arXiv subjects

Shengzhe Wang

Publications and source records attributed to Shengzhe Wang.

25 records · Page 2Linked to original sources

arXiv2025

Parallel $(1+ε)$-Approximate Multi-Commodity Mincost Flow in Almost Optimal Depth and Work

We present a parallel algorithm for computing $(1+ε)$-approximate mincost flow on an undirected graph with $m$ edges, where capacities and costs are assigned to both edges and vertices. Our algorithm achieves $\hat{O}(m)$ work and $\hat{O}(1)$ depth when $ε> 1/\mathrm{polylog}(m)$, making both the work and depth almost optimal, up to a subpolynomial factor. Previous algorithms with $\hat{O}(m)$ work required $Ω(m)$ depth, even for special cases of mincost flow with only edge capacities or max flow with vertex capacities. Our result generalizes prior almost-optimal parallel $(1+ε)$-approximation algorithms for these special cases, including shortest paths [Li, STOC'20] [Andoni, Stein, Zhong, STOC'20] [Rozhen, Haeupler, Marinsson, Grunau, Zuzic, STOC'23] and max flow with only edge capacities [Agarwal, Khanna, Li, Patil, Wang, White, Zhong, SODA'24]. Our key technical contribution is the first construction of length-constrained flow shortcuts with $(1+ε)$ length slack, $\hat{O}(1)$ congestion slack, and $\hat{O}(1)$ step bound. This provides a strict generalization of the influential concept of $(\hat{O}(1),ε)$-hopsets [Cohen, JACM'00], allowing for additional control over congestion. Previous length-constrained flow shortcuts [Haeupler, Hershkowitz, Li, Roeyskoe, Saranurak, STOC'24] incur a large constant in the length slack, which would lead to a large approximation factor. To enable our flow algorithms to work under vertex capacities, we also develop a close-to-linear time algorithm for computing length-constrained vertex expander decomposition. Building on Cohen's idea of path-count flows [Cohen, SICOMP'95], we further extend our algorithm to solve $(1+ε)$-approximate $k$-commodity mincost flow problems with almost-optimal $\hat{O}(mk)$ work and $\hat{O}(1)$ depth, independent of the number of commodities $k$.

cs.DS↗

arXiv2025

OneRec Technical Report

Recommender systems have been widely used in various large-scale user-oriented platforms for many years. However, compared to the rapid developments in the AI community, recommendation systems have not achieved a breakthrough in recent years. For instance, they still rely on a multi-stage cascaded architecture rather than an end-to-end approach, leading to computational fragmentation and optimization inconsistencies, and hindering the effective application of key breakthrough technologies from the AI community in recommendation scenarios. To address these issues, we propose OneRec, which reshapes the recommendation system through an end-to-end generative approach and achieves promising results. Firstly, we have enhanced the computational FLOPs of the current recommendation model by 10 $\times$ and have identified the scaling laws for recommendations within certain boundaries. Secondly, reinforcement learning techniques, previously difficult to apply for optimizing recommendations, show significant potential in this framework. Lastly, through infrastructure optimizations, we have achieved 23.7% and 28.8% Model FLOPs Utilization (MFU) on flagship GPUs during training and inference, respectively, aligning closely with the LLM community. This architecture significantly reduces communication and storage overhead, resulting in operating expense that is only 10.6% of traditional recommendation pipelines. Deployed in Kuaishou/Kuaishou Lite APP, it handles 25% of total queries per second, enhancing overall App Stay Time by 0.54% and 1.24%, respectively. Additionally, we have observed significant increases in metrics such as 7-day Lifetime, which is a crucial indicator of recommendation experience. We also provide practical lessons and insights derived from developing, optimizing, and maintaining a production-scale recommendation system with significant real-world impact.

cs.IR↗

arXiv2025

The H2 Glow of a Quiescent Molecular Cloud Observed with JWST

We report JWST MIRI/MRS observations of the H2 S(1) 17.04 micron transition in two regions in the boundary of the Taurus Molecular Cloud. The two regions, denoted Edge (near the relatively sharp boundary of the 13CO J=1-0 emission) and Peak (the location of the strongest H2 emission observed with Spitzer), have average intensities of 14.5 MJy/sr and 32.1 MJy/sr, respectively. We find small scale structures of characteristic size 1.0 to 2.5 arcseconds, corresponding to 140 AU to 350 AU, with characteristic intensity above the extended background of 10 MJy/sr, corresponding to a J = 3 column density of 1.6x1017/cm2. The most plausible explanation for the observed intensities from level 845 K above the J = 1 ortho-H2 ground state level is excitation by collisions with H2 molecules (the hydrogen in this region being predominantly molecular). Two mechanisms, turbulent dissipation and shocks, have been proposed for heating of localized regions of the ISM to temperatures ~1000 K to explain abundances of and emission from particular molecules. While we cannot determine unique values of density and kinetic temperature, the solutions in best agreement with predictions of shock models are H2 density = 370 /cm3 and kinetic temperature = 1000 K. The total H2 column density of the small-scale structures under these conditions is ~8x1017/cm2. This first direct detection of heated tiny scale structure in the quiescent molecular interstellar medium has significant implications for the physical structure of this phase of the ISM and maintaining supersonic motions within it.

