Provide the outlines of the paper

[kayhan-batmanghelich]

write the outlines in the following format:

Introduction

• introduce classification task with different data settings ordinary setting, weak supervision, noisy label, and complementary label learning.
• introduce recent generative models then gans and conditional gans.
• the reason why we apply conditional generative model on complementary learning.
  • the difficulties of proposed complementary gans.
  • semi-supervised learning on unlabeled data helps complementary learning.
• the results of experiments support our method.
• contribution paragraph:
  • We are the first method doing complementary in semi-supervised setting
  • We developed a new variant of conditional GAN that can use complementary information
  • We provide preliminary theoretical supporting under which the true conditional model is identifiable using complementary data

Method

Notation

Background

• cGAN:
  • classical gans formulation
  • conditional gans formulation
• complementary learning:
  • general method of complementary learning
  • simply introduce the method of transition matrix for mapping true label learning and complementary label learning.

Proposed method

• Complementary gans
  • the whole idea of how to generate true label data through only given complementary labels
  • the graph of complementary gans model
  • formulation of complementary gans
  • infer that proposed conditional model can match the true class conditioned data (refer to {https://papers.nips.cc/paper/8229-robustness-of-conditional-gans-to-noisy-labels.pdf})
• Semi-Supervised Learning with unlabeled data
  • cannot prove why it works mathematically but we can draw a figure to intuitively explain it. (improve estimation of p(x) will improve p(x|y) =p(x,y)/p(x))

Related Work

Experiments

• Mnist
  • training settings
  • generative examples vs true examples
  • ordinary acc vs complementary acc
  • semi complementary acc vs complementary acc
• Cifar 10
  • training settings
  • generative examples vs true examples
  • ordinary acc vs complementary acc
  • semi complementary acc vs complementary acc
• VGG face 100
  • training settings
  • generative examples vs true examples
  • ordinary acc vs complementary acc
  • semi complementary acc vs complementary acc

Ablation study and experiments

• complementary learning converges speed vs semi complementary learning converge speed (semi setting improves the converge speed by a big gap)
• the influence of the weight between class condition loss and adversarial loss

[kayhan-batmanghelich]

@xuyanwu this is not what I want. I asked you to complete bullet point not to write sentences or equations.

Do it again:

• Write the section and subsections
• For each subsection, you are going to have a couple of paragraph. For each paragraph you are going to write, write one bullet point. The bullet point is not the paragraph, it is a short sentence on what you are going to say in the paragraph.
• I want this for method section and experiment section
• No equation

[mgong2]

Hi Yanwu, Please find the attached outline of my paper. Hope it helps. Best, Mingming

…

-------------------------------------------------------------------------------- Mingming Gong Homepage: http:// <http://mingminggong.xyz>mgong2.github.io

On Fri, Mar 1, 2019 at 11:11 PM Kayhan Batmanghelich < ***@***.***> wrote: @xuyanwu <https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fxuyanwu&data=02%7C01%7Cgongm%40pitt.edu%7Cd0cd2d5a41d94277d2ec08d69ec513cc%7C9ef9f489e0a04eeb87cc3a526112fd0d%7C1%7C0%7C636870966617732298&sdata=ZOnkIWyu7YWpMLPWgpxxPWERe2h5NRWDrJTEEMMFAgk%3D&reserved=0> this is not what I want. I asked you to complete bullet point not to write sentences or equations. Do it again: - Write the section and subsections - For each subsection, you are going to have a couple of paragraph. For each paragraph you are going to write, write one bullet point. The bullet point is not the paragraph, it is a short sentence on what you are going to say in the paragraph. - I want this for method section and experiment section - No equation — You are receiving this because you were assigned. Reply to this email directly, view it on GitHub <https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fkayhan-batmanghelich%2FBatmanLab%2Fissues%2F70%23issuecomment-468881992&data=02%7C01%7Cgongm%40pitt.edu%7Cd0cd2d5a41d94277d2ec08d69ec513cc%7C9ef9f489e0a04eeb87cc3a526112fd0d%7C1%7C0%7C636870966617742303&sdata=z3czI5N9GCaZPszazNH1wVtvGDHCfMMeJ8f4bdOUWV4%3D&reserved=0>, or mute the thread <https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fnotifications%2Funsubscribe-auth%2FAdmbnQXMnVbwki9GPF1kZo6KKDpVKN10ks5vSfnSgaJpZM4bZ0-p&data=02%7C01%7Cgongm%40pitt.edu%7Cd0cd2d5a41d94277d2ec08d69ec513cc%7C9ef9f489e0a04eeb87cc3a526112fd0d%7C1%7C0%7C636870966617752312&sdata=8d%2Bn0DoU9LayL1EeP%2BPuJsxkbcjDYRRmbgDno8kvk90%3D&reserved=0> .

[mgong2]

Outline of the unpaired paper

Introduction

• a brief introduction of gene association methods, random effect model, HSIC, multi-phenotype
• the motivation of unpaired data, examples...
• how the proposed method make use of unpaired data
  • improve estimation of null distribution
  • explore low-rank structure of both genotype and phenotype features
• summarise experimental results and draw conclusions

Method

• schematic illustration
• random effect model, univariate phenotype, null distribution
  • show how estimation of null distribution can be immediately improved from unpaired data
• random effect model, multivariate phenotype
  • improve null distribution by better estimation of eigenvalues of phenotype covariance, give a bound
  • derive the new test statistic under low-rank assumption
    • infinite unpaired sample size
      • prove higher power
      • derive null distribution
    • finite unpaired sample size (incomplete)
      • prove xxx
      • derive xxx
• independence test, HSIC, multivariate phenotype
  • same as random effect model
• include covariates, conditional independence test

Experiment

• simulation 1.1

  • random effect model
  • univariate/multivariate phenotype
  • low-rank assumption of phenotype
  • type I , II error

• simulation 1.2

  • HSIC
  • univariate/multivariate phenotype
  • low-rank assumption of phenotype and genotype
  • type I, II error

• simulation 2

  • random effect model
  • genotype from COPD
  • multivariate phenotype
  • type I, II error

• simulation 3

  • COPD genotype and phenotype (including Smedha's)
  • p-value

• simulation 4

  • part of Uganda
  • p-value

• real data

  • part of Uganda with missing phenotypes
  • p-value