How does deep learning work?

Deep Learning, also known as “Deep Structured Learning” or “Hierarchical Learning”, is a subset of machine learning. Although the concept was first popularised in the 1980s, it has recently grown in popularity thanks to powerful software and hardware. Deep learning models copy neural networks of the human brain in structure and function. Immediately after setup, neural networks are comparable to an infant’s brain unconditioned by world knowledge. Triggered by the input of training data, the system learns basic concepts from which it in turn derives more complex concepts.

In deep learning, files go through a multi-level training model until the machine has gained experience. The many layers between input and output create a complex, deep structure, hence ‘deep learning’. The final result is an individual model that allows the recognition of patterns and the prediction of new phenomena. The deep learning process of a machine is limited to a precisely defined problem. This also means that the data used to train a neural network must be directly relevant to the problem to be solved.

Machine learning is used where patterns are to be derived from large amounts of data. It is most effective where unstructured data is available, e.g. for videos, images, audio files or text.

Examples of deep learning

An image-recognition application can learn what a cat looks like. The machine learning system receives a large number of images for this purpose. Each individual image is analysed and evaluated in just fractions of a second.

The first presentation layer can abstract the individual pixels of the cat image and then encode outlines or edges. In the second layer of the neural network, different arrangements of edges are assembled and coded. The third layer encodes elements such as the cat’s nose and eyes before the fourth layer recognizes that the image shows a face. As a result, the model assigns the image to the Cat or Non-Cat category.

The special thing about Deep Learning are the self-optimizing processes. The systems learn with increasing duration which features are most effectively transformed from the input data at which level. This has the advantage that people do not have to intervene. On the other hand, it is not always clear how the system learns. Machine learning systems such as deep learning models are therefore also regarded as black boxes.

Methods such as the object recognition described above are used, for example, to tag photos on social networks. Facebook’s “Deep Face” is a good example of a deep learning application in this area.

Deep learning applications

The strength of deep learning is in its ability to deal with large amounts of data, which it can classify or group, before making predictions. Common deep learning applications include image and speech recognition, predictive analysis, recommendation systems or anomaly detection.

Deep Learning classification is able to flag up suspicious or inappropriate content. Phrases often associated with Deep Learning include ‘This image shows inappropriate content’, ‘This financial transaction is fraudulent, and ‘This email is spam’.

The grouping of data allows results like “The user is most likely looking for this document” or “These two sounds are similar”. The latter example is a prerequisite for identifying sound clips in larger audio files and then transcribing them.

Deep learning would also provide a relevant prediction such as “Based on the data on user behavior to date, we predict that User XY will unsubscribe from our SaaS service”.

References & PDF:

https://www.deeplearningbook.org/front_matter.pdf

http://www.iro.umontreal.ca/~bengioy/talks/lisbon-mlss-19juillet2015.pdf