arXiv subjects

Kevin Hu

Publications and source records attributed to Kevin Hu.

12 recordsLinked to original sources

arXiv2026

Quantitative propagation of chaos and universality for asymmetric Langevin spin glass dynamics

We obtain quantitative estimates on quenched propagation of chaos for Langevin spin glass dynamics with i.i.d. disorder. Prior work in the case of Gaussian disorder established the qualitative convergence of the law of a single spin to a deterministic McKean-Vlasov limit. We prove convergence rates in expected Wasserstein distance and quantitative concentration rates for Lipschitz observables under the assumption that the disorder satisfies the T2 inequality. The proof uses a coupling argument, together with techniques from concentration of measure, filtering theory, and Malliavin calculus

math.PR↗

arXiv2025

fabric-lib: RDMA Point-to-Point Communication for LLM Systems

Emerging Large Language Model (LLM) system patterns, such as disaggregated inference, Mixture-of-Experts (MoE) routing, and asynchronous reinforcement fine-tuning, require flexible point-to-point communication beyond simple collectives. Existing implementations are locked to specific Network Interface Controllers (NICs), hindering integration into inference engines and portability across hardware providers. We present fabric-lib, which bridges the functionality of common NICs to expose a uniform interface. fabric-lib exposes one-sided WriteImm operations with a ImmCounter primitive for completion notification, without ordering assumptions of network transport, transparently managing multiple NICs per GPU. We demonstrate peak throughput of 400 Gbps on both NVIDIA ConnectX-7 and AWS Elastic Fabric Adapter (EFA). We showcase fabric-lib through three production systems: (1) KvCache transfer for disaggregated inference with dynamic scaling, (2) RL weight updates achieving 1.3 seconds for trillion-parameter models, and (3) MoE dispatch/combine implementation exceeding DeepEP decode latency on ConnectX-7, with the first viable latencies on EFA. We demonstrate that our portable point-to-point communication complements collectives while avoiding lock-in. fabric-lib is open-sourced at https://github.com/perplexityai/pplx-garden/

cs.DC↗

arXiv2025

A case study of the long-time behavior of the Gaussian local-field equation

For any integer $\kappa \geq 2$, the $\kappa$-local-field equation ($\kappa$-LFE) characterizes the limit of the neighborhood path empirical measure of interacting diffusions on $\kappa$-regular random graphs, as the graph size goes to infinity. It has been conjectured that the long-time behavior of the (in general non-Markovian) $\kappa$-LFE coincides with that of a certain more tractable Markovian analog, the Markov $\kappa$-local-field equation. In the present article, we prove this conjecture for the case when $\kappa = 2$ and the diffusions are one-dimensional with affine drifts. As a by-product of our proof, we also show that for interacting diffusions on the $n$-cycle (or 2-regular random graph on $n$ vertices), the limits $n \rightarrow \infty$ and $t\rightarrow \infty$ commute. Along the way, we also establish well-posedness of the Markov $\kappa$-local field equations with affine drifts for all $\kappa \geq 2$, which may be of independent interest.

math.PR↗

arXiv2024

An H-theorem for a conditional McKean-Vlasov process related to interacting diffusions on regular trees

We study the long-time behavior of the $\kappa$-Markov local-field equation ($\kappa$-MLFE), which is a conditional McKean-Vlasov equation associated with interacting diffusions on the $\kappa$-regular tree. Under suitable assumptions on the coefficients, we prove well-posedness of the $\kappa$-MLFE. We also establish an H-theorem by identifying an energy functional, referred to as the sparse free energy, whose derivative along the measure flow of the $\kappa$-MLFE is given by a nonnegative functional that can be viewed as a modified Fisher information. Moreover, we show that the zeros of the latter functional coincide with the set of stationary distributions of the $\kappa$-MLFE and are also marginals of splitting Gibbs measures on the $\kappa$-regular tree. Furthermore, we show that for a natural class of initial conditions, the corresponding measure flow converges to one of the stationary distributions, thus demonstrating that the sparse free energy acts as a global Lyapunov function. Under mild additional conditions, in the case $\kappa = 2$ we prove that the sparse free energy arises naturally as the renormalized limit of certain relative entropies. We exploit this characterization to prove a modified logarithmic Sobolev inequality and establish an exponential rate of convergence of the $2$-MLFE measure flow to its unique stationary distribution.

