Showing 1–26 of 26 results for author: Tsai, E

ELFS: Enhancing Label-Free Coreset Selection via Clustering-based Pseudo-Labeling

Authors: Haizhong Zheng, Elisa Tsai, Yifu Lu, Jiachen Sun, Brian R. Bartoldson, Bhavya Kailkhura, Atul Prakash

Abstract: High-quality human-annotated data is crucial for modern deep learning pipelines, yet the human annotation process is both costly and time-consuming. Given a constrained human labeling budget, selecting an informative and representative data subset for labeling can significantly reduce human annotation effort. Well-performing state-of-the-art (SOTA) coreset selection methods require ground-truth la… ▽ More High-quality human-annotated data is crucial for modern deep learning pipelines, yet the human annotation process is both costly and time-consuming. Given a constrained human labeling budget, selecting an informative and representative data subset for labeling can significantly reduce human annotation effort. Well-performing state-of-the-art (SOTA) coreset selection methods require ground-truth labels over the whole dataset, failing to reduce the human labeling burden. Meanwhile, SOTA label-free coreset selection methods deliver inferior performance due to poor geometry-based scores. In this paper, we introduce ELFS, a novel label-free coreset selection method. ELFS employs deep clustering to estimate data difficulty scores without ground-truth labels. Furthermore, ELFS uses a simple but effective double-end pruning method to mitigate bias on calculated scores, which further improves the performance on selected coresets. We evaluate ELFS on five vision benchmarks and show that ELFS consistently outperforms SOTA label-free baselines. For instance, at a 90% pruning rate, ELFS surpasses the best-performing baseline by 5.3% on CIFAR10 and 7.1% on CIFAR100. Moreover, ELFS even achieves comparable performance to supervised coreset selection at low pruning rates (e.g., 30% and 50%) on CIFAR10 and ImageNet-1K. △ Less

Submitted 6 June, 2024; originally announced June 2024.

CheXagent: Towards a Foundation Model for Chest X-Ray Interpretation

Authors: Zhihong Chen, Maya Varma, Jean-Benoit Delbrouck, Magdalini Paschali, Louis Blankemeier, Dave Van Veen, Jeya Maria Jose Valanarasu, Alaa Youssef, Joseph Paul Cohen, Eduardo Pontes Reis, Emily B. Tsai, Andrew Johnston, Cameron Olsen, Tanishq Mathew Abraham, Sergios Gatidis, Akshay S. Chaudhari, Curtis Langlotz

Abstract: Chest X-rays (CXRs) are the most frequently performed imaging test in clinical practice. Recent advances in the development of vision-language foundation models (FMs) give rise to the possibility of performing automated CXR interpretation, which can assist physicians with clinical decision-making and improve patient outcomes. However, develo** FMs that can accurately interpret CXRs is challengin… ▽ More Chest X-rays (CXRs) are the most frequently performed imaging test in clinical practice. Recent advances in the development of vision-language foundation models (FMs) give rise to the possibility of performing automated CXR interpretation, which can assist physicians with clinical decision-making and improve patient outcomes. However, develo** FMs that can accurately interpret CXRs is challenging due to the (1) limited availability of large-scale vision-language datasets in the medical image domain, (2) lack of vision and language encoders that can capture the complexities of medical data, and (3) absence of evaluation frameworks for benchmarking the abilities of FMs on CXR interpretation. In this work, we address these challenges by first introducing \emph{CheXinstruct} - a large-scale instruction-tuning dataset curated from 28 publicly-available datasets. We then present \emph{CheXagent} - an instruction-tuned FM capable of analyzing and summarizing CXRs. To build CheXagent, we design a clinical large language model (LLM) for parsing radiology reports, a vision encoder for representing CXR images, and a network to bridge the vision and language modalities. Finally, we introduce \emph{CheXbench} - a novel benchmark designed to systematically evaluate FMs across 8 clinically-relevant CXR interpretation tasks. Extensive quantitative evaluations and qualitative reviews with five expert radiologists demonstrate that CheXagent outperforms previously-developed general- and medical-domain FMs on CheXbench tasks. Furthermore, in an effort to improve model transparency, we perform a fairness evaluation across factors of sex, race and age to highlight potential performance disparities. Our project is at \url{https://stanford-aimi.github.io/chexagent.html}. △ Less

Submitted 22 January, 2024; originally announced January 2024.

Comments: 24 pages, 8 figures

Virtual Scientific Companion for Synchrotron Beamlines: A Prototype

Authors: Daniel Potemkin, Carlos Soto, Ruipeng Li, Kevin Yager, Esther Tsai

Abstract: The extraordinarily high X-ray flux and specialized instrumentation at synchrotron beamlines have enabled versatile in-situ and high throughput studies that are impossible elsewhere. Dexterous and efficient control of experiments are thus crucial for efficient beamline operation. Artificial intelligence and machine learning methods are constantly being developed to enhance facility performance, bu… ▽ More The extraordinarily high X-ray flux and specialized instrumentation at synchrotron beamlines have enabled versatile in-situ and high throughput studies that are impossible elsewhere. Dexterous and efficient control of experiments are thus crucial for efficient beamline operation. Artificial intelligence and machine learning methods are constantly being developed to enhance facility performance, but the full potential of these developments can only be reached with efficient human-computer-interaction. Natural language is the most intuitive and efficient way for humans to communicate. However, the low credibility and reproducibility of existing large language models and tools demand extensive development to be made for robust and reliable performance for scientific purposes. In this work, we introduce the prototype of virtual scientific companion (VISION) and demonstrate that it is possible to control basic beamline operations through natural language with open-source language model and the limited computational resources at beamline. The human-AI nature of VISION leverages existing automation systems and data framework at synchrotron beamlines. △ Less

Submitted 28 December, 2023; originally announced December 2023.

