GradVis: Visualization and Second Order Analysis of Optimization Surfaces during the Training of Deep Neural Networks

被引:2
|
作者
Chatzimichailidis, Avraam [1 ,3 ]
Pfreundt, Franz-Josef [1 ,2 ]
Gauger, Nicolas R. [3 ]
Keuper, Janis [1 ,4 ]
机构
[1] Fraunhofer ITWM, Competence Ctr High Performance Comp, Kaiserslautern, Germany
[2] Fraunhofer Ctr Machine Learning, Berlin, Germany
[3] TU Kaiserslautern, Chair Sci Comp, Kaiserslautern, Germany
[4] Offenburg Univ, Inst Machine Learning & Analyt, Offenburg, Germany
关键词
visualization; eigenvalues; parallelization; deeplearning; second-order;
D O I
10.1109/MLHPC49564.2019.00012
中图分类号
TP18 [人工智能理论];
学科分类号
081104 ; 0812 ; 0835 ; 1405 ;
摘要
Current training methods for deep neural networks boil down to very high dimensional and non-convex optimization problems which are usually solved by a wide range of stochastic gradient descent methods. While these approaches tend to work in practice, there are still many gaps in the theoretical understanding of key aspects like convergence and generalization guarantees, which are induced by the properties of the optimization surface (loss landscape). In order to gain deeper insights, a number of recent publications proposed methods to visualize and analyze the optimization surfaces. However, the computational cost of these methods are very high, making it hardly possible to use them on larger networks. In this paper, we present the GradVis Toolbox, an open source library for efficient and scalable visualization and analysis of deep neural network loss landscapes in Tensorflow and PyTorch. Introducing more efficient mathematical formulations and a novel parallelization scheme, GradVis allows to plot 2d and 3d projections of optimization surfaces and trajectories, as well as high resolution second order gradient information for large networks.
引用
收藏
页码:66 / 74
页数:9
相关论文
共 50 条
  • [1] Second-order Derivative Optimization Methods in Deep Learning Neural Networks
    Lim, Si Yong
    Lim, King Hann
    [J]. 2022 INTERNATIONAL CONFERENCE ON GREEN ENERGY, COMPUTING AND SUSTAINABLE TECHNOLOGY (GECOST), 2022, : 470 - 475
  • [2] DNNViz: Training Evolution Visualization for Deep Neural Networks
    Clavien, Gil
    Alberti, Michele
    Pondenkandath, Vinaychandran
    Ingold, Rolf
    Liwicki, Marcus
    [J]. 2019 6TH SWISS CONFERENCE ON DATA SCIENCE (SDS), 2019, : 19 - 24
  • [3] An Optimization Strategy for Deep Neural Networks Training
    Wu, Tingting
    Zeng, Peng
    Song, Chunhe
    [J]. 2022 INTERNATIONAL CONFERENCE ON IMAGE PROCESSING, COMPUTER VISION AND MACHINE LEARNING (ICICML), 2022, : 596 - 603
  • [4] An Efficient Optimization Technique for Training Deep Neural Networks
    Mehmood, Faisal
    Ahmad, Shabir
    Whangbo, Taeg Keun
    [J]. MATHEMATICS, 2023, 11 (06)
  • [5] MULTICOMPOSITE NONCONVEX OPTIMIZATION FOR TRAINING DEEP NEURAL NETWORKS
    Cui, Ying
    He, Ziyu
    Pang, Jong-Shi
    [J]. SIAM JOURNAL ON OPTIMIZATION, 2020, 30 (02) : 1693 - 1723
  • [6] A Second Order Training Algorithm for Multilayer Feedforward Neural Networks
    谭营
    何振亚
    邓超
    [J]. Journal of Southeast University(English Edition), 1997, (01) : 32 - 36
  • [7] Visualization in Deep Neural Network Training
    Kollias, Stefanos
    [J]. INTERNATIONAL JOURNAL ON ARTIFICIAL INTELLIGENCE TOOLS, 2022, 31 (03)
  • [8] TRAINING DEEP NEURAL NETWORKS VIA OPTIMIZATION OVER GRAPHS
    Zhang, Guoqiang
    Kleijn, W. Bastiaan
    [J]. 2018 IEEE INTERNATIONAL CONFERENCE ON ACOUSTICS, SPEECH AND SIGNAL PROCESSING (ICASSP), 2018, : 4119 - 4123
  • [9] ASYNCHRONOUS STOCHASTIC OPTIMIZATION FOR SEQUENCE TRAINING OF DEEP NEURAL NETWORKS
    Heigold, Georg
    McDermott, Erik
    Vanhoucke, Vincent
    Senior, Andrew
    Bacchiani, Michiel
    [J]. 2014 IEEE INTERNATIONAL CONFERENCE ON ACOUSTICS, SPEECH AND SIGNAL PROCESSING (ICASSP), 2014,
  • [10] Generalized backpropagation algorithm for training second-order neural networks
    Fan, Fenglei
    Cong, Wenxiang
    Wang, Ge
    [J]. INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN BIOMEDICAL ENGINEERING, 2018, 34 (05)