Showing 1–7 of 7 results for author: Cheng, T S

Characterizing Overfitting in Kernel Ridgeless Regression Through the Eigenspectrum

Authors: Tin Sum Cheng, Aurelien Lucchi, Anastasis Kratsios, David Belius

Abstract: We derive new bounds for the condition number of kernel matrices, which we then use to enhance existing non-asymptotic test error bounds for kernel ridgeless regression (KRR) in the over-parameterized regime for a fixed input dimension. For kernels with polynomial spectral decay, we recover the bound from previous work; for exponential decay, our bound is non-trivial and novel. Our contribution is… ▽ More We derive new bounds for the condition number of kernel matrices, which we then use to enhance existing non-asymptotic test error bounds for kernel ridgeless regression (KRR) in the over-parameterized regime for a fixed input dimension. For kernels with polynomial spectral decay, we recover the bound from previous work; for exponential decay, our bound is non-trivial and novel. Our contribution is two-fold: (i) we rigorously prove the phenomena of tempered overfitting and catastrophic overfitting under the sub-Gaussian design assumption, closing an existing gap in the literature; (ii) we identify that the independence of the features plays an important role in guaranteeing tempered overfitting, raising concerns about approximating KRR generalization using the Gaussian design assumption in previous literature. △ Less

Submitted 29 May, 2024; v1 submitted 2 February, 2024; originally announced February 2024.

A Theoretical Analysis of the Test Error of Finite-Rank Kernel Ridge Regression

Authors: Tin Sum Cheng, Aurelien Lucchi, Ivan Dokmanić, Anastasis Kratsios, David Belius

Abstract: Existing statistical learning guarantees for general kernel regressors often yield loose bounds when used with finite-rank kernels. Yet, finite-rank kernels naturally appear in several machine learning problems, e.g.\ when fine-tuning a pre-trained deep neural network's last layer to adapt it to a novel task when performing transfer learning. We address this gap for finite-rank kernel ridge regres… ▽ More Existing statistical learning guarantees for general kernel regressors often yield loose bounds when used with finite-rank kernels. Yet, finite-rank kernels naturally appear in several machine learning problems, e.g.\ when fine-tuning a pre-trained deep neural network's last layer to adapt it to a novel task when performing transfer learning. We address this gap for finite-rank kernel ridge regression (KRR) by deriving sharp non-asymptotic upper and lower bounds for the KRR test error of any finite-rank KRR. Our bounds are tighter than previously derived bounds on finite-rank KRR, and unlike comparable results, they also remain valid for any regularization parameters. △ Less

Submitted 3 October, 2023; v1 submitted 2 October, 2023; originally announced October 2023.

A Tale of Two Cultures: Comparing Interpersonal Information Disclosure Norms on Twitter

Authors: Mainack Mondal, Anju Punuru, Tyng-Wen Scott Cheng, Kenneth Vargas, Chaz Gundry, Nathan S Driggs, Noah Schill, Nathaniel Carlson, Josh Bedwell, Jaden Q Lorenc, Isha Ghosh, Yao Li, Nancy Fulda, Xinru Page

