What is Deep Learning?
Deep learning is a branch of machine learning that focuses on artificial neural networks and large amounts of data. Deep learning is used to make decisions and forecasts more accurate. Deep learning has gained relevance through big data and strong computing power and is developing into a new hype topic in computer science.
Neural networks have been around since the early 1940s and are not really a new topic, but big data and growing computing power and the use of graphics cards (GPUs) are rapidly gaining attention.
Differentiation of Deep Learning, Machine Learning and Artificial Intelligence. Machine learning and deep learning are part of artificial intelligence
Deep learning is part of artificial intelligence
The creation of deep learning models is very computationally intensive and the training can last for months until good predictions and decisions can be made.
Due to the complex architectures, it is often difficult to find the right model parameters, especially Google is working on automation processes (Auto-ML) to address these drawbacks.
Deep learning and artificial neural networks
Deep learning is based on the use of artificial neural networks. Artificial neural networks are algorithms modelled on the biological model of the human brain. These are used to recognize patterns or help us to form clusters and classify objects.
Of course, a deep learning algorithm will work just like everyone else algorithm of machine learning, trained by data. Artificial neural networks are often very complex, which makes the interpretation of individual decisions difficult to understand.
A very simple artificial neural network consists of an input layer, a hidden layer and an output layer.
The hidden layer neurons contain so-called weights and assign an output result to the various input signals.
In the process, input signals are viewed and transformed via activation functions.
Artificial neural networks can have simple or complex structures. They consist of Input Layer, Hidden Layer and Output Layer.
Deep learning and artificial neural networks can have simple or complex architectures.
As you can see in the diagram, a neural network can be built arbitrarily complex. For example, simple ANNs can have only one hidden layer, with more complex approaches then having 100 hidden layers.
When talking about deep learning, we mean models that definitely have more than one hidden layer.
Above all, there are many different approaches that always require different architectures.
Why Deep Learning? Big Data as a driver
Why should deep learning algorithms and artificial neural networks be used? There are problems, such as unstructured image and text recognition, which can be reproduced particularly well with neural networks.
Learning these more complex patterns is difficult with classic machine learning algorithms.
Here artificial neural networks play their strength. But also in the field of structured data, deep learning can achieve extremely good predictive qualities. Artificial intelligence brain robotic system.
Basically, the requirement of large amounts of data and corresponding technology, which allows the analysis.
The larger the amount of data, the better the deep learning works.
A good example is voice, text and image recognition. There are certain neural network architectures that are particularly well suited for this. For image recognition very often Convolutional Neural Networks (CNN) are used.
But also structured data can be processed well with neural networks, especially when many data are available.
Big data technology and the ever-growing amount of data play a crucial role in the success of these approaches. Without big data, neural networks often perform worse than traditional machine learning, as shown clearly in the graph by Andrew Ng.
Why deep learning works better with large amounts of data than traditional algorithms.
Why Deep Learning? Deep Learning Vs. standard procedure.
In practice, I find that many of my customers just do not have enough data or are just not far enough in their machine learning maturity. A disadvantage of neural networks is that it adds more complexity to the machine learning process.
Application examples of deep learning
Deep learning solves many problems, especially when analyzing unstructured data such as video, audio and text. But above all, large amounts of data are an advantage, because the models need a lot of data to make good predictions.
Well-known examples of deep learning:
Natural Language Processing for digital assistants Alexa or Siri
the recommendation and recruitment system of Spotify
the autonomous vehicles of Tesla and other vehicle manufacturers
the translation portal of DeepL
the object recognition and image analysis of Google
Of course, there are also application examples in CRM or e-commerce. Many recommendation systems use neural networks to cover complex customer behaviour and to include more features.