Many online courses and universities promote deep learning courses and bootcamps. This enthusiasm goes also to business decision makers and investors. And then, there comes the push to incorporate deep learning / artificial intelligence / machine learning models at all costs. For someone transitioning to data science, is deep learning a valuable skill to invest time and money to learn?
The truth is, unless you are working on images, audio or text, deep learning is not a very valuable skill. Deep learning excels at extracting patterns from high-dimensional data that is generated in a consistent way. Only in those cases makes sense to use deep learning. The typical data scientist that uses a mixture of SQL and classical data mining algorithms (logistic regression, decision trees, random forests) on tabular data is unlikely to benefit from it.
Even in cases where deep learning would normally work great, one should also be aware of the amount of data. If one does not have enough data, deep learning algorithms will not work. A rough estimate is 10 data points per parameter. This is a completely heuristic figure, which I cannot back up by theory. Modern neural networks have millions of parameters.
During my corporate training courses we often benchmark deep learning models against others. In most cases, deep learning models are way below the mark. One exception to this are outlier detection models. In this setting, we often get better results with autoencoders. But otherwise it is hard to make the case for deep learning.
Instead of investing time and money in learning advanced algorithms, newcomers to the field should brush up on their data gathering and analysis skills. That is definitely a differentiatior.