Abstract: We present an exploration of cultural norms surrounding online disclosure of information about one's interpersonal relationships (such as information about family members, colleagues, friends, or lovers) on Twitter. The literature identifies the cultural dimension of individualism versus collectivism as being a major determinant of offline communication differences in terms of emotion, topic, and… ▽ More We present an exploration of cultural norms surrounding online disclosure of information about one's interpersonal relationships (such as information about family members, colleagues, friends, or lovers) on Twitter. The literature identifies the cultural dimension of individualism versus collectivism as being a major determinant of offline communication differences in terms of emotion, topic, and content disclosed. We decided to study whether such differences also occur online in context of Twitter when comparing tweets posted in an individualistic (U.S.) versus a collectivist (India) society. We collected more than 2 million tweets posted in the U.S. and India over a 3 month period which contain interpersonal relationship keywords. A card-sort study was used to develop this culturally-sensitive saturated taxonomy of keywords that represent interpersonal relationships (e.g., ma, mom, mother). Then we developed a high-accuracy interpersonal disclosure detector based on dependency-parsing (F1-score: 86%) to identify when the words refer to a personal relationship of the poster (e.g., "my mom" as opposed to "a mom"). This allowed us to identify the 400K+ tweets in our data set which actually disclose information about the poster's interpersonal relationships. We used a mixed methods approach to analyze these tweets (e.g., comparing the amount of joy expressed about one's family) and found differences in emotion, topic, and content disclosed between tweets from the U.S. versus India. Our analysis also reveals how a combination of qualitative and quantitative methods are needed to uncover these differences; Using just one or the other can be misleading. This study extends the prior literature on Multi-Party Privacy and provides guidance for researchers and designers of culturally-sensitive systems. △ Less

Submitted 26 September, 2023; originally announced September 2023.

Comments: This work will be presented at the 26th ACM Conference on Computer-Supported Cooperative Work and Social Computing (CSCW 2023). This paper will also be published in The Proceedings of the ACM on Human Computer Interaction

Cathodoluminescence spectroscopy of monolayer hexagonal boron nitride

Authors: K. Shima, T. S. Cheng, C. J. Mellor, P. H. Beton, C. Elias, P. Valvin, B. Gil, G. Cassabois, S. V. Novikov, S. F. Chichibu

Abstract: Cathodoluminescence (CL) spectroscopy is a powerful technique for studying emission properties of optoelectronic materials because CL is free from excitable bandgap limits and from ambiguous signals due to simple light scattering and resonant Raman scattering potentially involved in the photoluminescence (PL) spectra. However, direct CL measurements of atomically thin two-dimensional materials, su… ▽ More Cathodoluminescence (CL) spectroscopy is a powerful technique for studying emission properties of optoelectronic materials because CL is free from excitable bandgap limits and from ambiguous signals due to simple light scattering and resonant Raman scattering potentially involved in the photoluminescence (PL) spectra. However, direct CL measurements of atomically thin two-dimensional materials, such as transition metal dichalcogenides and hexagonal boron nitride (hBN), have been difficult due to the small excitation volume that interacts with high-energy electron beams (e-beams). Herein, distinct CL signals from a monolayer hBN, namely mBN, epitaxial film grown on a highly oriented pyrolytic graphite substrate are shown by using a home-made CL system capable of large-area and surface-sensitive excitation by an e-beam. The spatially resolved CL spectra at 13 K exhibited a predominant 5.5-eV emission band, which has been ascribed to originate from multilayered aggregates of hBN, markedly at thicker areas formed on the step edges of the substrate. Conversely, a faint peak at 6.04 eV was routinely observed from atomically flat areas. Since the energy agreed with the PL peak of 6.05 eV at 10 K that has been assigned as being due to the recombination of phonon-assisted direct excitons of mBN by Elias et al. [Nat. Commun. 10, 2639 (2019)], the CL peak at 6.04 eV is attributed to originate from the mBN epilayer. The CL results support the transition from indirect bandgap in bulk hBN to direct bandgap in mBN, in analogy with molybdenum disulfide. The results also encourage to elucidate emission properties of other low-dimensional materials with reduced excitation volumes by using the present CL configuration. △ Less

Submitted 17 May, 2023; originally announced May 2023.

