3 présentations

• 
  15h30 - 15h55

  Hyperplane Arrangements and Matroids in Complementarity Problems: the B-differential of the componentwise Minimum

  • Baptiste Plaquevent-Jourdain, prés., Inria Paris - Université de Sherbrooke - Sorbonne Université
  • Jean-Pierre Dussault, Université de Sherbrooke
  • Jean Charles Gilbert, Inria Paris

  Systems of smooth nonlinear equations of the form H(x) = 0 can often be
  solved by the famous Newton method. When the equations involve nonsmooth
  functions, this algorithm is no longer adapted. The semismooth Newton
  method is a possible remedy to overcome this difficulty: it yields fast
  local convergence, provided the B-differential of H has only nonsingular
  Jacobians at the considered solution.

  An example of nonsmooth vector function H is encountered when
  complementarity problems are reformulated as nonsmooth equations
  thanks to the componentwise minimum function. This talk describes
  the content and properties of the B-differential of the componentwise
  minimum of two affine vector functions.

  Describing and computing this B-differential is a problem with
  connections to linear algebra, convex anaysis and discrete geometry
  (hyperplane arrangement), the latter being the approach chosen numerically.
  We shall emphasize the role of duality and matroid circuits in these links,
  and show how insight acquired by the theory serves the numerical aspect.

  The presentation will show significant improvements on the state of the art
  approach by Rada and Černý that solves numerically this combinatorial
  problem.
• 
  15h55 - 16h20

  PartiallySeparableNLPModels.jl: A Julia framework for partitioned quasi-Newton optimization

  • Paul Raynaud, prés., GERAD and Polytechnique Montréal
  • Dominique Orban, GERAD - Polytechnique Montréal
  • Jean Bigeon, Univ. Grenoble Alpes, CNRS, Grenoble INP, G-SCOP, F-38000 Grenoble, France

  In optimization applications, objectives and constraints often have a partially-separable structure: they are a sum of so-called element functions that depend on a small number of variables.
  It is well known that twice-differentiable functions with a sparse Hessian are partially-separable.
  We present a framework based on the JuliaSmoothOptimizers ecosystem to model and detect partially-separable problems, and to efficiently evaluate their derivatives by informing AD engines of their structure.
  We also implement partitioned trust-region quasi-Newton methods in which Hessian approximations are also informed by the partially-separable structure.
  Such methods were proposed in the 1980s and have rivaled limited-memory quasi-Newton methods in efficiency since.
  Our implementations feature novel aspects, including the automatic detection of convexity and its impact on the choice of a quasi-Newton approximation, and the use of limited-memory approximations for element functions that depend on a sufficiently large number of variables, which consumes significantly less memory than the dense matrices used in the 1980s and 1990s.
  Our numerical results illustrate the performance of a number of variants on a standard problem collection and demonstrate the viability of our methods for problems in which element functions depend on both a small or a substantial number of variables.
• 
  16h20 - 16h45

  Quadratic Regularization Optimizer in Low Precision for Deep Neural Networks: Implementation and Numerical Experience

  • Farhad Rahbarnia, prés., GERAD Student
  • Dominique Orban, GERAD - Polytechnique Montréal
  • Paul Raynaud, GERAD
  • Nathan Allaire, GERAD
  • Dominique Monet, GERAD

  We propose a Julia framework to automatically port a deterministic optimization method to the stochastic setting for training deep neural networks (DNNs). We illustrate the process using the quadratic regularization method R2, which is a variant of the gradient method with adaptive step size. We report on numerical experiments using the stochastic R2 method on classification networks in low-precision arithmetic.