Historically speaking, machine learning is the older and simpler technology. It works with an algorithm that adapts when it receives human feedback. One requirement for making use of this technology is the availability of structured data. First, the system is fed structured and categorised data, and in this way, it understands how to classify new data of the same type. Depending on the classification, the system then carries out programmed activities. For example, it can distinguish whether a photo features a dog or a cat, and allots the files to their respective folders.
An initial application phase is followed by the optimisation of the algorithm using human feedback – for this, the system is informed about any incorrect classifications and the correct categorisations.
With deep learning, structured data isn’t necessary. The system works with multi-layer neural networks that combine different algorithms that are modelled on the human brain. That’s why the system can also process unstructured data.
The approach is most suitable for complex tasks where not all aspects of objects can be categorised beforehand.
Important: In deep learning, the system finds suitable differentiation characteristics in the files by itself, with no need for any external categorisation. In other words: training by the developer isn’t necessary. The system itself considers whether to change classifications or produce new categories based on new input.
While machine learning can already work with a manageable data pool, deep learning requires much more data. For the system to produce reliable results, more than 100 million data points should be available.
The technology for deep learning is also more costly to implement. It takes more IT resources and is significantly more expensive than machine learning, meaning that – for now, at least – it isn't an option for mainstream businesses.