Corpus ID: 3647043

Learning One Convolutional Layer with Overlapping Patches

@article{Goel2018LearningOC,
  title={Learning One Convolutional Layer with Overlapping Patches},
  author={Surbhi Goel and Adam R. Klivans and Raghu Meka},
  journal={ArXiv},
  year={2018},
  volume={abs/1802.02547}
}
  • Surbhi Goel, Adam R. Klivans, Raghu Meka
  • Published 2018
  • Computer Science, Mathematics
  • ArXiv
  • We give the first provably efficient algorithm for learning a one hidden layer convolutional network with respect to a general class of (potentially overlapping) patches. Additionally, our algorithm requires only mild conditions on the underlying distribution. We prove that our framework captures commonly used schemes from computer vision, including one-dimensional and two-dimensional "patch and stride" convolutions. Our algorithm-- $Convotron$ -- is inspired by recent work applying isotonic… CONTINUE READING

    Citations

    Publications citing this paper.
    SHOWING 1-10 OF 41 CITATIONS

    On the Learning Dynamics of Two-layer Nonlinear Convolutional Neural Networks

    VIEW 3 EXCERPTS
    CITES BACKGROUND

    Improved Learning of One-hidden-layer Convolutional Neural Networks with Overlaps

    VIEW 19 EXCERPTS
    CITES METHODS

    Learning Two-layer Neural Networks with Symmetric Inputs

    VIEW 5 EXCERPTS
    CITES BACKGROUND & METHODS
    HIGHLY INFLUENCED

    Tight Sample Complexity of Learning One-hidden-layer Convolutional Neural Networks

    VIEW 18 EXCERPTS
    CITES BACKGROUND & METHODS
    HIGHLY INFLUENCED

    Guaranteed Recovery of One-Hidden-Layer Neural Networks via Cross Entropy

    VIEW 1 EXCERPT

    How Many Samples are Needed to Learn a Convolutional Neural Network?

    VIEW 2 EXCERPTS
    CITES BACKGROUND

    A Convergence Theory for Deep Learning via Over-Parameterization

    VIEW 16 EXCERPTS
    CITES BACKGROUND
    HIGHLY INFLUENCED

    Local Geometry of Cross Entropy Loss in Learning One-Hidden-Layer Neural Networks

    VIEW 1 EXCERPT

    How Many Samples are Needed to Estimate a Convolutional Neural Network?

    VIEW 2 EXCERPTS
    CITES BACKGROUND

    FILTER CITATIONS BY YEAR

    2018
    2020

    CITATION STATISTICS

    • 5 Highly Influenced Citations

    • Averaged 14 Citations per year from 2018 through 2020

    References

    Publications referenced by this paper.