Artificial intelligence (AI) is one of the biggest technology talking points of our time. Since the arrival of the computer, AI has been inspiring fiction and film – and as the levels of automation have been increasing, so has the fear of it replacing manual labor in the future been stoked. But what will happen next? What can thinking machines really do? How, what and why do they learn? Here we have the answers.
Picture: Microsoft (Cortana)
Mechanical or artificial intelligence?
Artificial intelligence generally represents the endeavor to build machines so that they can behave like humans. So that they solve problems not just "on command", but actually learn independently and draw logical conclusions. So that they "think". The key terms here are Machine Learning and Deep Learning.
What is Machine Learning?
Machine Learning is a component of artificial intelligence. The objective is for the machine to analyze data volumes based on automatically created models (algorithms) – i.e. to learn from sample data and recognize rules within it – and for it to be applied on unknown entities after a short learning phase.
During tasks such as the classification of data, the machine receives input data and output data and is to learn a rule which is as universal as possible. If the input data is film descriptions, for example, then an output could be the corresponding genre. This is known as supervised learning.
Unsupervised learning, on the other hand, does not produce output data. Instead, the output is in the form of recognizing patterns in the data independently and grouping them into clusters, for example. This process can be applied in areas such as predictive analytics (predicting events, damage/loss and behavior) or recommender systems (shopping cart analyses and purchase recommendations).
The level of intelligence in these examples is limited. Machine Learning is primarily about exact, automated pattern recognition based on empirical information and training data.
What is Deep Learning?
Deep Learning denotes both a current paradigm and specific technological processes of Machine Learning – and a real leap forward in the development of artificial intelligence. It differentiates itself from traditional Machine Learning by the following factors:
Deep Neural Networks
The architectures for Deep Learning algorithms are generally what are known as multilayered neural networks. The more layers there are, the deeper the neural network. (That is why there is also the term deep neural networks.) Each layer learns a representation of the inputted data. The output of the layer moves as an input into the next layer, in which a new representation of the data is learned. The further a layer is from the input data, the more abstract the information that is learned in it. That is how, for example, image-based recognition of vehicles is learned in the layers near to the data, while the layers far from the data recognize objects such as wheels and windshield.
In each layer a nonlinear transformation of data occurs. Nonlinear functions are more complex than the linear functions that are often applied in traditional processes. By connecting the layers, these complicated functions are effectively series-connected and therefore enable even more complex transformation of the data. It has been mathematically proven that Deep Learning networks like these are able to handle such complex mathematical functions; they do not even have to be particularly "deep". This is a sign of their universality and, therefore, also their huge potential.
Automated feature learning
With traditional Machine Learning approaches, the raw data must be processed in order to extract useful features. That is why for image-based recognition of vehicles, for example, typical features such as wheels and windshields first had to be manually specified and combined.
Extracting and combining the features is often the costliest stage when creating a Machine Learning algorithm. However, Deep Learning algorithms enable this process to be done automatically, without "manual" intervention. This contributes to neural networks being able to independently learn which combinations of features are useful and which are not.
Where is Deep Learning put into action?
Example 1: Facial recognition
A familiar application of Deep Learning is the recognition of things in images (visual object recognition). The multilayered model of neural networks is a crucial factor for successful face recognition, since, similarly to the brain, it enables certain features to be identified separately from each other. (Another crucial factor is the computing power available today. Deep Learning requires many samples and large data volumes to learn from.)
Picture: Knowledgebase Business Informatics
Example 2: Object recognition
Another example of Deep Learning in action is the automatic, quick and convenient recognition of spare parts on an industrial scale, as described recently in a LEADdigital publication featuring *um.
Example 3: Speech recognition
The insights and methodology of Deep Learning have long been in the virtual wizards on our smartphones and in digital communication. Today, all technology companies use this type of software – from the speech recognition modules for Apple’s Siri or Microsoft’s Cortana to text analysis and automated translations on Skype.
How will this intelligence develop?
Despite the many recent successes and the huge leaps forward in development, artificial intelligence is still very much a technology of the future. Its possibilities are currently being utilized only to a limited extent. For the foreseeable future, Deep Learning is artificial intelligence’s most important component. However, whether it can meet all demands is questionable. For more complex and strategic tasks in a more automated future there will be more developed solutions available in the future. That said, the initial demands on artificial intelligence remain the same – to build "thinking machines" that simplify and improve our lives. We are working on it.
In the meantime, here is something productive & applicable for you:
*um on Machine Learning/Deep Learning as a Service
*um on Advanced Machine Learning/applicable Deep Learning
*um Data Scientists are creating key foundations for Machine Learning and Industrial Internet