Statistical Characteristics of the Electron Isotropy Boundary

Authors: Colin Wilkins, Vassilis Angelopoulos, Andrei Runov, Anton Artemyev, Xiao-Jia Zhang, Jiang Liu, Ethan Tsai

Abstract: Utilizing observations from the ELFIN satellites, we present a statistical study of $\sim$2000 events in 2019-2020 characterizing the occurrence in magnetic local time (MLT) and latitude of $\geq$50 keV electron isotropy boundaries (IBs) at Earth, and the dependence of associated precipitation on geomagnetic activity. The isotropy boundary for an electron of a given energy is the magnetic latitude… ▽ More Utilizing observations from the ELFIN satellites, we present a statistical study of $\sim$2000 events in 2019-2020 characterizing the occurrence in magnetic local time (MLT) and latitude of $\geq$50 keV electron isotropy boundaries (IBs) at Earth, and the dependence of associated precipitation on geomagnetic activity. The isotropy boundary for an electron of a given energy is the magnetic latitude poleward of which persistent isotropized pitch-angle distributions ($J_{prec}/J_{perp}\sim 1$) are first observed to occur, interpreted as resulting from magnetic field-line curvature scattering (FLCS) in the equatorial magnetosphere. We find that energetic electron IBs can be well-recognized on the nightside from dusk until dawn, under all geomagnetic activity conditions, with a peak occurrence rate of almost 90% near $\sim$22 hours in MLT, remaining above 80% from 21 to 01 MLT. The IBs span a wide range of IGRF magnetic latitudes from $60^\circ$-$74^\circ$, with a maximum occurrence between $66^\circ$-$71^\circ$ (L of 6-8), shifting to lower latitudes and pre-midnight local times with activity. The precipitating energy flux of $\geq$50 keV electrons averaged over the IB-associated latitudes varies over four orders of magnitude, up to $\sim$1 erg/cm$^2$-s, and often includes electron energies exceeding 1 MeV. The local time distribution of IB-associated energies and precipitating fluxes also exhibit peak values near midnight for low activity, shifting toward pre-midnight for elevated activity. The percentage of the total energy deposited over the high-latitude regions ($55^\circ$ to $80^\circ$; or IGRF $L\gtrsim 3$) attributed to IBs is 10-20%, on average, or about 10 MW of total atmospheric power input, but at times can be up to $\sim$100% of the total $\geq$50 keV electron energy deposition over the entire sub-auroral and auroral zone region, exceeding 1 GW in atmospheric power input. △ Less

Submitted 25 May, 2023; originally announced May 2023.

CERTainty: Detecting DNS Manipulation at Scale using TLS Certificates

Authors: Elisa Tsai, Deepak Kumar, Ram Sundara Raman, Gavin Li, Yael Eiger, Roya Ensafi

Abstract: DNS manipulation is an increasingly common technique used by censors and other network adversaries to prevent users from accessing restricted Internet resources and hijack their connections. Prior work in detecting DNS manipulation relies largely on comparing DNS resolutions with trusted control results to identify inconsistencies. However, the emergence of CDNs and other cloud providers practicin… ▽ More DNS manipulation is an increasingly common technique used by censors and other network adversaries to prevent users from accessing restricted Internet resources and hijack their connections. Prior work in detecting DNS manipulation relies largely on comparing DNS resolutions with trusted control results to identify inconsistencies. However, the emergence of CDNs and other cloud providers practicing content localization and load balancing leads to these heuristics being inaccurate, paving the need for more verifiable signals of DNS manipulation. In this paper, we develop a new technique, CERTainty, that utilizes the widely established TLS certificate ecosystem to accurately detect DNS manipulation, and obtain more information about the adversaries performing such manipulation. We find that untrusted certificates, mismatching hostnames, and blockpages are powerful proxies for detecting DNS manipulation. Our results show that previous work using consistency-based heuristics is inaccurate, allowing for 72.45% false positives in the cases detected as DNS manipulation. Further, we identify 17 commercial DNS filtering products in 52 countries, including products such as SafeDNS, SkyDNS, and Fortinet, and identify the presence of 55 ASes in 26 countries that perform ISP-level DNS manipulation. We also identify 226 new blockpage clusters that are not covered by previous research. We are integrating techniques used by CERTainty into active measurement platforms to continuously and accurately monitor DNS manipulation. △ Less

Submitted 14 May, 2023; originally announced May 2023.

Comments: To Appear in: Privacy Enhancing Technologies Symposium (PETS), July 2023

Temporal Egonet Subgraph Transitions

Authors: Daniel Gonzalez Cedre, Sophia Abraham, Lucas Parzianello, Eric Tsai

Abstract: How do we summarize dynamic behavioral interactions? We introduce a possible node-embedding-based solution to this question: temporal egonet subgraph transitions. How do we summarize dynamic behavioral interactions? We introduce a possible node-embedding-based solution to this question: temporal egonet subgraph transitions. △ Less

Submitted 26 March, 2023; originally announced March 2023.

