Publications

Many of the most fundamental examples in probability involve the pose statistics of coins and dice as they are dropped on a flat surface. For these parts, the probability assigned to each stable face is justified based on part symmetry, although most gamblers are familiar with the possibility of loaded dice. In industrial part feeding, parts also arrive in random orientations. We consider the following problem: given part geometry and parameters such as center of mass, estimate the probability of encountering each stable pose of the part. We describe three estimators for solving this problem for polyhedral parts with known center of mass. The first estimator uses a quasistatic motion model that is computed in time O(n log n) for a part with n vertices. The second estimator has the same time complexity but takes into account a measure of dynamic stability based on perturbation. The third estimator uses repeated Monte Carlo experiments with a mechanics simulation package. To evaluate these estimators, we used a robot and computer vision system to record the pose statistics based on 3595 physical drop experiments with four different parts. We compare this data to the results from each estimator. We believe this is the first paper to systematically compare alternative estimators and to correlate their performance with statistically significant experiments on industrial parts.

Modern machine learning algorithms are increasingly computationally demanding, requiring specialized hardware and distributed computation to achieve high performance in a reasonable time frame. Many hyperparameter search algorithms have been proposed for improving the efficiency of model selection, however their adaptation to the distributed compute environment is often ad-hoc. We propose Tune, a unified framework for model selection and training that provides a narrow-waist interface between training scripts and search algorithms. We show that this interface meets the requirements for a broad range of hyperparameter search algorithms, allows straightforward scaling of search to large clusters, and simplifies algorithm implementation. We demonstrate the implementation of several state-of-the-art hyperparameter search algorithms in Tune. Tune is available at http://ray.readthedocs.io/en/latest/tune.html.

Abstract Understanding thermal transport across metal/semiconductor interfaces is crucial for the heat dissipation of electronics. The dominant heat carriers in non-metals, phonons, are thought to transport elastically across most interfaces, except for a few extreme cases where the two materials that formed the interface are highly dissimilar with a large difference in Debye temperature. In this work, we show that even for two materials with similar Debye temperatures (Al/Si, Al/GaN), a substantial portion of phonons will transport inelastically across their interfaces at high temperatures, significantly enhancing interface thermal conductance. Moreover, we find that interface sharpness strongly affects phonon transport process. For atomically sharp interfaces, phonons are allowed to transport inelastically and interface thermal conductance linearly increases at high temperatures. With a diffuse interface, inelastic phonon transport diminishes. Our results provide new insights on phonon transport across interfaces and open up opportunities for engineering interface thermal conductance specifically for materials of relevance to microelectronics.

A recent trend in routing research is to avoid inefficiencies in network-level routing by allowing hosts to either choose routes themselves (e.g., source routing) or use overlay routing networks (e.g., Detour or RON). Such approaches result in selfish routing, because routing decisions are no longer based on system-wide criteria but are instead designed to optimize host-based or overlay-based metrics. A series of theoretical results showing that selfish routing can result in suboptimal system behavior have cast doubts on this approach. In this paper, we use a game-theoretic approach to investigate the performance of selfish routing in Internet-like environments based on realistic topologies and traffic demands in our simulations. We show that in contrast to theoretical worst cases, selfish routing achieves close to optimal average latency in such environments. However, such performance benefits come at the expense of significantly increased congestion on certain links. Moreover, the adaptive nature of selfish overlays can significantly reduce the effectiveness of traffic engineering by making network traffic less predictable.

Using a new mAb, 2F1, we characterize a mouse natural killer (NK) cell antigen termed `killer cell lectin-like receptor G1' (KLRG1; formerly mouse MAFA or 2F1-Ag). KLRG1 is expressed on 30 – 60 % of murine NK cells, and a small fraction of T cells, and is composed of a homodimer of glycosylated 30 – 38-kDa subunits. Strikingly, cell surface expression of KLRG1 by NK cells was substantially down-regulated in mice deficient for expression of class I molecules, in contrast to the Ly49 lectin-like NK receptors, which are up-regulated in class Ideficient mice. We could not demonstrate binding of KLRG1 to class I molecules in a cell-cell adhesion assay. Transgenic expression of KLRG1 under heterologous transcription elements was unaffected by class I deficiency, indicating that class I molecules do not affect the KLRG1 protein directly, and suggesting that regulation is at the level of expression of the endogenous KLRG1 gene. Evidence is presented that class I molecules regulate KLRG1 via interactions with class I-specific inhibitory Ly49 molecules and SHP-1 signaling. Thus, although KLRG1 and Ly49 molecules are both lectin-like inhibitory receptors that are regulated by class I expression, the effects of class I on the cell surface expression of the molecules are opposing, and the underlying regulatory mechanisms are distinct.

PyTorch \texttt{2.x} introduces a compiler designed to accelerate deep learning programs. However, for machine learning researchers, adapting to the PyTorch compiler to full potential can be challenging. The compiler operates at the Python bytecode level, making it appear as an opaque box. To address this, we introduce \texttt{depyf}, a tool designed to demystify the inner workings of the PyTorch compiler. \texttt{depyf} decompiles bytecode generated by PyTorch back into equivalent source code, and establishes connections between in-memory code objects and their on-disk source code counterparts. This feature enables users to step through the source code line by line using debuggers, thus enhancing their understanding of the underlying processes. Notably, \texttt{depyf} is non-intrusive and user-friendly, primarily relying on two convenient context managers for its core functionality. The project is \href{https://github.com/thuml/depyf}{ openly available} and is recognized as a \href{https://pytorch.org/ecosystem/}{PyTorch ecosystem project}.

A memristor emulator which imitates the behavior of a TiO <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub> memristor is presented. Our emulator is built from off-the-shelf solid state components. To develop real world memristor circuit applications, the emulator can be used for breadboard experiments in real time. Two or more memristor emulators can be connected in serial, in parallel, or in hybrid (serial and parallel combined) with identical or opposite polarities. With a simple change of connection, each memristor emulator can be switched between a decremental configuration or an incremental configuration. The hardware and spice simulation of the proposed emulator showed promising results that provides an alternative solution of hp TiO <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub> memristor model in real circuit.

Abstract Sodium‐ion battery (SIB) shows great potential for large‐scale energy storage, high‐power and low‐temperature supplies. Pitch is a promising precursor for preparing carbon anode materials for SIBs owing to its low cost, high carbon yield, and structural tunability. In this review, the preparation and application of pitch‐derived soft‐ and hard‐carbons for SIBs are discussed from a practical perspective. We first introduce the full‐slope sodium storage behavior of soft carbon, mainly focused on the storage mechanism, high‐rate feature, modification strategies, application scenario, and existing challenges. Thereafter, we present the preparation of pitch‐derived hard carbon mainly through pre‐oxidation and carbonization. The crucial influencing factors, including pitch property, pre‐oxidation temperature, holding time, heating rate, atmosphere, and carbonization conditions, as well as the underlying mechanisms, are discussed. From scientific and industrial points of view, we also propose the challenges for application of pitch‐derived carbon materials. Future efforts should be paid on clarifying the sodium storage mechanisms and electrochemical kinetics of carbon, revealing the screening criteria of pitch and the “preparation‐structure‐performance” relationships through a combination of experiments, artificial intelligence, and big data techniques. Material design and preparation guidance should be obtained to push forward pitch‐based carbon for practical SIBs.