math.PR↗

arXiv2024

The fundamental martingale with applications to Markov Random Fields

We consider collections of SDEs indexed by a graph. Each SDE is driven by an additive Gaussian noise and each drift term interacts with all other SDEs within the graph neighbourhood. We derive the fundamental martingale for a class of Gaussian processes and use this to prove a Girsanov type theorem. Further, we use this to construct a clique factorisation to prove that the law of the interacting SDEs forms a 2-Markov Random Field.

math.PR↗

arXiv2024

A Mimicking Theorem for processes driven by fractional Brownian motion

In this paper, we prove a mimicking theorem for stochastic processes with an additive Gaussian noise along with some entropy and transport type estimates. As an application of these results, we prove sharp quantitative propagation of chaos result and derive a formula for the marginal dynamics of collections of locally interacting stochastic differential equations with additive Gaussian noise.

math.PR↗

arXiv2022

Supply Chain Characteristics as Predictors of Cyber Risk: A Machine-Learning Assessment

This paper provides the first large-scale data-driven analysis to evaluate the predictive power of different attributes for assessing risk of cyberattack data breaches. Furthermore, motivated by rapid increase in third party enabled cyberattacks, the paper provides the first quantitative empirical evidence that digital supply-chain attributes are significant predictors of enterprise cyber risk. The paper leverages outside-in cyber risk scores that aim to capture the quality of the enterprise internal cybersecurity management, but augment these with supply chain features that are inspired by observed third party cyberattack scenarios, as well as concepts from network science research. The main quantitative result of the paper is to show that supply chain network features add significant detection power to predicting enterprise cyber risk, relative to merely using enterprise-only attributes. Particularly, compared to a base model that relies only on internal enterprise features, the supply chain network features improve the out-of-sample AUC by 2.3\%. Given that each cyber data breach is a low probability high impact risk event, these improvements in the prediction power have significant value. Additionally, the model highlights several cybersecurity risk drivers related to third party cyberattack and breach mechanisms and provides important insights as to what interventions might be effective to mitigate these risks.

q-fin.RM↗

arXiv2021

Evolutionary Dynamics of Gig Economy Labor Strategies under Technology, Policy and Market Influence

The emergence of the modern gig economy introduces a new set of employment considerations for firms and laborers that include various trade-offs. With a game-theoretical approach, we examine the influences of technology, policy and markets on firm and worker preferences for gig labor. Theoretically, we present a new extension to the replicator equation and model oscillating dynamics in two-player asymmetric bi-matrix games with time-evolving environments, introducing concepts of the attractor arc, trapping zone and escape. We demonstrate how changing market conditions result in distinct evolutionary patterns for gig-labor preferences across high and low skill work-forces, which we explain through their differing sensitivities to market-driven consumer demand and financial incentives among other considerations. Informing tensions regarding the future of this new employment category, we present a novel payoff framework to analyze the role of technology on the growth of the gig economy. Finally, we explore regulatory implications within the gig economy, demonstrating how intervals of lenient and strict policy alter firm and worker sensitivities between gig and employee labor strategies.

