Error Rate
The error rate refers to the proportion of prediction errors in the prediction. The calculation formula is generally: 1 – Accuracy (%)
The trained model can generally be used to measure the error rate of a model in a data set, where three numbers are important:
- Bayes Optimal Error: An ideal unmeasurable limit value that can be used to approximate the human error rate in image recognition.
- Train Error: The error rate of the model used on the Train Set;
- Dev Error: The error rate of the model used on the Dev Set.
Strategies to reduce error rates
1) Reduce Bias
- Try a larger model, such as a neural network with more layers, more neurons, etc.
- Extend training time;
- Adjust the optimization algorithm, such as trying Momentum, RMS Prop, ADOM, etc.;
- Switch to neural network models such as CNN and RNN.
2) Reduce Variance
- More data added;
- Add constraints to make the fitted function smoother;
- Switch to CNN and RNN models.
References: