arXiv subjects

Yejin Lee

Publications and source records attributed to Yejin Lee.

33 records · Page 2Linked to original sources

arXiv2025

Characterizing and Efficiently Accelerating Multimodal Generation Model Inference

Generative artificial intelligence (AI) technology is revolutionizing the computing industry. Not only its applications have broadened to various sectors but also poses new system design and optimization opportunities. The technology is capable of understanding and responding in multiple modalities. However, the advanced capability currently comes with significant system resource demands. To sustainably scale generative AI capabilities to billions of users in the world, inference must be fast and efficient. This paper pinpoints key system design and optimization opportunities by characterizing a family of emerging multi-modal generation models on real systems. Auto-regressive token generation is a critical latency performance bottleneck, typically dominated by GPU idle time. In addition to memory-intensive attention across the generative AI models, linear operations constitute significant inference latency due to the feed forward networks in Transformer-based models. We demonstrate that state-of-the-art optimization levers, spanning from applications to system software and hardware, set a 3.88x better baseline.

cs.LG↗

arXiv2025

Exploring van der Waals cuprate superconductors using a hybrid microwave circuit

The advent of two-dimensional van der Waals materials is a frontier of condensed matter physics and quantum devices. However, characterizing such materials remains challenging due to the limitations of bulk material techniques, necessitating the development of specialized methods. In this study, we investigate the superconducting properties of Bi2Sr2CaCu2O8+x flakes by integrating them with a hybrid superconducting microwave resonator. The hybrid resonator is significantly modified by the interaction with the flake while maintaining a high quality factor. We also observe a significant upshift of the resonator frequency with increasing temperature, as well as a positive nonlinearity. These effects originate from a presently unknown microscopic mechanism within the flake, and can be modeled as a two-level system bath interacting with the resonant mode. Our findings open a path for high quality hybrid circuits with van der Waals flakes for exploring novel materials and developing new devices for quantum technology.

cond-mat.supr-con↗

arXiv2025

Color Flow Imaging Microscopy Improves Identification of Stress Sources of Protein Aggregates in Biopharmaceuticals

Protein-based therapeutics play a pivotal role in modern medicine targeting various diseases. Despite their therapeutic importance, these products can aggregate and form subvisible particles (SvPs), which can compromise their efficacy and trigger immunological responses, emphasizing the critical need for robust monitoring techniques. Flow Imaging Microscopy (FIM) has been a significant advancement in detecting SvPs, evolving from monochrome to more recently incorporating color imaging. Complementing SvP images obtained via FIM, deep learning techniques have recently been employed successfully for stress source identification of monochrome SvPs. In this study, we explore the potential of color FIM to enhance the characterization of stress sources in SvPs. To achieve this, we curate a new dataset comprising 16,000 SvPs from eight commercial monoclonal antibodies subjected to heat and mechanical stress. Using both supervised and self-supervised convolutional neural networks, as well as vision transformers in large-scale experiments, we demonstrate that deep learning with color FIM images consistently outperforms monochrome images, thus highlighting the potential of color FIM in stress source classification compared to its monochrome counterparts.

cs.CV↗

arXiv2024

Evolution of dissipative regimes in atomically thin $\text{Bi}_{2}\text{Sr}_{2}\text{CaCu}_{2}\text{O}_{8+x}$ superconductor

Thermoelectric transport has been widely used to study Abrikosov vortex dynamics in unconventional superconductors. However, only a few thermoelectric studies have been conducted near the dimensional crossover that occurs when the vortex-vortex interaction length scale becomes comparable to the sample size. Here we report the effects of finite size on the dissipation mechanisms of the Nernst effect in the optimally doped $\text{Bi}_{2}\text{Sr}_{2}\text{CaCu}_{2}\text{O}_{8+x}$ high-temperature superconductor, down to the atomic length limit. To access this regime, we develop a new generation of thermoelectric chips based on silicon nitride microprinted circuit boards. These chips ensure optimized signals while preventing sample deterioration. Our results demonstrate that lateral confinement at the nanoscale can effectively reduce vortex dissipation. Investigating vortex dissipation at the micro- and nano-scale is essential for creating stable, miniaturized superconducting circuits.

