Search

In image classification, "debiasing" aims to train a classifier to be less susceptible to dataset bias, the strong correlation between peripheral attributes of data samples and a target class. For example, even if the frog class in the dataset mainly consists of frog images with a swamp background (i.e., bias-aligned samples), a debiased classifier should be able to correctly classify a frog at a beach (i.e., bias-conflicting samples). Recent debiasing approaches commonly use two components for debiasing, a biased model $f_B$ and a debiased model $f_D$. $f_B$ is trained to focus on bias-aligned samples (i.e., overfitted to the bias) while $f_D$ is mainly trained with bias-conflicting samples by concentrating on samples which $f_B$ fails to learn, leading $f_D$ to be less susceptible to the dataset bias. While the state-of-the-art debiasing techniques have aimed to better train $f_D$, we focus on training $f_B$, an overlooked component until now. Our empirical analysis reveals that removing the bias-conflicting samples from the training set for $f_B$ is important for improving the debiasing performance of $f_D$. This is due to the fact that the bias-conflicting samples work as noisy samples for amplifying the bias for $f_B$ since those samples do not include the bias attribute. To this end, we propose a simple yet effective data sample selection method which removes the bias-conflicting samples to construct a bias-amplified dataset for training $f_B$. Our data sample selection method can be directly applied to existing reweighting-based debiasing approaches, obtaining consistent performance boost and achieving the state-of-the-art performance on both synthetic and real-world datasets., Accepted to AAAI 2023

Why do leaders of states engage in face-to-face diplomacy? The existing answers to this question largely focus on the private effect of the leaders engaging in the face-to-face meeting or on the public effects created by the mere occurrences of the meetings. This research aims to innovate by looking at the public and the performance side of face-to-face summit diplomacy in the context of the bilateral alliance. The non-substantive visual performances of face-to-face summit meetings can signal the relational status of the summits to the public. I argue that the performances of equivalence (peer-ness) or hierarchy (deference) between the summits affect the domestic public opinion, which ultimately affects the prospect for alliance coordination. In a bilateral alliance, a government can face a dilemma between the conflicting demands of the ally and of its own citizens, which I will coin as the “alliance-public dilemma.” In the Korea-U.S. alliance, the Korean public wants its government to maintain a close and friendly relationship with the U.S. strategically but does not want it to show a weaker and losing stance vis-à-vis the U.S. emotionally. This dual demand puts the Korean government in a dilemma between (1)having to strategically coordinate with (make concessions to) the U.S. and (2)satisfying its citizens, that are highly sensitive to the implicit hierarchy of the two countries. The Korean government can alleviate the dilemma by winning public support through deliberately performing friendliness and peer-ness in face-to-face summit meetings. I studied 17 Korea-U.S. summit meetings during 2002-2021 held in bilateral visits and provided 5 selected case studies. The cases show that a high level of peer-ness performed is linked to high public approval of summitry or to the increased presidential approval rate, while the high level of deference performed is linked to the opposite result, on the condition that the two summits showed friendly relationship.

Despite the impressive performance of deep networks in vision, language, and healthcare, unpredictable behaviors on samples from the distribution different than the training distribution cause severe problems in deployment. For better reliability of neural-network-based classifiers, we define a new task, natural attribute-based shift (NAS) detection, to detect the samples shifted from the training distribution by some natural attribute such as age of subjects or brightness of images. Using the natural attributes present in existing datasets, we introduce benchmark datasets in vision, language, and medical for NAS detection. Further, we conduct an extensive evaluation of prior representative out-of-distribution (OOD) detection methods on NAS datasets and observe an inconsistency in their performance. To understand this, we provide an analysis on the relationship between the location of NAS samples in the feature space and the performance of distance- and confidence-based OOD detection methods. Based on the analysis, we split NAS samples into three categories and further suggest a simple modification to the training objective to obtain an improved OOD detection method that is capable of detecting samples from all NAS categories.

We evaluate the out-of-distribution (OOD) detection performance of self-supervised learning (SSL) techniques with a new evaluation framework. Unlike the previous evaluation methods, the proposed framework adjusts the distance of OOD samples from the in-distribution samples. We evaluate an extensive combination of OOD detection algorithms on three different implementations of the proposed framework using simulated samples, images, and text. SSL methods consistently demonstrated the improved OOD detection performance in all evaluation settings., Accepted to NeurIPS 2020 Workshop: Self-Supervised Learning - Theory and Practice

We show that the stability theorem of the depolarizing channel holds for the output quantum $p$-R��nyi entropy for $p \ge 2$ or $p=1$, which is an extension of the well known case $p=2$. As an application, we present a protocol in which Bob determines whether Alice prepares a pure quantum state close to a product state. In the protocol, Alice transmits to Bob multiple copies of a pure state through a depolarizing channel, and Bob estimates its output quantum $p$-R��nyi entropy. By using our stability theorem, we show that Bob can determine whether her preparation is appropriate., 9 pages, no figure

It was shown [T.S. Cubitt et al., IEEE Trans. Inform. Theory 57, 8114 (2011)] that there exist quantum channels where a single use cannot transmit classical information perfectly yet two uses can. This phenomenon is called the superactivation of the zero-error classical capacity which does not occur in classical channels. In this paper, it is shown that qubit channels cannot generate the superactivation., Comment: 4 pages