image_pdfimage_print

Retreat notes and progress untill 05.09.2022

Agenda

Next steps after AAAI submission

Upcoming research questions to answer

  1. Normalize total loss
  2. What is the performance of CR-VAE with ResNet architecture on MNIST and CIFAR-10 Datasets?
  3. What is the performance of MoCo on MNIST and CIFAR-10 Datasets?
  4. How does CR-VAE-BIG compare with MoCo?
  5. What is better, SGD or Adam? Why?
  6. What is better, E2E or Modular? Why?
  7. How can we train on ImageNet? Maybe alternative datasets?
  8. New architecture: decoder input -> concatenated latent representations from q and k encoders.
  9. Can we incorporate all representation techniques into one?

post paper submission setbacks

  • KL divergence computation was wrong. When fixed, performance was different
  • With a weight factor of 1 for the KL divergence, the learned features performance in classification task diminish.
  • This report shows this problem.
  • Same behavior for CR-VAE. Untill the reconstruction and the contrastive losses are in the same scale with the KLD loss, the performance will continue to deviate. This happens because KLD dominates numerically the total loss.
  • Way to mitigate it:
    • Descending beta value
      • currently exploring different scheduling techniques
      • report
  • Note: CR-VAE does not seem novel now
  • Normalizing total loss (weight loss inversively with their magnitude) might lead to better performance

Meeting Notes -Melanie-Prof Rueckert 22.07.2022

Agenda

  1. Data / Visualization framework in Python
    • API framework/usage guidelines
    • input  format
    • output format
    • Gui to inspect the data.
      • Original Data view
      • subsection view (using the rotated and subselected image parts).
      • Slider to adjust the time.
      • Play function.
      • Replay speed adjustment.
      • Basic statistics of the shown data (e.g. histogramms of the two images, min, max, mean, boxplots, number of blobs [1], …).
  2. Symmetry measurements
    • Develop measures and visualization tools to detect asymmetries between the two images.
    • Find examples of such asymmetries.
    • Analyze them.
  3. Occlusion removal
    • Classical CV approach
    • Learning-based approach(De-Occlusion)
  4. Abnormality detection
  5. Thesis writing

Topic 1: Data / Visualization framework in Python

Deliverables due to September

use pandas dataframe library

Topic 2: Symmetry measurements 

Deliverables due to October:

  • Develop automatic detection methods and selection tools for your guidance
    • e.g., highlight these events in your time-line in the gui with red bars, or create a list of events that can be selected

Topic 3: Occlusion removal

Concerned about its appicability to this project. we cannot assume it will work with non-face data

If we don’t know the dynamics of the liquid we cannot reconstruct maintaining the true underlying information.

due to November

https://arxiv.org/pdf/1612.08534.pdf

https://github.com/zhaofang0627/face-deocc-lstm

Next Steps

  • Abnormality detection
    • due December
  • Thesis writing
    • due January

Next Meeting: TBA

Meeting Notes – Melanie 19.07.2022

Agenda

  • Presentation on Metallurgy review
  • Next Steps
    • Study dense NN
      • MNIST dataset
    • Study CNNs
      • Classification
      • Bounding Box
      • Segmentation
      • Feature matching
    • Autoencoders
      • Anomaly detection
    • FlowNet 2.0

Topic 1: Presentation on Metallurgy review

Great introductory presentation on each book’s(3) content

No need to study the math on the properties of mixtures on the second book

Next Steps

  1. Present an introduction to NN/CNNs
  2. Small jupyter tutorial on DNN/CNNs
  3. Presentation of FlowNet paper

Next Meeting: Tue 26 July

Meeting Notes 15.07.2022

Agenda

  • Present Paper Concept: Contrastive VAE
    • https://docs.google.com/presentation/d/1zBnog1A9mlHpZ4sFhwS12w6UBMrNYRfKgxCbLvcysUA/edit?usp=sharing

Topic 1: ConVAE

Present concept, math and next steps

Notes

  • Add an intermediate step in the introduction: “Why we want the latent representations”
  • Motivate Unsupervised Learning
  • Consider using a different distribution for the prior: L1 norm for example
  • What has changed in the behavior of the ConVAE in comparison with VAE from an Information Theory perspective
  • Mathematical proof that it can work in all datasets
  • Train later on ImageNet

Meeting Notes – Melanie 14.07.2022

Agenda

  • Dataset updates
  • Progress
    • Metallurgy review
  • Next Steps
    • Study dense NN
    • Study CNNs
      • Classification
      • Bounding Box
      • Segmentation
      • Feature matching
    • Autoencoders
      • Anomaly detection
    • FlowNet 2.0
  • Tools
    • Pycharm
    • Pytotch
  • meeting schedule
    • day/time
    • place

Topic 1: Dataset updates

On Friday’s meeting with Voest

Topic 2: Progress

Focus on the part of the data comes from.

make a presentation. 6-7 slides each book

Topic 3: Next steps

  • Study dense NN
  • Study CNNs
    • Classification
  • FlowNet 2.0

 

Topic 3: Tools

  • Pycharm
  • Pytotch

 

Topic 4: Meeting Schedule

Next Steps

  1. Presentation of Metallourgy SOTA
  2. Present an introduction to NN/CNNs
  3. Small jupyter tutorial on DNN/CNNs
  4. Presentation of FlowNet paper

Next Meeting: Tue 19 July