Temporal Scales of Electron Precipitation Driven by Whistler-Mode Waves

Authors: Xiao-Jia Zhang, Vassilis Angelopoulos, Anton Artemyev, Didier Mourenas, Oleksiy Agapitov, Ethan Tsai, Colin Wilkins

Abstract: Electron resonant scattering by whistler-mode waves is one of the most important mechanisms responsible for electron precipitation to the Earth's atmosphere. We investigate temporal and spatial scales of such precipitation with measurements from the two low-altitude ELFIN CubeSats. We compare the variations in energetic electron precipitation at the same L-shells but on successive data collection… ▽ More Electron resonant scattering by whistler-mode waves is one of the most important mechanisms responsible for electron precipitation to the Earth's atmosphere. We investigate temporal and spatial scales of such precipitation with measurements from the two low-altitude ELFIN CubeSats. We compare the variations in energetic electron precipitation at the same L-shells but on successive data collection orbit tracks by the two ELFIN satellites. Variations seen at the smallest inter-satellite separations are likely associated with whistler-mode chorus elements or with the scale of chorus wave packets (0.1 - 1 s in time and 100 km in space at the equator). Variations between precipitation L-shell profiles at intermediate inter-satellite separations are likely associated with whistler-mode wave power modulations by ultra-low frequency (ULF) waves, i.e., with the wave source region (from a few to tens of seconds to a few minutes in time and 1000km in space at the equator). During these two types of variations, consecutive crossings are associated with precipitation L-shell profiles very similar to each other. Therefore the spatial and temporal variations at those scales do not change the net electron loss from the outer radiation belt. Variations at the largest range of inter-satellite separations, several minutes to more than 10 min, are likely associated with mesoscale equatorial plasma structures that are affected by convection (at minutes to tens of minutes temporal variations and [1000,10000]km spatial scales). The latter type of variations results in appreciable changes in the precipitation L-shell profiles and can significantly modify the net electron losses during successive tracks. △ Less

Submitted 29 November, 2022; originally announced November 2022.

Energetic electron precipitation driven by electromagnetic ion cyclotron waves from ELFIN's low altitude perspective

Authors: V. Angelopoulos, X. -J. Zhang, A. V. Artemyev, D. Mourenas, E. Tsai, C. Wilkins, A. Runov, J. Liu, D. L. Turner, W. Li, K. Khurana, R. E. Wirz, V. A. Sergeev, X. Meng, J. Wu, M. D. Hartinger, T. Raita, Y. Shen, X. An, X. Shi, M. F. Bashir, X. Shen, L. Gan, M. Qin, L. Capannolo , et al. (61 additional authors not shown)

Abstract: We review comprehensive observations of electromagnetic ion cyclotron (EMIC) wave-driven energetic electron precipitation using data from the energetic electron detector on the Electron Losses and Fields InvestigatioN (ELFIN) mission, two polar-orbiting low-altitude spinning CubeSats, measuring 50-5000 keV electrons with good pitch-angle and energy resolution. EMIC wave-driven precipitation exhibi… ▽ More We review comprehensive observations of electromagnetic ion cyclotron (EMIC) wave-driven energetic electron precipitation using data from the energetic electron detector on the Electron Losses and Fields InvestigatioN (ELFIN) mission, two polar-orbiting low-altitude spinning CubeSats, measuring 50-5000 keV electrons with good pitch-angle and energy resolution. EMIC wave-driven precipitation exhibits a distinct signature in energy-spectrograms of the precipitating-to-trapped flux ratio: peaks at 0.5 MeV which are abrupt (bursty) with significant substructure (occasionally down to sub-second timescale). Multiple ELFIN passes over the same MLT sector allow us to study the spatial and temporal evolution of the EMIC wave - electron interaction region. Using two years of ELFIN data, we assemble a statistical database of 50 events of strong EMIC wave-driven precipitation. Most reside at L=5-7 at dusk, while a smaller subset exists at L=8-12 at post-midnight. The energies of the peak-precipitation ratio and of the half-peak precipitation ratio (our proxy for the minimum resonance energy) exhibit an L-shell dependence in good agreement with theoretical estimates based on prior statistical observations of EMIC wave power spectra. The precipitation ratio's spectral shape for the most intense events has an exponential falloff away from the peak (i.e., on either side of 1.45 MeV). It too agrees well with quasi-linear diffusion theory based on prior statistics of wave spectra. Sub-MeV electron precipitation observed concurrently with strong EMIC wave-driven 1MeV precipitation has a spectral shape that is consistent with efficient pitch-angle scattering down to 200-300 keV by much less intense higher frequency EMIC waves. These results confirm the critical role of EMIC waves in driving relativistic electron losses. Nonlinear effects may abound and require further investigation. △ Less

Submitted 28 November, 2022; originally announced November 2022.

