No account yet ?

How to detect and solve overfitting in Machine Learning ?

Intuitively, if your model performs too well with the training data, and it fails to predict correctly when it is in production, then there is an overfitting problem.

Bias and variance:

Bias is an error corresponding to the general solution found with a model that learned to detect rectangular shapes when you wanted to detect vehicles in images.

The variance is an error in sensitivity to the training data, the result of which will vary greatly depending on the data.

The model is not stable. When the model suffers from a lot of variance, then there is overfitting.

The solution goes through a good organization of the initial data by carrying out a cross-validation.

For normal training, you divide your data into training data and test data, while for cross-validation, you will perform several trainings by doing iterations where you will vary the distribution of training and testing.

Thus, at the end of the iterations, all your data will have been used for both learning and testing.

Since the model has been controlled through these iterations, you are less likely to end up with an overfitting model.

How to detect and solve underfitting in Machine Learning?

A machine learning algorithm is considered to be underfitting when it cannot capture the underlying trend of the data.

Underfitting destroys the accuracy of the machine learning model.

Its occurrence means that our pattern does not match the data.

This usually happens when we have less data to build an accurate model and also when we try to build a linear model with less nonlinear data.

In such cases, the rules of the machine learning model are too simple and flexible to be applied to such minimal data and therefore the model is likely to make a lot of wrong predictions.

Underfitting can be avoided by using more data and also reducing features by feature selection.

Underfitting = High bias and low variance