Deep Learning Essentials

When not to use fine-tuning

There are a few cases where you may want to resort to other techniques rather than fine-tuning:

• When the new data/domain is drastically different from the one used for training the existing model. In such cases, training the model from scratch could be a better solution, given there are enough data volumes.
• When the new dataset is very small, for example, less than 2 thousands of samples. In such cases, continuing to train with such a small dataset may still lead to overfitting. However, one could use the existing network as a feature learner to extract a feature from the early layers and feed into traditional machine learning models, such as SVM. For example, one could take the output of the intermediate layer ...