Improving the Factual Correctness of Radiology Report Generation with Semantic Rewards

Authors: Jean-Benoit Delbrouck, Pierre Chambon, Christian Bluethgen, Emily Tsai, Omar Almusa, Curtis P. Langlotz

Abstract: Neural image-to-text radiology report generation systems offer the potential to improve radiology reporting by reducing the repetitive process of report drafting and identifying possible medical errors. These systems have achieved promising performance as measured by widely used NLG metrics such as BLEU and CIDEr. However, the current systems face important limitations. First, they present an incr… ▽ More Neural image-to-text radiology report generation systems offer the potential to improve radiology reporting by reducing the repetitive process of report drafting and identifying possible medical errors. These systems have achieved promising performance as measured by widely used NLG metrics such as BLEU and CIDEr. However, the current systems face important limitations. First, they present an increased complexity in architecture that offers only marginal improvements on NLG metrics. Secondly, these systems that achieve high performance on these metrics are not always factually complete or consistent due to both inadequate training and evaluation. Recent studies have shown the systems can be substantially improved by using new methods encouraging 1) the generation of domain entities consistent with the reference and 2) describing these entities in inferentially consistent ways. So far, these methods rely on weakly-supervised approaches (rule-based) and named entity recognition systems that are not specific to the chest X-ray domain. To overcome this limitation, we propose a new method, the RadGraph reward, to further improve the factual completeness and correctness of generated radiology reports. More precisely, we leverage the RadGraph dataset containing annotated chest X-ray reports with entities and relations between entities. On two open radiology report datasets, our system substantially improves the scores up to 14.2% and 25.3% on metrics evaluating the factual correctness and completeness of reports. △ Less

Submitted 21 October, 2022; originally announced October 2022.

Comments: Findings of EMNLP 2022

Tens to hundreds of keV electron precipitation driven by kinetic Alfvén waves during an electron injection

Authors: Y. Shen, A. V. Artemyev, X. -J. Zhang, V. Angelopoulos, I. Vasko, D. Turner, E. Tsai, C. Wilkins, J. Weygand, C. T. Russell, R. E. Ergun, B. L. Giles

Abstract: Electron injections are critical processes associated with magnetospheric substorms, which deposit significant electron energy into the ionosphere. Although wave scattering of $<$10 keV electrons during injections has been well studied, the link between magnetotail electron injections and energetic ($\geq$100 keV) electron precipitation remains elusive. Using conjugate observations between the ELF… ▽ More Electron injections are critical processes associated with magnetospheric substorms, which deposit significant electron energy into the ionosphere. Although wave scattering of $<$10 keV electrons during injections has been well studied, the link between magnetotail electron injections and energetic ($\geq$100 keV) electron precipitation remains elusive. Using conjugate observations between the ELFIN and Magnetospheric Multiscale (MMS) missions, we present evidence of tens to hundreds of keV electron precipitation to the ionosphere potentially driven by kinetic Alfvén waves (KAWs) associated with magnetotail electron injections and magnetic field gradients. Test particle simulations adapted to observations show that dipolarization-front magnetic field gradients and associated $\nabla B$ drifts allow Doppler-shifted Landau resonances between the injected electrons and KAWs, producing electron spatial scattering across the front which results in pitch-angle decreases and subsequent precipitation. Test particle results show that such KAW-driven precipitation can account for ELFIN observations below $\sim$300 keV. △ Less

Submitted 18 July, 2022; originally announced July 2022.

Comments: 25 pages, 5 figures, with supporting information, the manuscript has been accepted for publication by JGR space physics

Relativistic electron precipitation by EMIC waves: importance of nonlinear resonant effects

Authors: Veronika S. Grach, Anton V. Artemyev, Andrei G. Demekhov, Xiao-Jia Zhang, Jacob Bortnik, Vassilis Angelopoulos, R. Nakamura, E. Tsai, C. Wilkins, O. W. Roberts

Abstract: Relativistic electron losses in Earth's radiation belts are usually attributed to electron resonant scattering by electromagnetic waves. One of the most important wave mode for such scattering is the electromagnetic ion cyclotron (EMIC) mode. Within the quasi-linear diffusion framework, the cyclotron resonance of relativistic electrons with EMIC waves results in very fast electron precipitation to… ▽ More Relativistic electron losses in Earth's radiation belts are usually attributed to electron resonant scattering by electromagnetic waves. One of the most important wave mode for such scattering is the electromagnetic ion cyclotron (EMIC) mode. Within the quasi-linear diffusion framework, the cyclotron resonance of relativistic electrons with EMIC waves results in very fast electron precipitation to the atmosphere. However, wave intensities often exceed the threshold for nonlinear resonant interaction, and such intense EMIC waves have been shown to transport electrons away from the loss cone due to the force bunching effect. In this study we investigate if this transport can block electron precipitation. We combine test particle simulations, low-altitude ELFIN observations of EMIC-driven electron precipitation, and ground-based EMIC observations. Comparing simulations and observations, we show that, despite of the low pitch-angle electrons being transported away from the loss cone, the scattering at higher pitch angles results in the loss cone filling and electron precipitation. △ Less

Submitted 1 May, 2022; originally announced May 2022.

