Dealing with an Underfitted or Overfitted Model

Building a deep learning solution is not just a matter of defining an architecture and then training a model using the input data; on the contrary, most would agree that this is the easy part. The art of creating high-tech models entails achieving high levels of accuracy that surpass human performance. This section will introduce the topic of error analysis, which is commonly used to diagnose a trained model to discover what actions are more likely to have a positive impact on the performance of the model.

Error Analysis

Error analysis refers to the initial analysis of the error rate over the training and validation sets of data. This analysis is then used to determine the best course of action to improve the performance of the model.

In order to perform error analysis, it is necessary to determine the Bayes error (also known as the irreducible error), which is the minimum achievable error. Several decades ago, the Bayes error...