Limitations of Machine Learning
- a model cannot be created without input data
- a model can only learn based on patterns found in input data
- input data must be labelled
- models predictions will reflect existing biases in the data
- a predictive policing model based on past arrest data doesn't predict crime, it predicts arrests
- existing biases in arrests will be perpetuated
- feedback loops
- when model results influence future model training, existing biases will be amplified
- when predictive policing model is deployed nad used, additional arrests may be focused on areas that the model already over-emphasised
- feeding this back into future training exacerbates the problem