Reframing Neural Networks: Deep Structure in Overcomplete Representations

被引:1
作者
Murdock, Calvin [1 ]
Cazenavette, George [1 ]
Lucey, Simon [2 ]
机构
[1] Carnegie Mellon Univ, Pittsburgh, PA 15213 USA
[2] Univ Adelaide, Adelaide, SA, Australia
基金
美国国家科学基金会;
关键词
Machine learning; vision and scene understanding; deep learning; generalization; overcomplete frames; sparsity; SPARSE; ALGORITHM; RECOVERY;
D O I
10.1109/TPAMI.2022.3149445
中图分类号
TP18 [人工智能理论];
学科分类号
081104 ; 0812 ; 0835 ; 1405 ;
摘要
In comparison to classical shallow representation learning techniques, deep neural networks have achieved superior performance in nearly every application benchmark. But despite their clear empirical advantages, it is still not well understood what makes them so effective. To approach this question, we introduce deep frame approximation: a unifying framework for constrained representation learning with structured overcomplete frames. While exact inference requires iterative optimization, it may be approximated by the operations of a feed-forward deep neural network. We indirectly analyze how model capacity relates to frame structures induced by architectural hyperparameters such as depth, width, and skip connections. We quantify these structural differences with the deep frame potential, a data-independent measure of coherence linked to representation uniqueness and stability. As a criterion for model selection, we show correlation with generalization error on a variety of common deep network architectures and datasets. We also demonstrate how recurrent networks implementing iterative optimization algorithms can achieve performance comparable to their feed-forward approximations while improving adversarial robustness. This connection to the established theory of overcomplete representations suggests promising new directions for principled deep network architecture design with less reliance on ad-hoc engineering.
引用
收藏
页码:964 / 979
页数:16
相关论文
共 65 条
  • [1] Alvarez M, 2016, P ADV NEUR INF PROC
  • [2] Arpit D, 2017, PR MACH LEARN RES, V70
  • [3] NEURAL NETWORKS AND PRINCIPAL COMPONENT ANALYSIS - LEARNING FROM EXAMPLES WITHOUT LOCAL MINIMA
    BALDI, P
    HORNIK, K
    [J]. NEURAL NETWORKS, 1989, 2 (01) : 53 - 58
  • [4] Dictionary Learning for Sparse Coding: Algorithms and Convergence Analysis
    Bao, Chenglong
    Ji, Hui
    Quan, Yuhui
    Shen, Zuowei
    [J]. IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, 2016, 38 (07) : 1356 - 1369
  • [5] Finite normalized tight frames
    Benedetto, JJ
    Fickus, M
    [J]. ADVANCES IN COMPUTATIONAL MATHEMATICS, 2003, 18 (2-4) : 357 - 385
  • [6] Representation Learning: A Review and New Perspectives
    Bengio, Yoshua
    Courville, Aaron
    Vincent, Pascal
    [J]. IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, 2013, 35 (08) : 1798 - 1828
  • [7] Fast Convolutional Sparse Coding
    Bristow, Hilton
    Eriksson, Anders
    Lucey, Simon
    [J]. 2013 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR), 2013, : 391 - 398
  • [8] On the Uniqueness of Nonnegative Sparse Solutions to Underdetermined Systems of Equations
    Bruckstein, Alfred M.
    Elad, Michael
    Zibulevsky, Michael
    [J]. IEEE TRANSACTIONS ON INFORMATION THEORY, 2008, 54 (11) : 4813 - 4820
  • [9] Cao Y, 2019, 33 C NEURAL INFORM P, V32
  • [10] Casazza P., 2009, P INT C SAMPL THEOR