In which cases would SVM be better than deep learning?

Classical machine learning models, including SVM, are usually better options compared to deep learning models when the amount of data is very less.

This is becuase normally deep learning models have a lot of weights (free variables) that need to be tuned with data. If the number of weights are more than (or around the same as) the number of training examples, the deep models just ‘memorize’ the data, leading to overfitting.

Moreover, if computational power is a limitation or predictability of the outcome of the learning model is a concern it is advisable to use classical machine learning first before opting for deep architectures.