Taming Binarized Neural Networks and Mixed-Integer Programs

被引:0
|
作者
Aspman, Johannes [1 ]
Korpas, Georgios [1 ,2 ]
Marecek, Jakub [1 ]
机构
[1] Czech Tech Univ, Dept Comp Sci, Prague, Czech Republic
[2] HSBC Holdings, HSBC Lab, Innovat & Ventures, London, England
关键词
FIELD;
D O I
暂无
中图分类号
TP18 [人工智能理论];
学科分类号
081104 ; 0812 ; 0835 ; 1405 ;
摘要
There has been a great deal of recent interest in binarized neural networks, especially because of their explainability. At the same time, automatic differentiation algorithms such as back-propagation fail for binarized neural networks, which limits their applicability. We show that binarized neural networks admit a tame representation by reformulating the problem of training binarized neural networks as a subadditive dual of a mixed-integer program, which we show to have nice properties. This makes it possible to use the framework of Bolte et al. for implicit differentiation, which offers the possibility for practical implementation of backpropagation in the context of binarized neural networks. This approach could also be used for a broader class of mixed-integer programs, beyond the training of binarized neural networks, as encountered in symbolic approaches to AI and beyond.
引用
收藏
页码:10935 / 10943
页数:9
相关论文
共 50 条
  • [31] Coefficient strengthening: a tool for reformulating mixed-integer programs
    Kent Andersen
    Yves Pochet
    Mathematical Programming, 2010, 122 : 121 - 154
  • [32] SOLVING MIXED-INTEGER NONLINEAR PROGRAMS BY OUTER APPROXIMATION
    FLETCHER, R
    LEYFFER, S
    MATHEMATICAL PROGRAMMING, 1994, 66 (03) : 327 - 349
  • [33] Global solution of nonlinear mixed-integer bilevel programs
    Alexander Mitsos
    Journal of Global Optimization, 2010, 47 : 557 - 582
  • [34] Mixed-integer nonlinear programs featuring “on/off” constraints
    Hassan Hijazi
    Pierre Bonami
    Gérard Cornuéjols
    Adam Ouorou
    Computational Optimization and Applications, 2012, 52 : 537 - 558
  • [35] Distances between optimal solutions of mixed-integer programs
    Joseph Paat
    Robert Weismantel
    Stefan Weltge
    Mathematical Programming, 2020, 179 : 455 - 468
  • [36] On Robustness of Mixed-Integer reformulations of Generalized Disjunctive Programs
    Bogataj, Milos
    Kravanja, Zdravko
    29TH EUROPEAN SYMPOSIUM ON COMPUTER AIDED PROCESS ENGINEERING, PT A, 2019, 46 : 1117 - 1122
  • [37] Distances between optimal solutions of mixed-integer programs
    Paat, Joseph
    Weismantel, Robert
    Weltge, Stefan
    MATHEMATICAL PROGRAMMING, 2020, 179 (1-2) : 455 - 468
  • [38] Distributed Solving of Mixed-Integer Programs with GLPK and Thrift
    Gurski, Frank
    Rethmann, Jochen
    OPERATIONS RESEARCH PROCEEDINGS 2016, 2018, : 599 - 605
  • [39] Solving hard mixed-integer programs for electricity generation
    Ceria, S
    NEXT GENERATION OF ELECTRIC POWER UNIT COMMITMENT MODELS, 2001, 36 : 153 - 166
  • [40] Extending the QCR method to general mixed-integer programs
    Alain Billionnet
    Sourour Elloumi
    Amélie Lambert
    Mathematical Programming, 2012, 131 : 381 - 401