Comments: 7 pages, 3 figures

Band gap measurements of monolayer h-BN and insights into carbon-related point defects

Authors: Ricardo Javier Peña Román, Fábio J R Costa Costa, Alberto Zobelli, Christine Elias, Pierre Valvin, Guillaume Cassabois, Bernard Gil, Alex Summerfield, Tin S Cheng, Christopher J Mellor, Peter H Beton, Sergei V Novikov, Luiz F Zagonel

Abstract: Being a flexible wide band gap semiconductor, hexagonal boron nitride (h-BN) has great potential for technological applications like efficient deep ultraviolet light sources, building block for two-dimensional heterostructures and room temperature single photon emitters in the ultraviolet and visible spectral range. To enable such applications, it is mandatory to reach a better understanding of th… ▽ More Being a flexible wide band gap semiconductor, hexagonal boron nitride (h-BN) has great potential for technological applications like efficient deep ultraviolet light sources, building block for two-dimensional heterostructures and room temperature single photon emitters in the ultraviolet and visible spectral range. To enable such applications, it is mandatory to reach a better understanding of the electronic and optical properties of h-BN and the impact of various structural defects. Despite the large efforts in the last years, aspects such as the electronic band gap value, the exciton binding energy and the effect of point defects remained elusive, particularly when considering a single monolayer. Here, we directly measured the density of states of a single monolayer of h-BN epitaxially grown on highly oriented pyrolytic graphite, by performing low temperature scanning tunneling microscopy (STM) and spectroscopy (STS). The observed h-BN electronic band gap on defect-free regions is $(6.8\pm0.2)$ eV. Using optical spectroscopy to obtain the h-BN optical band gap, the exciton binding energy is determined as being of $(0.7\pm0.2)$ eV. In addition, the locally excited cathodoluminescence and photoluminescence show complex spectra that are typically associated to intragap states related to carbon defects. Moreover, in some regions of the monolayer h-BN we identify, using STM, point defects which have intragap electronic levels around 2.0 eV below the Fermi level. △ Less

Submitted 16 July, 2021; originally announced July 2021.

Comments: 50 Pages, 8 Figures, 100+ references

Journal ref: 2D Material 8 044001 (2021)

Identifying Carbon as the Source of Visible Single Photon Emission from Hexagonal Boron Nitride

Authors: Noah Mendelson, Dipankar Chugh, Jeffrey R. Reimers, Tin S. Cheng, Andreas Gottscholl, Hu Long, Christopher J. Mellor, Alex Zettl, Vladimir Dyakonov, Peter H. Beton, Sergei V. Novikov, Chennupati Jagadish, Hark Hoe Tan, Michael J. Ford, Milos Toth, Carlo Bradac, Igor Aharonovich

Abstract: Single photon emitters (SPEs) in hexagonal boron nitride (hBN) have garnered significant attention over the last few years due to their superior optical properties. However, despite the vast range of experimental results and theoretical calculations, the defect structure responsible for the observed emission has remained elusive. Here, by controlling the incorporation of impurities into hBN and by… ▽ More Single photon emitters (SPEs) in hexagonal boron nitride (hBN) have garnered significant attention over the last few years due to their superior optical properties. However, despite the vast range of experimental results and theoretical calculations, the defect structure responsible for the observed emission has remained elusive. Here, by controlling the incorporation of impurities into hBN and by comparing various synthesis methods, we provide direct evidence that the visible SPEs are carbon related. Room temperature optically detected magnetic resonance (ODMR) is demonstrated on ensembles of these defects. We also perform ion implantation experiments and confirm that only carbon implantation creates SPEs in the visible spectral range. Computational analysis of hundreds of potential carbon-based defect transitions suggest that the emission results from the negatively charged VBCN- defect, which experiences long-range out-of-plane deformations and is environmentally sensitive. Our results resolve a long-standing debate about the origin of single emitters at the visible range in hBN and will be key to deterministic engineering of these defects for quantum photonic devices. △ Less

Submitted 20 April, 2020; v1 submitted 2 March, 2020; originally announced March 2020.

Concept of a Value in Multilevel Security Databases

Authors: Jia Tao, Shashi Gadia, Tsz Shing Cheng

Abstract: This paper has been withdrawn. This paper has been withdrawn. △ Less

Submitted 29 June, 2009; v1 submitted 21 March, 2007; originally announced March 2007.

Comments: This paper has been withdrawn