Young‐Hoon Kim

Jaewoo Yang, Hayun Kim, Younghoon Kim

Modern large language models (LLMs) have established state-of-the-art performance through architectural improvements, but still require significant computational cost for inference. In an effort to reduce the inference cost, post-training quantization (PTQ) has become a popular approach, quantizing weights and activations to lower precision, such as INT8. In this paper, we reveal the challenges of activation quantization in GLU variants, which are widely used in feed-forward network (FFN) of modern LLMs, such as LLaMA family. The problem is that severe local quantization errors, caused by excessive magnitudes of activation in GLU variants, significantly degrade the performance of the quantized LLM. We denote these activations as activation spikes. Our further observations provide a systematic pattern of activation spikes: 1) The activation spikes occur in the FFN of specific layers, particularly in the early and late layers, 2) The activation spikes are dedicated to a couple of tokens, rather than being shared across a sequence. Based on our observations, we propose two empirical methods, Quantization-free Module (QFeM) and Quantization-free Prefix (QFeP), to isolate the activation spikes during quantization. Our extensive experiments validate the effectiveness of the proposed methods for the activation quantization, especially with coarse-grained scheme, of latest LLMs with GLU variants, including LLaMA-2/3, Mistral, Mixtral, SOLAR, and Gemma. In particular, our methods enhance the current alleviation techniques (e.g., SmoothQuant) that fail to control the activation spikes. Code is available at https://github.com/onnoo/activation-spikes.

Hoon Kim, Minje Jang, Wonjun Yoon et al.

We introduce a co-designed approach for human portrait relighting that combines a physics-guided architecture with a pre-training framework. Drawing on the Cook-Torrance reflectance model, we have meticulously configured the architecture design to precisely simulate light-surface interactions. Furthermore, to overcome the limitation of scarce high-quality lightstage data, we have developed a self-supervised pre-training strategy. This novel combination of accurate physical modeling and expanded training dataset establishes a new benchmark in relighting realism.

Younghoon Kim, Jeffrey Heer

Complex animated transitions may be easier to understand when divided into separate, consecutive stages. However, effective staging requires careful attention to both animation semantics and timing parameters. We present Gemini2, a system for creating staged animations from a sequence of chart keyframes. Given only a start state and an end state, Gemini2 can automatically recommend intermediate keyframes for designers to consider. The Gemini2 recommendation engine leverages Gemini, our prior work, and GraphScape to itemize the given complex change into semantic edit operations and to recombine operations into stages with a guided order for clearly conveying the semantics. To evaluate Gemini2's recommendations, we conducted a human-subject study in which participants ranked recommended animations from both Gemini2 and Gemini. We find that Gemini2's animation recommendation ranking is well aligned with subjects' preferences, and Gemini2 can recommend favorable animations that Gemini cannot support.

Dongmyung Shin, Younghoon Kim, Chungseok Oh et al.

Carefully engineered radiofrequency (RF) pulses play a key role in a number of systems such as mobile phone, radar, and magnetic resonance imaging. The design of an RF waveform, however, is often posed as an inverse problem with no general solution. As a result, various design methods each with a specific purpose have been developed based on the intuition of human experts. In this work, we propose an artificial intelligence (AI)-powered RF pulse design framework, DeepRF, which utilizes the self-learning characteristics of deep reinforcement learning to generate a novel RF pulse. The effectiveness of DeepRF is demonstrated using four types of RF pulses that are commonly used. The DeepRF-designed pulses successfully satisfy the design criteria while reporting reduced energy. Analyses demonstrate the pulses utilize new mechanisms of magnetization manipulation, suggesting the potentials of DeepRF in discovering unseen design dimensions beyond human intuition. This work may lay the foundation for an emerging field of AI-driven RF waveform design.