Publications

Abstract The authors add a missing reference to the article published in phys. stat. sol. (b) 241 , 2775 (2004), referring to one of their own earlier papers in Nano Lett. 3 , 1063 (2003). (© 2005 WILEY‐VCH Verlag GmbH & Co. KGaA, Weinheim)

If you ask a human to describe an image, they might do so in a thousand different ways. Traditionally, image captioning models are trained to generate a single "best" (most like a reference) image caption. Unfortunately, doing so encourages captions that are "informationally impoverished," and focus on only a subset of the possible details, while ignoring other potentially useful information in the scene. In this work, we introduce a simple, yet novel, method: "Image Captioning by Committee Consensus" (IC3), designed to generate a single caption that captures high-level details from several annotator viewpoints. Humans rate captions produced by IC3 at least as helpful as baseline SOTA models more than two thirds of the time, and IC3 can improve the performance of SOTA automated recall systems by up to 84%, outperforming single human-generated reference captions, and indicating significant improvements over SOTA approaches for visual description. Code is available at https://davidmchan.github.io/caption-by-committee/

We propose an economic model based on user utility to study the adoption of new network architectures such as IPv6. We use mathematical analysis and simulation studies to understand the role of various factors such as user and network benefits, switching costs, and the impact of converters on the adoption of new network architectures. In addition to corroborating various commonly held beliefs about new network architecture adoption, our analysis and simulation studies also reveal several surprising and non-intuitive results. For example, while in general, increasing the efficiency of converters hastens the adoption of new network architectures, there are cases in which more efficient converters hinder the adoption of such architectures. This and other results in the paper increase our understanding of new network architecture adoption and guide the design and implementation of mechanisms to hasten new network architecture adoption.

The evaluation of reaction energies between solids using density functional theory (DFT) is of practical importance in many technological fields and paramount in the study of the phase stability of known and predicted compounds. In this work, we present a comparison between reaction energies provided by experiments and computed by DFT in the generalized gradient approximation (GGA), using a Hubbard U parameter for some transition metal elements (GGA+U). We use a data set of 135 reactions involving the formation of ternary oxides from binary oxides in a broad range of chemistries and crystal structures. We find that the computational errors can be modeled by a normal distribution with a mean close to zero and a standard deviation of 24 meV/atom. The significantly smaller error compared to the more commonly reported errors in the formation energies from the elements is related to the larger cancellation of errors in energies when reactions involve chemically similar compounds. This result is of importance for phase diagram computations for which the relevant reaction energies are often not from the elements but from chemically close phases (e.g., ternary oxides versus binary oxides). In addition, we discuss the distribution of computational errors among chemistries and show that the use of a Hubbard U parameter is critical to the accuracy of reaction energies involving transition metals even when no major change in formal oxidation state is occurring.

Humanoid robots hold significant potential in accomplishing daily tasks across diverse environments thanks to their flexibility and human-like morphology. Recent works have made significant progress in humanoid whole-body control and loco-manipulation leveraging optimal control or reinforcement learning. However, these methods require tedious task-specific tuning for each task to achieve satisfactory behaviors, limiting their versatility and scalability to diverse tasks in daily scenarios. To that end, we introduce SkillBlender, a novel hierarchical reinforcement learning framework for versatile humanoid loco-manipulation. SkillBlender first pretrains goal-conditioned task-agnostic primitive skills, and then dynamically blends these skills to accomplish complex loco-manipulation tasks with minimal task-specific reward engineering. We also introduce SkillBench, a parallel, cross-embodiment, and diverse simulated benchmark containing three embodiments, four primitive skills, and eight challenging loco-manipulation tasks, accompanied by a set of scientific evaluation metrics balancing accuracy and feasibility. Extensive simulated experiments show that our method significantly outperforms all baselines, while naturally regularizing behaviors to avoid reward hacking, resulting in more accurate and feasible movements for diverse loco-manipulation tasks in our daily scenarios. Our code and benchmark will be open-sourced to the community to facilitate future research. Project page: https://usc-gvl.github.io/SkillBlender-web/.

Lack of proper English pronunciations is a major problem for immigrant\npopulation in developed countries like U.S. This poses various problems,\nincluding a barrier to entry into mainstream society. This paper presents a\nresearch study that explores the use of speech technologies merged with\nactivity-based and arcade-based games to do pronunciation feedback for Hispanic\nchildren within the U.S. A 3-month long study with immigrant population in\nCalifornia was used to investigate and analyze the effectiveness of computer\naided pronunciation feedback through games. In addition to quantitative\nfindings that point to statistically significant gains in pronunciation\nquality, the paper also explores qualitative findings, interaction patterns and\nchallenges faced by the researchers in dealing with this community. It also\ndescribes the issues involved in dealing with pronunciation as a competency.\n

We present broad-band observations and analysis of Swift gamma-ray burst (GRB) 120119A. Our early-time afterglow detections began under 15 s after the burst in the host frame (redshift z = 1.73), and they yield constraints on the burst energetics and local environment. Late-time afterglow observations of the burst show evidence for a moderate column of dust (A_V ≈ 1.1 mag) similar to, but statistically distinct from, dust seen along Small Magellanic Cloud sightlines. Deep late-time observations reveal a dusty, rapidly star-forming host galaxy. Most notably, our early-time observations exhibit a significant red-to-blue colour change in the first ∼200 s after the trigger at levels heretofore unseen in GRB afterglows. This colour change, which is coincident with the final phases of the prompt emission, is a hallmark prediction of the photodestruction of dust in GRB afterglows. We test whether dust-destruction signatures are significantly distinct from other sources of colour change, namely a change in the intrinsic spectral index β. We find that a time-varying power-law spectrum alone cannot adequately describe the observed colour change, and allowing for dust destruction (via a time-varying A_V) significantly improves the fit. While not definitively ruling out other possibilities, this event provides the best support yet for the direct detection of dust destruction in the local environment of a GRB.