astro-ph.GA↗

arXiv2025

Length-Constrained Directed Expander Decomposition and Length-Constrained Vertex-Capacitated Flow Shortcuts

We show the existence of length-constrained expander decomposition in directed graphs and undirected vertex-capacitated graphs. Previously, its existence was shown only in undirected edge-capacitated graphs [Haeupler-Räcke-Ghaffari, STOC 2022; Haeupler-Hershkowitz-Tan, FOCS 2024]. Along the way, we prove the multi-commodity maxflow-mincut theorems for length-constrained expansion in both directed and undirected vertex-capacitated graphs. Based on our decomposition, we build a length-constrained flow shortcut for undirected vertex-capacitated graphs, which roughly speaking is a set of edges and vertices added to the graph so that every multi-commodity flow demand can be routed with approximately the same vertex-congestion and length, but all flow paths only contain few edges. This generalizes the shortcut for undirected edge-capacitated graphs from [Haeupler-Hershkowitz-Li-Roeyskoe-Saranurak, STOC 2024]. Length-constrained expander decomposition and flow shortcuts have been crucial in the recent algorithms in undirected edge-capacitated graphs [Haeupler-Hershkowitz-Li-Roeyskoe-Saranurak, STOC 2024; Haeupler-Long-Saranurak, FOCS 2024]. Our work thus serves as a foundation to generalize these concepts to directed and vertex-capacitated graphs.

cs.DS↗

arXiv2024

LearnedFTL: A Learning-Based Page-Level FTL for Reducing Double Reads in Flash-Based SSDs

We present LearnedFTL, a new on-demand page-level flash translation layer (FTL) design, which employs learned indexes to improve the address translation efficiency of flash-based SSDs. The first of its kind, it reduces the number of double reads induced by address translation in random read accesses. LearnedFTL proposes three key techniques: an in-place-update linear model to build learned indexes efficiently, a virtual PPN representation to obtain contiguous PPNs for sorted LPNs, and a group-based allocation and model training via GC/rewrite strategy to reduce the training overhead. By tightly integrating the aforementioned key techniques, LearnedFTL considerably speeds up address translation while reducing the number of flash read accesses caused by the address translation. Our extensive experiments on a FEMU-based prototype show that LearnedFTL can reduce up to 55.5\% address translation-induced double reads. As a result, LearnedFTL reduces the P99 tail latency by 2.9$\times$ $\sim$ 12.2$\times$ with an average of 5.5$\times$ and 8.2$\times$ compared to the state-of-the-art TPFTL and LeaFTL schemes, respectively.

cs.AR↗

arXiv2023

A continuous cold rubidium atomic beam with enhanced flux and tunable velocity

We present a cold atomic beam source based on a two-dimensional (2D)+ magneto-optical trap (MOT), capable of generating a continuous cold beam of 87Rb atoms with a flux up to 4.3*10^9 atoms/s, a mean velocity of 10.96(2.20) m/s, and a transverse temperature of 16.90(1.56) uK. Investigating the influence of high cooling laser intensity, we observe a significant population loss of atoms to hyperfine-level dark states. To account for this, we employ a multiple hyperfine level model to calculate the cooling efficiency associated with the population in dark states, subsequently modifying the scattering force. Simulations of beam flux at different cooling and repumping laser intensities using the modified scattering force are in agreement with experimental results. Optimizing repumping and cooling intensities enhances the flux by 50%. The influence of phase modulation on both the pushing and cooling lasers is experimentally studied, revealing that the mean velocity of cold atoms can be tuned from 9.5 m/s to 14.6 m/s with a phase-modulated pushing laser. The versatility of this continuous beam source, featuring high flux, controlled velocity, and narrow transverse temperature, renders it valuable for applications in atom interferometers and clocks, ultimately enhancing bandwidth, sensitivity, and signal contrast in these devices.

physics.atom-ph↗

arXiv2022

Temporal Convolution Domain Adaptation Learning for Crops Growth Prediction

Existing Deep Neural Nets on crops growth prediction mostly rely on availability of a large amount of data. In practice, it is difficult to collect enough high-quality data to utilize the full potential of these deep learning models. In this paper, we construct an innovative network architecture based on domain adaptation learning to predict crops growth curves with limited available crop data. This network architecture overcomes the challenge of data availability by incorporating generated data from the developed crops simulation model. We are the first to use the temporal convolution filters as the backbone to construct a domain adaptation network architecture which is suitable for deep learning regression models with very limited training data of the target domain. We conduct experiments to test the performance of the network and compare our proposed architecture with other state-of-the-art methods, including a recent LSTM-based domain adaptation network architecture. The results show that the proposed temporal convolution-based network architecture outperforms all benchmarks not only in accuracy but also in model size and convergence rate.

cs.LG↗