9. Working Group Meeting

[Simsso]

9. Working Group Meeting (26. August 2018)

(aka. "Half-time")

Assignments

Still, it remains mission critical to find a good classifier architecture and parametrization #23.

• @Doktorgibson ops #9 (implied #11)
• @florianpfisterer adapt and train Inception net #29 (sub task of #23)
• @simsso enhance ResNet architecture #31 (sub task of #23)

[FlorianPfisterer]

Minutes

• @doktorgibson:
  • deployment pipeline 'almost' done, only needs some testing (issue #9 done)
• @FlorianPfisterer:
  • tried Inception v3 architecture, first an adaption of the original and then a pre-trained architecture
  • with full dataset augmentation, validation accuracy of up to 50% was achieved
  • the results of a Stanford student paper (70% validation accuracy after 5 mins of training) were not achieved, even though the architecture was tried to be followed exactly (needs to be debugged!)
• @Simsso:
  • refactored crowdai ResNet code
  • tried different parameters for this architecture for maximum validation accuracy, which was around 55%
  • the crowdai-supplied weights yield 58.92% validation accuracy, more than our own training ever achieved

Next Steps

• @Simsso / @FlorianPfisterer:
  • @Simsso read the Stanford paper, discuss what was wrong with our approach so far with @FlorianPfisterer
  • try larger batch sizes with ResNet
  • try pre-trained ResNet architecture transfer learning
  • investigate last year's results accuracies
• @doktorgibson:
  • fixes to current pipeline
  • create Go application that displays relevant information (#11)
  • documentation of how the pipeline works and concretely how to use it

Goals

• @Simsso and @FlorianPfisterer: >60% validation accuracy for own and >70% for pre-trained architecture