5 research outputs found

    Faster Rates for Compressed Federated Learning with Client-Variance Reduction

    No full text
    Due to the communication bottleneck in distributed and federated learning applications, algorithms using communication compression have attracted significant attention and are widely used in practice. Moreover, the huge number, high heterogeneity and limited availability of clients result in high client-variance. This paper addresses these two issues together by proposing compressed and client-variance reduced methods COFIG and FRECON. We prove an O((1+ω)3/2NSϵ2+(1+ω)N2/3Sϵ2)O(\frac{(1+\omega)^{3/2}\sqrt{N}}{S\epsilon^2}+\frac{(1+\omega)N^{2/3}}{S\epsilon^2}) bound on the number of communication rounds of COFIG in the nonconvex setting, where NN is the total number of clients, SS is the number of clients participating in each round, ϵ\epsilon is the convergence error, and ω\omega is the variance parameter associated with the compression operator. In case of FRECON, we prove an O((1+ω)NSϵ2)O(\frac{(1+\omega)\sqrt{N}}{S\epsilon^2}) bound on the number of communication rounds. In the convex setting, COFIG converges within O((1+ω)NSϵ)O(\frac{(1+\omega)\sqrt{N}}{S\epsilon}) communication rounds, which, to the best of our knowledge, is also the first convergence result for compression schemes that do not communicate with all the clients in each round. We stress that neither COFIG nor FRECON needs to communicate with all the clients, and they enjoy the first or faster convergence results for convex and nonconvex federated learning in the regimes considered. Experimental results point to an empirical superiority of COFIG and FRECON over existing baselines.Comment: Accepted by SIAM Journal on Mathematics of Data Science (SIMODS

    Teaching Virtual Characters to use Body Language

    No full text
    Non-verbal communication, or “body language”, is a critical component in constructing believable virtual characters. Most often, body language is implemented by a set of ad-hoc rules.We propose a new method for authors to specify and refine their character’s body-language responses. Using our method, the author watches the character acting in a situation, and provides simple feedback on-line. The character then learns to use its body language to maximize the rewards, based on a reinforcement learning algorithm

    DOUBLE RESONANCE EXCITATION OF THE RUBIDIUM DIMER : THE 2 1Πg^{1}\Pi_g STATE

    No full text
    Author Institution: Department of Chemistry, Moscow State University, 119991; Moscow, Russia\; Institut Lumiere Matiere, Universite Lyon 1 \& CNRS UMR5306, Universite de Lyon, FranceWe have performed a series of optical-optical double resonance experiments with one or two cw Ti:sapphire lasers, to excite the 2~1Πg^{1}\Pi_g state of Rb2_2, recording infrared fluorescence from 2~1Πg^{1}\Pi_g on a Fourier transform spectrometer. Fluorescence from the lower vibrational levels of 2~1Πg^{1}\Pi_g (Te_e = 22069.56 cm1^{-1}) is dominated by transitions to the B 1Πu^{1}\Pi_u state studied by Amiot and Verges, Chem. Phys. Lett. 294, 91-98 (1997). Vibrational and rotational relaxation from laser-pumped levels v' 35, occurs also to the 0+^+ components of the A~1Σu+^{1}\Sigma_u^+ \sim b 3Πu~^{3}\Pi_u complex. Fitting all available 2~1Πg^{1}\Pi_g \rightarrow B~1Πu^{1}\Pi_u data for 85^{85}Rb2_2 and 85^{85}Rb87^{87}Rb (several thousand transitions) has also given an improved description of the bottom of the B 1Πu^{1}\Pi_u state potential well. The 2~1Πg^{1}\Pi_g state correlates at long-range with Rb 5s + Rb 4d 2D3/2^2D_{3/2} atoms (A.-R. Allouche, M. Aubert-Frecon, J. Chem Phys 136, 37-41 (2012)), giving a dissociation energy of 1279.6 cm1^{-1}. Most new data lie below v = 45, 250 cm1^{-1} below this dissociation threshold

    THE 43Δg4^{3} \Delta_{g} STATE IN K2K_{2} - INVESTIGATING A POSSIBLE GATEWAY TO CORE NON-PENETRATING RYDBERG STATES

    No full text
    Author Institution: Department of Physics, United States Military Academy; Department of Physics, Temple University; Department of Chemistry, University of Illinois at Chicago; Key Lab Atom and Molecular Nanoscience, Tsinghua University; Laboratoire de Physique des Atomes, Lasers, Mol\'{e}cules et Surfaces, (PALMES), CNRS et Universit\'{e}; Laboratoire de Spectrom\'etrie Ionique et Moleculaire (L.A.S.I.M), CNRS et Universit\'e Lyon (UMR5579)Core non-penetrating Rydberg states can give useful information on the electronic structure of the ion core; however, core non-penetrating states are difficult to observe since these states hardly penetrate the more accessible ion core and the electronic angular momentum quantum number, l, is large, for the core non-penetrating states thus the transition dipole moment to the core non-penetrating states is small. The core penetrating 43Δg4^{3} \Delta_{g} state (atomic limit: 4s+5d4s+5d) and the core non-penetrating 33Δg3^{3}\Delta_{g} state (atomic limit: 4s+4f4s+4f) perturb each other since they have the same symmetry and overlapping energy states thus creating the possibility of a gateway to other core non-penetrating states
    corecore