cond-mat.supr-con↗

arXiv2024

Real-Time Human Action Recognition on Embedded Platforms

With advancements in computer vision and deep learning, video-based human action recognition (HAR) has become practical. However, due to the complexity of the computation pipeline, running HAR on live video streams incurs excessive delays on embedded platforms. This work tackles the real-time performance challenges of HAR with four contributions: 1) an experimental study identifying a standard Optical Flow (OF) extraction technique as the latency bottleneck in a state-of-the-art HAR pipeline, 2) an exploration of the latency-accuracy tradeoff between the standard and deep learning approaches to OF extraction, which highlights the need for a novel, efficient motion feature extractor, 3) the design of Integrated Motion Feature Extractor (IMFE), a novel single-shot neural network architecture for motion feature extraction with drastic improvement in latency, 4) the development of RT-HARE, a real-time HAR system tailored for embedded platforms. Experimental results on an Nvidia Jetson Xavier NX platform demonstrated that RT-HARE realizes real-time HAR at a video frame rate of 30 frames per second while delivering high levels of recognition accuracy.

cs.CV↗

arXiv2024

Generative AI Beyond LLMs: System Implications of Multi-Modal Generation

As the development of large-scale Generative AI models evolve beyond text (1D) generation to include image (2D) and video (3D) generation, processing spatial and temporal information presents unique challenges to quality, performance, and efficiency. We present the first work towards understanding this new system design space for multi-modal text-to-image (TTI) and text-to-video (TTV) generation models. Current model architecture designs are bifurcated into 2 categories: Diffusion- and Transformer-based models. Our systematic performance characterization on a suite of eight representative TTI/TTV models shows that after state-of-the-art optimization techniques such as Flash Attention are applied, Convolution accounts for up to 44% of execution time for Diffusion-based TTI models, while Linear layers consume up to 49% of execution time for Transformer-based models. We additionally observe that Diffusion-based TTI models resemble the Prefill stage of LLM inference, and benefit from 1.1-2.5x greater speedup from Flash Attention than Transformer-based TTI models that resemble the Decode phase. Since optimizations designed for LLMs do not map directly onto TTI/TTV models, we must conduct a thorough characterization of these workloads to gain insights for new optimization opportunities. In doing so, we define sequence length in the context of TTI/TTV models and observe sequence length can vary up to 4x in Diffusion model inference. We additionally observe temporal aspects of TTV workloads pose unique system bottlenecks, with Temporal Attention accounting for over 60% of total Attention time. Overall, our in-depth system performance characterization is a critical first step towards designing efficient and deployable systems for emerging TTI/TTV workloads.

cs.DC↗

arXiv2024

Is Flash Attention Stable?

Training large-scale machine learning models poses distinct system challenges, given both the size and complexity of today's workloads. Recently, many organizations training state-of-the-art Generative AI models have reported cases of instability during training, often taking the form of loss spikes. Numeric deviation has emerged as a potential cause of this training instability, although quantifying this is especially challenging given the costly nature of training runs. In this work, we develop a principled approach to understanding the effects of numeric deviation, and construct proxies to put observations into context when downstream effects are difficult to quantify. As a case study, we apply this framework to analyze the widely-adopted Flash Attention optimization. We find that Flash Attention sees roughly an order of magnitude more numeric deviation as compared to Baseline Attention at BF16 when measured during an isolated forward pass. We then use a data-driven analysis based on the Wasserstein Distance to provide upper bounds on how this numeric deviation impacts model weights during training, finding that the numerical deviation present in Flash Attention is 2-5 times less significant than low-precision training.

cs.LG↗

arXiv2024

CHAI: Clustered Head Attention for Efficient LLM Inference

Large Language Models (LLMs) with hundreds of billions of parameters have transformed the field of machine learning. However, serving these models at inference time is both compute and memory intensive, where a single request can require multiple GPUs and tens of Gigabytes of memory. Multi-Head Attention is one of the key components of LLMs, which can account for over 50% of LLMs memory and compute requirement. We observe that there is a high amount of redundancy across heads on which tokens they pay attention to. Based on this insight, we propose Clustered Head Attention (CHAI). CHAI combines heads with a high amount of correlation for self-attention at runtime, thus reducing both memory and compute. In our experiments, we show that CHAI is able to reduce the memory requirements for storing K,V cache by up to 21.4% and inference time latency by up to 1.73x without any fine-tuning required. CHAI achieves this with a maximum 3.2% deviation in accuracy across 3 different models (i.e. OPT-66B, LLAMA-7B, LLAMA-33B) and 5 different evaluation datasets.

