Deep learning can do wonders for complex data, with thousands to millions of features and a large history of labeled examples to use as training sets. The rapid advancements in image recognition has as much to do with the vast trove of identified images that Google and others have collected over the years, as to the advances in the deep learning algorithms used.

For IoT data, this can limit the usefulness of deep learning techniques. Most IoT data is relatively new without a long history of labeled examples. Many IoT devices have only a few sensors, so the feature set is not as complex. In these situations, many of the ML techniques discussed previously can do as good, if not better, job of prediction than deep learning techniques. Deep learning is also more computationally expensive in both time and compute power (that is, higher cost).

However, if the IoT data flow includes a large number of features and hundreds of thousands to millions of labeled training data is available, deep learning techniques may be able to provide a significant boost in predictive power. This is clearly the case in autonomous vehicle development. It could also do wonders where images are captured as part of the device functionality (static or video).

Deep learning packages tend to be interacted with best by Python as opposed to R. They are also relatively new, so expect documentation and tutorials to be limited. This makes them a little more finicky to work with than the prior examples in this chapter. You will need more time and expertise to develop deep learning models than for more well-trodden ML models, such as random forest.

Use what method makes the most sense for the individual use case and available training data. Consider the effectiveness requirements and see whether using a deep learning modeling technique is likely to provide enough boost in accuracy to warrant the expense.