Why fastai?

This is my first blog post discussing about fastai, it’s uses and why it’s important for you as a beginner in machine learning to start using fastai. Let me tell you, you might be late to the party, but you won’t be slow with fastai ;)

• Simple top-down approach

• Build products for applications

• Challenges
  • Working with videos
  • Handling nighttime images
  • low-resolution images
  • speedy inferences for usefulness
  • out-of-domain data and domain shift in data over time
• Data Ethics
  • Questions to be asked: Given a situation (jury is yet to be out on ethical or unethical), How would I deal with it? What would I look out for?
  • Ask questions - understand the data flow and steps towards building a product/algorithm - Say NO if you find something is fishy.
  • Some data product design questions
    • What level of aggregation will you store data?
    • What loss functions, validation and training splits?
    • Focus : Simplicity OR Speedy inference OR Accuracy?
    • Challenges : Do you focus on out-of-domain data?
    • Should it be fine-tuned or train from scratch?
  • Recourse and Accountability [Arkansas healthcare]
    • blame-game - government and creator of the algorithm
    • data contains errors → audits and correction
  • Feedback Loops
    • YouTube story about pedophile recommendations → Build ethical metrics, maybe?
    • Using or rather selecting features that does not create the biases on metrics based on observations. Ex: Evan Estola’s meetup idea
  • Bias
    • Historical, Representation, Measurement, Aggregation, Evaluation, Deployment
    • Ailments
      • Diverse data helps Aggregation bias, but no help to historical or measurement bias
      • proper documentation of datasets, limitations and how decisions are made using context and biases
  • Disinformation
  • Identifying and Addressing Ethical Issues
    • Analyzing the project that you’re working on
      • Should we even do this?
      • Biases in data?
      • Code and Data – auditable?
      • Errors for different subgroups
      • Accuracy of simpler algorithm
      • Mistake and appeals handling
      • team diversity
    • Implement processes to find and address ethical risks
      • Consult over assumptions
      • knowing/asking stake-holders interests and direct effect
      • wrongful/unintended usage and purposes
    • Support good policy
      • rights approach
      • justice approach
      • utilitarian approach
      • common good approach
      • virtue approach
    • Increase diversity