Data

• Register on kaggle.com
• Select and enter one of the completed Kaggle competitions based on topics/code learned in-class. Examples:
  • Kannada MNIST - handwritten digit recognition written using the Kannada script
  • Kuzushiji Recognition - transcribe ancient Kuzushiji into contemporary Japanese characters
  • Aerial Cactus Identification - determine whether an image contains a columnar cactus
  • Movie Review Sentiment Analysis - classify the sentiment of sentences from the Rotten Tomatoes dataset
  • Dog Breed Identification - determine the breed of a dog in an image
• Explore Kaggle datasets for other interesting projects
• Consult with the instructor if you would like to use your own data

Proposal Submission (due 06/26/2020)

• Register on GitHub. Learn about GitHub in the section 4 - Git and GitHub of the Reproducible research tools
• Create a repository for your project
  • If teamwork, add your teammate as a collaborator
• Add README file describing the selected project. Address the following points:
  • Data type (numerical, text, or images), link to the data source
  • Problem type (binary or multiclass classification, regression)
  • Proposed network architecture (feed-forward, CNN, RNN, etc.), testing multiple architectures encouraged

Solution

• Create your code in RMarkdown
• Describe and implement data download and processing
• Select, justify, and implement a network architecture suitable for the data
• Train and evaluate the performance of the network
• Tweak network architecture, add regularization, if needed, to improve model performance
• Explain all steps, describe the achieved performance

Many Kaggle competitions allow for late submissions in the form of text files with, e.g., samples and predicted classes. Such submissions will be benchmarked against previous submissions, and you will be able to compare how your model performs with respect to previous solutions.