PAM: Pose Attention Module for Pose-Invariant Face Recognition

Authors: En-Jung Tsai, Wei-Chang Yeh

Abstract: Pose variation is one of the key challenges in face recognition. Conventional techniques mainly focus on face frontalization or face augmentation in image space. However, transforming face images in image space is not guaranteed to preserve the lossless identity features of the original image. Moreover, these methods suffer from more computational costs and memory requirements due to the additiona… ▽ More Pose variation is one of the key challenges in face recognition. Conventional techniques mainly focus on face frontalization or face augmentation in image space. However, transforming face images in image space is not guaranteed to preserve the lossless identity features of the original image. Moreover, these methods suffer from more computational costs and memory requirements due to the additional models. We argue that it is more desirable to perform feature transformation in hierarchical feature space rather than image space, which can take advantage of different feature levels and benefit from joint learning with representation learning. To this end, we propose a lightweight and easy-to-implement attention block, named Pose Attention Module (PAM), for pose-invariant face recognition. Specifically, PAM performs frontal-profile feature transformation in hierarchical feature space by learning residuals between pose variations with a soft gate mechanism. We validated the effectiveness of PAM block design through extensive ablation studies and verified the performance on several popular benchmarks, including LFW, CFP-FP, AgeDB-30, CPLFW, and CALFW. Experimental results show that our method not only outperforms state-of-the-art methods but also effectively reduces memory requirements by more than 75 times. It is noteworthy that our method is not limited to face recognition with large pose variations. By adjusting the soft gate mechanism of PAM to a specific coefficient, such semantic attention block can easily extend to address other intra-class imbalance problems in face recognition, including large variations in age, illumination, expression, etc. △ Less

Submitted 23 November, 2021; originally announced November 2021.

Comments: 10 pages, 5 figures

Thermal history of matrix forsterite grains from Murchison based on high-resolution tomography

Authors: Giulia Perotti, Henning O. Sørensen, Henning Haack, Anja C. Andersen, Dario Ferreira Sanchez, Elishevah M. M. E. van Kooten, Esther H. R. Tsai, Kim N. Dalby, Mirko Holler, Daniel Grolimund, Tue Hassenkam

Abstract: Protoplanetary disks are dust- and gas-rich structures surrounding protostars. Depending on the distance from the protostar, this dust is thermally processed to different degrees and accreted to form bodies of varying chemical compositions. The primordial accretion processes occurring in the early protoplanetary disk such as chondrule formation and metal segregation are not well understood. One wa… ▽ More Protoplanetary disks are dust- and gas-rich structures surrounding protostars. Depending on the distance from the protostar, this dust is thermally processed to different degrees and accreted to form bodies of varying chemical compositions. The primordial accretion processes occurring in the early protoplanetary disk such as chondrule formation and metal segregation are not well understood. One way to constrain them is to study the morphology and composition of forsteritic grains from the matrix of carbonaceous chondrites. Here, we present high-resolution ptychographic X-ray nanotomography and multimodal chemical micro-tomography (X-ray diffraction and X-ray fluorescence) to reveal the early history of forsteritic grains extracted from the matrix of the Murchison CM2.5 chondrite. The 3D electron density maps revealed, at unprecedented resolution (64~nm), spherical inclusions containing Fe-Ni, very little silica-rich glass and void caps (i.e., volumes where the electron density is consistent with conditions close to vacuum) trapped in forsterite. The presence of the voids along with the overall composition, petrological textures, and shrinkage calculations is consistent with the grains experiencing one or more heating events with peak temperatures close to the melting point of forsterite ($\sim$2100~K) and subsequently cooled and contracted, in agreement with chondrule-forming conditions. △ Less

Submitted 14 September, 2021; originally announced September 2021.

Comments: 15 pages, 11 figures, accepted for publication in ApJ

Amazon SageMaker Clarify: Machine Learning Bias Detection and Explainability in the Cloud

Authors: Michaela Hardt, Xiaoguang Chen, Xiaoyi Cheng, Michele Donini, Jason Gelman, Satish Gollaprolu, John He, Pedro Larroy, Xinyu Liu, Nick McCarthy, Ashish Rathi, Scott Rees, Ankit Siva, ErhYuan Tsai, Keerthan Vasist, Pinar Yilmaz, Muhammad Bilal Zafar, Sanjiv Das, Kevin Haas, Tyler Hill, Krishnaram Kenthapadi

Abstract: Understanding the predictions made by machine learning (ML) models and their potential biases remains a challenging and labor-intensive task that depends on the application, the dataset, and the specific model. We present Amazon SageMaker Clarify, an explainability feature for Amazon SageMaker that launched in December 2020, providing insights into data and ML models by identifying biases and expl… ▽ More Understanding the predictions made by machine learning (ML) models and their potential biases remains a challenging and labor-intensive task that depends on the application, the dataset, and the specific model. We present Amazon SageMaker Clarify, an explainability feature for Amazon SageMaker that launched in December 2020, providing insights into data and ML models by identifying biases and explaining predictions. It is deeply integrated into Amazon SageMaker, a fully managed service that enables data scientists and developers to build, train, and deploy ML models at any scale. Clarify supports bias detection and feature importance computation across the ML lifecycle, during data preparation, model evaluation, and post-deployment monitoring. We outline the desiderata derived from customer input, the modular architecture, and the methodology for bias and explanation computations. Further, we describe the technical challenges encountered and the tradeoffs we had to make. For illustration, we discuss two customer use cases. We present our deployment results including qualitative customer feedback and a quantitative evaluation. Finally, we summarize lessons learned, and discuss best practices for the successful adoption of fairness and explanation tools in practice. △ Less

Submitted 7 September, 2021; originally announced September 2021.

