Journals Articles Conferences News About CLP
Submit a paper Email Alert
Search
Search
Articles
Journals
News
Advanced Search
Top Searches
2D MaterialsTransformation opticsQuantum PhotonicsSmall lasersSemiconductor UV PhotonicsSupersymmetric microring laser arrays

Acta Optica Sinica, Volume. 38, Issue 6, 0620002(2018)

An Neural Network Framework of Self-Learning Uncertainty

Hanqing Sun* and Yanwei Pang
Author Affiliations
  • School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China
    • show less
    AbstractGet PDF(in Chinese)
    Figures&Tables (11)
    Equations (0)
    References (29)
    Cited By (2)
    Tools
    Paper Information
    Topics
    Figures & Tables(11)
    Architecture of ULNN

    Fig. 1. Architecture of ULNN

    Download full size
    (a) Input data construction procedure of ULNN-rep; (b) input data construction procedure of ULNN-aug

    Fig. 2. (a) Input data construction procedure of ULNN-rep; (b) input data construction procedure of ULNN-aug

    Download full size
    Components of an UL layer

    Fig. 3. Components of an UL layer

    Download full size
    Curves of uncertainty learning with different βs (on CIFAR-10, vertical-axis is a logarithmic coordinate)

    Fig. 4. Curves of uncertainty learning with different βs (on CIFAR-10, vertical-axis is a logarithmic coordinate)

    Download full size
    Curves of UL (CIFAR-10)

    Fig. 5. Curves of UL (CIFAR-10)

    Download full size

    • Table 1. Comparison between ULNN-rep and ULNN-aug

      View table

      Table 1. Comparison between ULNN-rep and ULNN-aug

      CIFAR-10CIFAR-100
      AccuracyUncertaintyAccuracyUncertainty
      DenseNet-rep91.6%0.035668.3%0.1166
      ULNN-rep92.4%0.001069.3%0.0066
      ULNN-aug94.3%0.082874.2%0.1480

    • Table 2. Impact of input repetition number nu

      View table

      Table 2. Impact of input repetition number nu

      nuCIFAR-10CIFAR-100
      AccuracyUncertaintyAccuracyUncertainty
      294.8%0.057975.9%0.0753
      394.9%0.065975.6%0.0762
      495.0%0.071775.7%0.0743
      594.8%0.076175.6%0.0740

    • Table 3. Impact of UL weight β

      View table

      Table 3. Impact of UL weight β

      βCIFAR-10CIFAR-100
      1.5(=nu/2)0.07180.1076
      10.08280.1480
      0.10.21070.8285
      0.010.69362.4774

    • Table 4. Impact of Dropout on ULNN-aug

      View table

      Table 4. Impact of Dropout on ULNN-aug

      Dropout ratioCIFAR-10CIFAR-100
      AccuracyUncertaintyAccuracyUncertainty
      0.294.8%0.075775.2%0.1417
      0.194.9%0.068775.6%0.1077
      0.0595.1%0.057975.9%0.0753

    • Table 5. Comparison of validation result with original training parameters in DenseNet

      View table

      Table 5. Comparison of validation result with original training parameters in DenseNet

      CIFAR-10CIFAR-100
      AccuracyUncertaintyAccuracyUncertainty
      DenseNet + data aug.94.8%1.519075.6%1.3564
      ULNN-aug95.1%0.057975.9%0.0753

    • Table 6. Semantic segmentation results of ULNN SegNet on CamVid

      View table

      Table 6. Semantic segmentation results of ULNN SegNet on CamVid

      AccuracymIoUUncertainty
      Bayesian SegNet85.9%51.4%452.7
      ULNN SegNet87.1%52.0%50.63
    Tools

    Get Citation

    Copy Citation Text

    Hanqing Sun, Yanwei Pang. An Neural Network Framework of Self-Learning Uncertainty[J]. Acta Optica Sinica, 2018, 38(6): 0620002

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Set citation alerts for article
    Save article for my favorites
    Paper Information

    Category: Optics in Computing

    Received: Dec. 29, 2017

    Accepted: --

    Published Online: Jul. 9, 2018

    The Author Email: Sun Hanqing (HQSun@tju.edu.cn)

    DOI:10.3788/AOS201838.0620002

    Topics

    laser devices and laser physics
    Lasers and Laser Optics
    Laser physics
    laser manufacturing
    Instrumentation, Measurement and Metrology