cs.LG↗

arXiv2024

Twisted cuprate van der Waals heterostructures with controlled Josephson coupling

Twisted van der Waals (vdW) heterostructures offer a unique platform for engineering the efficient Josephson coupling between cuprate thin crystals harboring the nodal superconducting order parameter. Preparing vdW heterostructures-based Josephson junction comprising stacked cuprates requires maintaining an ordered interface with preserved surface superconductivity. Here, we report the preparation of the Josephson junction out of the stacked Bi2Sr2CuCa2O8+d crystals using the cryogenic dry transfer technique and encapsulating the junction with an insulating layer, that protects the interface during the electrical contacts evaporation at the 10-6 mbar base pressure. We find that the Josephson critical current Ic has a maximum at low twist angles, comparable to that of the bulk intrinsic Josephson junctions, and is reduced by two orders of magnitude at twist angles close to 45°. The reduction of Ic occurs due to a mismatch between superconducting d-wave order parameters, which suppresses the direct Cooper pair tunneling.

cond-mat.supr-con↗

arXiv2023

Hall-effect in the MnBi_2Te_4 crystal using silicon nitride nanomembrane via contacts

Utilizing an interplay between band topology and intrinsic magnetism, the two-dimensional van der Waals (vdW) system MnBi_2Te_4 provides an ideal platform for realizing exotic quantum phenomena and offers great opportunities in the emerging field of antiferromagnetic spintronic technology. Yet, the fabrication of MnBi_2Te_4-based nanodevices is hindered by the high sensitivity of this material, which quickly degrades when exposed to air or to elevated temperatures. Here, we demonstrate an alternative route of fabricating vdW-MnBi_2Te_4-based electronic devices using the cryogenic dry transfer of a printable circuit embedded in an inorganic silicon nitride membrane. The electrical connections between the thin crystal and the top surface of the membrane are established through via contacts. Our magnetotransport study reveals that this innovative via contact approach enables exploring the MnBi_2Te_4-like sensitive 2D materials and engineering synthetic heterostructures as well as complex circuits based on the two-dimensional vdW systems.

cond-mat.mtrl-sci↗

arXiv2023

2D high temperature superconductor integration in contact printed circuit boards

Inherent properties of superconducting Bi2Sr2CaCu2O8+x films, such as the high superconducting transition temperature Tc, efficient Josephson coupling between neighboring CuO layers, and fast quasiparticle relaxation dynamics, make them a promising platform for advances in quantum computing and communication technologies. However, preserving two-dimensional superconductivity during device fabrication is an outstanding experimental challenge because of the fast degradation of the superconducting properties of two-dimensional flakes when they are exposed to moisture, organic sol vents, and heat. Here, to realize superconducting devices utilizing two-dimensional superconducting films, we develop a novel fabrication technique relying on the cryogenic dry transfer of printable circuits embedded into a silicon nitride membrane. This approach separates the circuit fabrication stage requiring chemically reactive substances and ionizing physical processes from the creation of the thin superconducting structures. Apart from providing electrical contacts in a single transfer step, the membrane encapsulates the surface of the crystal shielding it from the environment. The fabricated atomically thin Bi2Sr2CaCu2O8+x-based devices show high superconducting transition temperature Tc ~ 91 K close to that of the bulk crystal and demonstrate stable super conducting properties

cond-mat.supr-con↗

arXiv2023

Anomalous Hall effect and magnetoresistance in micro-ribbons of the magnetic Weyl semimetal candidate PrRhC2