Journal ref: In Proc. ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD), 2974-2983 (2021)

Improving Factual Completeness and Consistency of Image-to-Text Radiology Report Generation

Authors: Yasuhide Miura, Yuhao Zhang, Emily Bao Tsai, Curtis P. Langlotz, Dan Jurafsky

Abstract: Neural image-to-text radiology report generation systems offer the potential to improve radiology reporting by reducing the repetitive process of report drafting and identifying possible medical errors. However, existing report generation systems, despite achieving high performances on natural language generation metrics such as CIDEr or BLEU, still suffer from incomplete and inconsistent generati… ▽ More Neural image-to-text radiology report generation systems offer the potential to improve radiology reporting by reducing the repetitive process of report drafting and identifying possible medical errors. However, existing report generation systems, despite achieving high performances on natural language generation metrics such as CIDEr or BLEU, still suffer from incomplete and inconsistent generations. Here we introduce two new simple rewards to encourage the generation of factually complete and consistent radiology reports: one that encourages the system to generate radiology domain entities consistent with the reference, and one that uses natural language inference to encourage these entities to be described in inferentially consistent ways. We combine these with the novel use of an existing semantic equivalence metric (BERTScore). We further propose a report generation system that optimizes these rewards via reinforcement learning. On two open radiology report datasets, our system substantially improved the F1 score of a clinical information extraction performance by +22.1 (Delta +63.9%). We further show via a human evaluation and a qualitative analysis that our system leads to generations that are more factually complete and consistent compared to the baselines. △ Less

Submitted 12 April, 2021; v1 submitted 20 October, 2020; originally announced October 2020.

Comments: Accepted to NAACL-HLT 2021

The ELFIN Mission

Authors: V. Angelopoulos, E. Tsai, L. Bingley, C. Shaffer, D. L. Turner, A. Runov, W. Li, J. Liu, A. V. Artemyev, X. -J. Zhang, R. J. Strangeway, R. E. Wirz, Y. Y. Shprits, V. A. Sergeev, R. P. Caron, M. Chung, P. Cruce, W. Greer, E. Grimes, K. Hector, M. J. Lawson, D. Leneman, E. V. Masongsong, C. L. Russell, C. Wilkins , et al. (57 additional authors not shown)

Abstract: The Electron Loss and Fields Investigation with a Spatio-Temporal Ambiguity-Resolving option (ELFIN-STAR, or simply: ELFIN) mission comprises two identical 3-Unit (3U) CubeSats on a polar (~93deg inclination), nearly circular, low-Earth (~450 km altitude) orbit. Launched on September 15, 2018, ELFIN is expected to have a >2.5 year lifetime. Its primary science objective is to resolve the mechanism… ▽ More The Electron Loss and Fields Investigation with a Spatio-Temporal Ambiguity-Resolving option (ELFIN-STAR, or simply: ELFIN) mission comprises two identical 3-Unit (3U) CubeSats on a polar (~93deg inclination), nearly circular, low-Earth (~450 km altitude) orbit. Launched on September 15, 2018, ELFIN is expected to have a >2.5 year lifetime. Its primary science objective is to resolve the mechanism of storm-time relativistic electron precipitation, for which electromagnetic ion cyclotron (EMIC) waves are a prime candidate. From its ionospheric vantage point, ELFIN uses its unique pitch-angle-resolving capability to determine whether measured relativistic electron pitch-angle and energy spectra within the loss cone bear the characteristic signatures of scattering by EMIC waves or whether such scattering may be due to other processes. Pairing identical ELFIN satellites with slowly-variable along-track separation allows disambiguation of spatial and temporal evolution of the precipitation over minutes-to-tens-of-minutes timescales, faster than the orbit period of a single low-altitude satellite (~90min). Each satellite carries an energetic particle detector for electrons (EPDE) that measures 50keV to 5MeV electrons with deltaE/E<40% and a fluxgate magnetometer (FGM) on a ~72cm boom that measures magnetic field waves (e.g., EMIC waves) in the range from DC to 5Hz Nyquist (nominally) with <0.3nT/sqrt(Hz) noise at 1Hz. The spinning satellites (T_spin~3s) are equipped with magnetorquers that permit spin-up/down and reorientation maneuvers. The spin axis is placed normal to the orbit plane, allowing full pitch-angle resolution twice per spin. An energetic particle detector for ions (EPDI) measures 250keV-5MeV ions, addressing secondary science. Funded initially by CalSpace and the University Nanosat Program, ELFIN was selected for flight with joint support from NSF and NASA between 2014 and 2018. △ Less

Submitted 16 June, 2020; v1 submitted 13 June, 2020; originally announced June 2020.

Comments: Submitted to Space Science Reviews April 2020. 51 pages, 7 tables, 21 figures

Optimizing the Factual Correctness of a Summary: A Study of Summarizing Radiology Reports

Authors: Yuhao Zhang, Derek Merck, Emily Bao Tsai, Christopher D. Manning, Curtis P. Langlotz

Abstract: Neural abstractive summarization models are able to generate summaries which have high overlap with human references. However, existing models are not optimized for factual correctness, a critical metric in real-world applications. In this work, we develop a general framework where we evaluate the factual correctness of a generated summary by fact-checking it automatically against its reference us… ▽ More Neural abstractive summarization models are able to generate summaries which have high overlap with human references. However, existing models are not optimized for factual correctness, a critical metric in real-world applications. In this work, we develop a general framework where we evaluate the factual correctness of a generated summary by fact-checking it automatically against its reference using an information extraction module. We further propose a training strategy which optimizes a neural summarization model with a factual correctness reward via reinforcement learning. We apply the proposed method to the summarization of radiology reports, where factual correctness is a key requirement. On two separate datasets collected from hospitals, we show via both automatic and human evaluation that the proposed approach substantially improves the factual correctness and overall quality of outputs over a competitive neural summarization system, producing radiology summaries that approach the quality of human-authored ones. △ Less

Submitted 27 April, 2020; v1 submitted 6 November, 2019; originally announced November 2019.