physics.soc-ph↗

arXiv2019

Sherlock: A Deep Learning Approach to Semantic Data Type Detection

Correctly detecting the semantic type of data columns is crucial for data science tasks such as automated data cleaning, schema matching, and data discovery. Existing data preparation and analysis systems rely on dictionary lookups and regular expression matching to detect semantic types. However, these matching-based approaches often are not robust to dirty data and only detect a limited number of types. We introduce Sherlock, a multi-input deep neural network for detecting semantic types. We train Sherlock on $686,765$ data columns retrieved from the VizNet corpus by matching $78$ semantic types from DBpedia to column headers. We characterize each matched column with $1,588$ features describing the statistical properties, character distributions, word embeddings, and paragraph vectors of column values. Sherlock achieves a support-weighted F$_1$ score of $0.89$, exceeding that of machine learning baselines, dictionary and regular expression benchmarks, and the consensus of crowdsourced annotations.

cs.LG↗

arXiv2019

VizNet: Towards A Large-Scale Visualization Learning and Benchmarking Repository

Researchers currently rely on ad hoc datasets to train automated visualization tools and evaluate the effectiveness of visualization designs. These exemplars often lack the characteristics of real-world datasets, and their one-off nature makes it difficult to compare different techniques. In this paper, we present VizNet: a large-scale corpus of over 31 million datasets compiled from open data repositories and online visualization galleries. On average, these datasets comprise 17 records over 3 dimensions and across the corpus, we find 51% of the dimensions record categorical data, 44% quantitative, and only 5% temporal. VizNet provides the necessary common baseline for comparing visualization design techniques, and developing benchmark models and algorithms for automating visual analysis. To demonstrate VizNet's utility as a platform for conducting online crowdsourced experiments at scale, we replicate a prior study assessing the influence of user task and data distribution on visual encoding effectiveness, and extend it by considering an additional task: outlier detection. To contend with running such studies at scale, we demonstrate how a metric of perceptual effectiveness can be learned from experimental results, and show its predictive power across test datasets.

cs.HC↗

arXiv2015

Pantheon 1.0, a manually verified dataset of globally famous biographies

We present the Pantheon 1.0 dataset: a manually verified dataset of individuals that have transcended linguistic, temporal, and geographic boundaries. The Pantheon 1.0 dataset includes the 11,341 biographies present in more than 25 languages in Wikipedia and is enriched with: (i) manually verified demographic information (place and date of birth, gender) (ii) a taxonomy of occupations classifying each biography at three levels of aggregation and (iii) two measures of global popularity including the number of languages in which a biography is present in Wikipedia (L), and the Historical Popularity Index (HPI) a metric that combines information on L, time since birth, and page-views (2008-2013). We compare the Pantheon 1.0 dataset to data from the 2003 book, Human Accomplishments, and also to external measures of accomplishment in individual games and sports: Tennis, Swimming, Car Racing, and Chess. In all of these cases we find that measures of popularity (L and HPI) correlate highly with individual accomplishment, suggesting that measures of global popularity proxy the historical impact of individuals.

physics.soc-ph↗

arXiv2013

Characteristic Direction Approach to Identify Differentially Expressed Genes

Genome-wide gene expression profiles, as measured with microarrays or RNA-Seq experiments, have revolutionized biological and biomedical research by providing a quantitative measure of the entire mRNA transcriptome. Typically, researchers set up experiments where control samples are compared to a treatment condition, and using the t-test they identify differentially expressed genes upon which further analysis and ultimately biological discovery from such experiments is based. Here we describe an alternative geometrical approach to identify differentially expressed genes. We show that this alternative method, called the Characteristic Direction, is capable of identifying more relevant genes. We evaluate our approach in three case studies. In the first two, we match transcription factor targets determined by ChIP-seq profiling with differentially expressed genes after the same transcription factor knockdown or over-expression in mammalian cells. In the third case study, we evaluate the quality of enriched terms when comparing normal epithelial cells with cancer stem cells. In conclusion, we demonstrate that the Characteristic Direction approach is much better in calling the significantly differentially expressed genes and should replace the widely currently in used t-test method for this purpose. Implementations of the method in MATLAB, Python and Mathematica are available at: http://www.maayanlab.net/CD.

stat.AP↗