PrRhC2 belongs to the rare-earth carbides family whose properties are of special interest among topological semimetals due to the simultaneous breaking of both inversion and time-reversal symmetry. The concomitant absence of both symmetries grants the possibility to tune the Weyl nodes chirality and to enhance topological effects like the chiral anomaly. In this work, we report on the synthesis and compare the magnetotransport measurements of a poly- and single crystalline PrRhC2 sample. Using a remarkable and sophisticated technique, the PrRhC2 single crystal is prepared via focused ion beam cutting from the polycrystalline material. Our magnetometric and specific heat analyses reveal a non-collinear antiferromagnetic state below 20K, as well as short-range magnetic correlations and/or magnetic fluctuations well above the onset of the magnetic transition. The transport measurements on the PrRhC2 single crystal display an electrical resistivity peak at 3K and an anomalous Hall effect below 6K indicative of a net magnetization component in the ordered state. Furthermore, we study the angular variation of magnetoresistivities as a function of the angle between the in-plane magnetic field and the injected electrical current. We find that both the transverse and the longitudinal resistivities exhibit fourfold angular dependencies due to higher-order terms in the resistivity tensor, consistent with the orthorhombic crystal symmetry of PrRhC2. Our experimental results may be interpreted as features of topological Weyl semimetallic behavior in the magnetotransport properties.

cond-mat.mtrl-sci↗

arXiv2023

Few-shot Fine-tuning is All You Need for Source-free Domain Adaptation

Recently, source-free unsupervised domain adaptation (SFUDA) has emerged as a more practical and feasible approach compared to unsupervised domain adaptation (UDA) which assumes that labeled source data are always accessible. However, significant limitations associated with SFUDA approaches are often overlooked, which limits their practicality in real-world applications. These limitations include a lack of principled ways to determine optimal hyperparameters and performance degradation when the unlabeled target data fail to meet certain requirements such as a closed-set and identical label distribution to the source data. All these limitations stem from the fact that SFUDA entirely relies on unlabeled target data. We empirically demonstrate the limitations of existing SFUDA methods in real-world scenarios including out-of-distribution and label distribution shifts in target data, and verify that none of these methods can be safely applied to real-world settings. Based on our experimental results, we claim that fine-tuning a source pretrained model with a few labeled data (e.g., 1- or 3-shot) is a practical and reliable solution to circumvent the limitations of SFUDA. Contrary to common belief, we find that carefully fine-tuned models do not suffer from overfitting even when trained with only a few labeled data, and also show little change in performance due to sampling bias. Our experimental results on various domain adaptation benchmarks demonstrate that the few-shot fine-tuning approach performs comparatively under the standard SFUDA settings, and outperforms comparison methods under realistic scenarios. Our code is available at https://github.com/daintlab/fewshot-SFDA .

cs.CV↗

arXiv2023

Encapsulating high-temperature superconducting twisted van der Waals heterostructures blocks detrimental effects of disorder

High-temperature cuprate superconductors-based van der Waals (vdW) heterostructures hold high technological promise. One of the obstacles hindering progress is the detrimental effect of disorder on the properties of Josephson junctions (JJ) realized by the vdW devices. Here we report the new method of fabricating twisted vdW heterostructures made of Bi2Sr2CuCa2O8+d, crucially improving the JJ characteristics, pushing them up to those of the intrinsic JJs in bulk samples. The method combines a cryogenic stacking using a solvent-free stencil mask technique and covering interface by the insulating hexagonal boron nitride crystals. Despite the high-vacuum condition down to 10-6 mbar in the evaporation chamber, the interface appears to be protected from water molecules during the in-situ metal deposition only when fully encapsulated. Comparing the current-voltage curves of encapsulated and unencapsulated interfaces, we reveal that the encapsulated interfaces' characteristics are crucially improved so that the corresponding JJs demonstrate high critical currents and sharpness of the superconducting transition comparable to those of the intrinsic JJs. Finally, we show that the encapsulated heterostructures are more stable in time.

cond-mat.supr-con↗

arXiv2023

Probing magnetic properties at the nanoscale: In-situ Hall measurements in a TEM

We report on advanced in-situ magneto-transport measurements in a transmission electron microscope. The approach allows for concurrent magnetic imaging and high resolution structural and chemical characterization of the same sample. Proof-of-principle in-situ Hall measurements on presumably undemanding nickel thin films supported by micromagnetic simulations reveal that in samples with non-trivial structures and/or compositions, detailed knowledge of the latter is indispensable for a thorough understanding and reliable interpretation of the magneto-transport data. The proposed in-situ approach is thus expected to contribute to a better understanding of the Hall signatures in more complex magnetic textures.

cond-mat.other↗