Comments: ACL2020. 13 pages with appendices

Higgs-Photon-Photon Amplitude in the Unitary Gauge

Authors: Er-Cheng Tsai

Abstract: The physical process of Higgs decaying to two photons is the most important mode in the discovery of the Higgs particle. Theoretical calculation of $H\rightarrowγγ$ amplitude therefore provides us with an avenue to explore the underlying theory. Yet some confusion has arisen over the gauge invariant property of $H\rightarrowγγ$ in the standard model. In this paper, we show that if dimensional regu… ▽ More The physical process of Higgs decaying to two photons is the most important mode in the discovery of the Higgs particle. Theoretical calculation of $H\rightarrowγγ$ amplitude therefore provides us with an avenue to explore the underlying theory. Yet some confusion has arisen over the gauge invariant property of $H\rightarrowγγ$ in the standard model. In this paper, we show that if dimensional regularization is adopted, the on-shell 1-loop $H\rightarrowγγ$ amplitude calculated in the standard model is gauge invariant. In particular, we prove that the physical amplitude calculated in the unitary gauge is the same as that calculated in the Feynman gauge. △ Less

Submitted 17 October, 2019; originally announced October 2019.

Ghost Loops are Indispensable in Unitary Gauge

Authors: Er-Cheng Tsai

Abstract: It is conventionally taken for granted that the unitary gauge formulation of quantum gauge field theory has the advantage of preservation unitarity because only physical fields are involved but has the disadvantage of losing renormalizability because of severe ultraviolet divergences due to vector meson propagators. In this paper, we show how to handle the ultraviolet divergent loops so that the p… ▽ More It is conventionally taken for granted that the unitary gauge formulation of quantum gauge field theory has the advantage of preservation unitarity because only physical fields are involved but has the disadvantage of losing renormalizability because of severe ultraviolet divergences due to vector meson propagators. In this paper, we show how to handle the ultraviolet divergent loops so that the physical amplitudes remain gauge invariant. One of the consequences we arrive at is that ghost loops are needed to cancel the divergences due to vector mesons and to give gauge invariant physical amplitudes. △ Less

Submitted 17 October, 2019; originally announced October 2019.

ParaNet - Using Dense Blocks for Early Inference

Authors: Joseph Chuang, Eric Tsai, Kevin Huang, Jay Fetter

Abstract: DenseNets have been shown to be a competitive model among recent convolutional network architectures. These networks utilize Dense Blocks, which are groups of densely connected layers where the output of a hidden layer is fed in as the input of every other layer following it. In this paper, we aim to improve certain aspects of DenseNet, especially when it comes to practicality. We introduce ParaNe… ▽ More DenseNets have been shown to be a competitive model among recent convolutional network architectures. These networks utilize Dense Blocks, which are groups of densely connected layers where the output of a hidden layer is fed in as the input of every other layer following it. In this paper, we aim to improve certain aspects of DenseNet, especially when it comes to practicality. We introduce ParaNet, a new architecture that constructs three pipelines which allow for early inference. We additionally introduce a cascading mechanism such that different pipelines are able to share parameters, as well as logit matching between the outputs of the pipelines. We separately evaluate each of the newly introduced mechanisms of ParaNet, then evaluate our proposed architecture on CIFAR-100. △ Less

Submitted 24 August, 2018; originally announced August 2018.

Ion motion in a polarized current sheet

Authors: Ethan Tsai, Anton Artemyev, Vassilis Angelopoulos

Abstract: We consider the effects of a polarization electric field on transient ion motion in a thin current sheet. Using adiabatic invariants, we analytically describe a variety of ion trajectories in current sheet configurations which include a local minimum or maximum of the scalar potential in the central region. Ions in the current sheet can either be trapped or ejected more efficiently than in an unpo… ▽ More We consider the effects of a polarization electric field on transient ion motion in a thin current sheet. Using adiabatic invariants, we analytically describe a variety of ion trajectories in current sheet configurations which include a local minimum or maximum of the scalar potential in the central region. Ions in the current sheet can either be trapped or ejected more efficiently than in an unpolarized current sheet, depending on the sign and magnitude of the polarization electric field. We derive an expression for the relative phase space volume filled by transient particles as a function of the electric field amplitude. This expression allows us to estimate the dependence of transient particle and current densities on the electric field. We discuss applicability of these results for current sheets observed in planetary magnetospheres. △ Less

Submitted 22 January, 2017; v1 submitted 20 December, 2016; originally announced December 2016.

Comments: 12 Pages, 17 figures, appears in AIP POP Feb 2017 Vol 24 Issue 2. Relevant Topics: solar-system plasma, magnetosphere, solar-system plasma, magnetotail, magnetic reconnection, mathematical foundations, kinetic theory, mathematical foundations, particle dynamics

Muon Anomalous Magnetic Moment and Gauge Symmetry in the Standard Model

Authors: Er-Cheng Tsai

Abstract: No gauge invariant regularization is available for the perturbative calculation of the standard model. One has to add finite counter terms to restore gauge symmetry for the renormalized amplitudes. The muon anomalous magnetic moment can be accurately measured but the experimental result does not entirely agree with the theoretical calculation from the standard model. This paper is to compute the c… ▽ More No gauge invariant regularization is available for the perturbative calculation of the standard model. One has to add finite counter terms to restore gauge symmetry for the renormalized amplitudes. The muon anomalous magnetic moment can be accurately measured but the experimental result does not entirely agree with the theoretical calculation from the standard model. This paper is to compute the contributions to the muon gyromagnetic ratio $g_μ$ due to the finite counter terms. The result obtained is found to be far from sufficient to explain the discrepancy between theory and experiment. △ Less

Submitted 24 October, 2014; originally announced October 2014.

