request rates, latency SLOs, and memory constraints and efficiency, and a local scheduler that adaptively adjusts batch sizes according to real-time request fluctuations. Experimental results demonstrate that FineServe achieves up to 2.2x higher SLO attainment and 1.8x higher token generation throughput compared to the state-of-the-art GPU sharing systems.

Taehyeon Eum, Jieun Choi, Tae-Kyun Kim

Diffusion-based methods have achieved significant successes in T2I generation, providing realistic images from text prompts. Despite their capabilities, these models face persistent challenges in generating realistic human hands, often producing images with incorrect finger counts and structurally deformed hands. MGHanD addresses this challenge by applying multi-modal guidance during the inference process. For visual guidance, we employ a discriminator trained on a dataset comprising paired real and generated images with captions, derived from various hand-in-the-wild datasets. We also employ textual guidance with LoRA adapter, which learns the direction from `hands' towards more detailed prompts such as `natural hands', and `anatomically correct fingers' at the latent level. A cumulative hand mask which is gradually enlarged in the assigned time step is applied to the added guidance, allowing the hand to be refined while maintaining the rich generative capabilities of the pre-trained model. In the experiments, our method achieves superior hand generation qualities, without any specific conditions or priors. We carry out both quantitative and qualitative evaluations, along with user studies, to showcase the benefits of our approach in producing high-quality hand images.

Daehyeon Baek, Jieun Choi, Jimyoung Son et al.

As large language models become increasingly prevalent, memory bandwidth constraints significantly limit inference throughput, motivating post-training quantization (PTQ). In this paper, we propose FireQ, a co-designed PTQ framework and an INT4-FP8 matrix multiplication kernel that accelerates LLM inference across all linear layers. Specifically, FireQ quantizes linear layer weights and key-values to INT4, and activations and queries to FP8, significantly enhancing throughput. Additionally, we introduce a three-stage pipelining for the prefill phase, which modifies the FlashAttention-3 kernel, effectively reducing time-to-first-token in the prefill phase. To minimize accuracy loss from quantization, we develop novel outlier smoothing techniques tailored separately for linear and attention layers. In linear layers, we explicitly use per-tensor scaling to prevent underflow caused by the FP8 quantization scaling factor of INT4 quantization, and channel-wise scaling to compensate for coarse granularity of INT4. In attention layers, we address quantization challenges posed by rotary positional embeddings (RoPE) by combining pre-RoPE and post-RoPE scaling strategies. FireQ significantly outperforms state-of-the-art methods, achieving 1.68x faster inference in feed-forward network layers on Llama2-7B and 1.26x faster prefill phase performance on Llama3-8B compared to QServe, with negligible accuracy loss.

Cheongjae Jang, Sang-Kyun Ko, Yung-Kyun Noh et al.

The $\text{t}\bar{\text{t}}\text{H}(\text{b}\bar{\text{b}})$ process is an essential channel to reveal the Higgs properties but has an irreducible background from the $\text{t}\bar{\text{t}}\text{b}\bar{\text{b}}$ process, which produces a top quark pair in association with a b quark pair. Therefore, understanding the $\text{t}\bar{\text{t}}\text{b}\bar{\text{b}}$ process is crucial for improving the sensitivity of a search for the $\text{t}\bar{\text{t}}\text{H}(\text{b}\bar{\text{b}})$ process. To this end, when measuring the differential cross-section of the $\text{t}\bar{\text{t}}\text{b}\bar{\text{b}}$ process, we need to distinguish the b-jets originated from top quark decays, and additional b-jets originated from gluon splitting. Since there are no simple identification rules, we adopt deep learning methods to learn from data to identify the additional b-jets from the $\text{t}\bar{\text{t}}\text{b}\bar{\text{b}}$ events. Specifically, by exploiting the special structure of the $\text{t}\bar{\text{t}}\text{b}\bar{\text{b}}$ event data, we propose several loss functions that can be minimized to directly increase the matching efficiency, the accuracy of identifying additional b-jets. We discuss the difference between our method and another deep learning-based approach based on binary classification arXiv:1910.14535 using synthetic data. We then verify that additional b-jets can be identified more accurately by increasing matching efficiency directly rather than the binary classification accuracy, using simulated $\text{t}\bar{\text{t}}\text{b}\bar{\text{b}}$ event data in the lepton+jets channel from pp collision at $\sqrt{s}$ = 13 TeV.