Ethics of Deep Learning¶

Types of errors¶

• Accuracy doesn't tell the whole story
• Type 1: False positive
  • Unnecessary surgery
  • Slam on the brakes for no reason

• Type 2: False negative

  • Untreated conditions
  • You crash into the car in front of you

• Think about the ramifications of different types of errors from your model, tune it accordingly.

Hidden biases¶

• Just because your model isn't human doesn't mean it's inherently fair
• Example: train a model on what sort of job applicants get hired, use it to screen resumes.
  • Past biases toward gender / age / race will be reflected in your model, because it was reflected in the data your trained the model with.

Is it really better than a human?¶

• Don't oversell the capabilities of an algorithm in your excitement
• Example: medical diagnostics that are almost, but not quite, as good as a human doctor
• Another example: self-driving cars that can kill people

Unintended applications of your research¶

• Think of how can this be twisted and be used in a different way.