Maintaining Gauge Symmetry in Renormalizing Chiral Gauge Theories

Authors: Er-Cheng Tsai

Abstract: It is known that the $γ_{5}$ scheme of Breitenlohner and Maison (BM) in dimensional regularization requires finite counter-term renormalization to restore gauge symmetry and implementing this finite renormalization in practical calculation is a daunting task even at 1-loop order. In this paper, we show that there is a simple and straightforward method to obtain these finite counter terms by using… ▽ More It is known that the $γ_{5}$ scheme of Breitenlohner and Maison (BM) in dimensional regularization requires finite counter-term renormalization to restore gauge symmetry and implementing this finite renormalization in practical calculation is a daunting task even at 1-loop order. In this paper, we show that there is a simple and straightforward method to obtain these finite counter terms by using the rightmost $γ_{5}$ scheme in which we move all the $γ_{5}$ matrices to the rightmost position before analytically continuing the dimension. For any 1-loop Feynman diagram, the difference between the amplitude regularized in the rightmost $γ_{5}$ scheme and the amplitude regularized in the BM scheme can be easily calculated. The differences for all 1-loop diagrams in the chiral Abelian-Higgs gauge theory and in the chiral non-Abelian gauge theory are shown to be the same as the amplitudes due to the finite counter terms that are required to restore gauge symmetry. △ Less

Submitted 15 December, 2010; originally announced December 2010.

Journal ref: Phys.Rev.D83:065011,2011

Gauge Invariant Treatment of $γ_{5}$ in the Scheme of 't Hooft and Veltman

Authors: Er-Cheng Tsai

Abstract: We propose moving all the $γ_{5}$ matrices to the rightmost position before continuing the dimension, and show that this simple prescription will enable the dimension regularization scheme proposed by 't Hooft and Veltman to be consistent with gauge invariance. We propose moving all the $γ_{5}$ matrices to the rightmost position before continuing the dimension, and show that this simple prescription will enable the dimension regularization scheme proposed by 't Hooft and Veltman to be consistent with gauge invariance. △ Less

Submitted 29 November, 2010; v1 submitted 11 May, 2009; originally announced May 2009.

Comments: 33 pages, LaTeX; Added references

Journal ref: Phys.Rev.D83:025020,2011

The Advantage of Rightmost Ordering for gamma5 in Dimensional Regularization

Authors: Er-Cheng Tsai

Abstract: We propose a gamma5 scheme in dimensional regularization by analytically continuing the dimension after all the gamma5 matrices have been moved to the rightmost position. All Feynman amplitudes corresponding to diagrams with no fermion loops regulated in this manner automatically satisfy the Ward-Takahashi identities. This is in contrast to the scheme of Breitenlohner and Maison, in which finite… ▽ More We propose a gamma5 scheme in dimensional regularization by analytically continuing the dimension after all the gamma5 matrices have been moved to the rightmost position. All Feynman amplitudes corresponding to diagrams with no fermion loops regulated in this manner automatically satisfy the Ward-Takahashi identities. This is in contrast to the scheme of Breitenlohner and Maison, in which finite counter-terms are needed to restore gauge invariance. This rightmost gamma5 scheme also has an advantage over the naive dimensional regularization scheme which does not have a definitive prescription consistent with gauge symmetry. Diagrams with fermion loops can be handled by selecting a proper cut point on each fermion loop to play the role of the point of the rightmost position. △ Less

Submitted 10 May, 2009; originally announced May 2009.

Possibility of $m_t < M_W + m_b$ for SUSY-type of Two Higgs Doublet Model

Authors: Wei-Shu Hou, Ern-Bin Tsai, Chao-Qiang Geng, Paul Turcotte

Abstract: Allowing for realistic uncertainties in input parameters, we demonstrate that the present CLEO limit of $5.4\times 10^{-4}$ for inclusive $b\to sγ$ decay does not yet fully exclude the $t\to bH^+$ decay possibility in supersymmetric type of two Higgs doublet models. Combined with direct search for $t\to bH^+$ via $H^+\to τ^+ν$ at the Tevatron, we conclude that $\tanβ\sim 1$ is the ``allowed" win… ▽ More Allowing for realistic uncertainties in input parameters, we demonstrate that the present CLEO limit of $5.4\times 10^{-4}$ for inclusive $b\to sγ$ decay does not yet fully exclude the $t\to bH^+$ decay possibility in supersymmetric type of two Higgs doublet models. Combined with direct search for $t\to bH^+$ via $H^+\to τ^+ν$ at the Tevatron, we conclude that $\tanβ\sim 1$ is the ``allowed" window for $m_{H^+} < m_t < M_W$. The possibility becomes excluded, however, {\it if} the CLEO limit is pushed below $4\times 10^{-4}$. △ Less

Submitted 18 July, 1994; originally announced July 1994.

Comments: 11 pages + 3 Figures (not included), ReVTeX, NTUTH-94-15 and NHCU-HEP-94-16

Journal ref: Prog.Theor.Phys.94:1089-1096,1995