- Imagenet error rates reduced over period of time due to deep learning approach
- Cultural things not captured
- Dataset is also culture dependent, dress, styles are dependant on geographic and cultural aspects
- Stereotype may result based on limited exposure to dataset
- Training and test distribution does not reflect all the combinations that exist in real world
- An overwhelming proportion of dataset only from US. Limited dataset from India / China / Rest of world
- Global diversity representation is needed to reflect / address real world challenges
- More Reads - https://developers.google.com/machine-learning/fairness-overview/
Keep Learning, Keep Growing!!!
No comments:
Post a Comment