What are online machine learning services

Artificial Intelligence

Machine learning is becoming more and more mainstream today. Until a few years ago, self-learning programs were exclusively a topic for universities, research institutions and some technology companies, but today they are increasingly being used in normal products and solutions. Our everyday and business life is increasingly determined by intelligent programs that learn from data and generalize what has been learned.

Speech recognition on cell phones such as the iPhone or Google cell phones, for example, is largely controlled by machine learning algorithms - just like spam filters in PCs and notebooks or facial recognition when managing photos. We are often in contact with learning systems without knowing it - for example with personalized online advertising. And more and more companies are recognizing the value of machine learning when it comes to optimizing their business and saving costs.

For most IT users, machine learning is a confusing field because there are many different conceptual, methodological and theoretical approaches. This FAQ clarifies the most important questions and introduces the central terms, methods and applications.

  1. Facebook faces
    Computers can learn to distinguish human faces. Facebook uses this for automatic face recognition.
  2. Machine learning
    Contrary to what the picture suggests, machine learning is a sub-area of ​​Artificial Intelligence - albeit a very important one.
  3. AlphaGo
    Machine beats people: In 2016, Google's machine learning system AlphaGo defeated the world champion in the game Go.
  4. Graphics processors GPU Nvidia
    The leading companies in machine learning use graphics processors (GPUs) - for example from Nvidia - for the parallel processing of data.
  5. Deep learning
    Deep learning processes first learn low-level elements such as brightness values, then medium-level elements and finally high-level elements such as whole faces.
  6. IBM Watson
    IBM Watson integrates several artificial intelligence methods: In addition to machine learning, there are algorithms for natural language processing and information retrieval, knowledge representation and automatic inference.

What is machine learning?

To put it bluntly, machine learning is the art of making a computer do useful things without specifically programming it to do so. To put it more precisely, machine learning is the acquisition of new knowledge through an artificial system. Like a human being, the computer independently generates knowledge from experience and can independently find solutions to new and unknown problems.

For this purpose, a computer program analyzes examples and tries to recognize certain patterns and regularities in the data with the help of self-learning algorithms. The aim of machine learning is to intelligently link data with one another, recognize relationships, draw conclusions and make predictions.

How does machine learning work in principle?

In principle, similar to human learning. In the same way as, for example, a child learns that certain objects can be seen in pictures, a computer can also "learn" to identify objects or to distinguish between people. To do this, the learning software is first fed with data and trained. For example, the programmers tell the system that one particular object is "a dog" and another is "not a dog".

The learning software continuously receives feedback from the programmer, which the algorithm uses to adapt and optimize the model: With each new data set, the model gets better and can ultimately clearly distinguish dogs from non-dogs.

What are the advantages of machine learning?

Machine learning helps people work more efficiently and creatively. For example, they can use machine learning to organize and manipulate their images faster. With machine learning, you can also leave boring or time-consuming work to the computer. Learning software can independently scan, save and file paper documents such as invoices.

Above all, self-learning machines are able to take on very complex tasks for humans - such as the detection of error patterns or possible damage in production (predictive maintenance). Even with the detection of cancerous tumors in medicine and with therapy recommendations, self-learning programs now help - and often outperform the best human experts. This ability to process complex relationships between the input and output of large amounts of data is one of the main advantages of machine learning.

Is machine learning the same as artificial intelligence?

No. Machine learning is a branch of Artificial Intelligence (AI). In the same sense, logic, analysis and stochastics are sub-areas of mathematics; Mechanics, thermodynamics and quantum physics are sub-areas of physics. Artificial Intelligence itself is a sub-discipline of computer science and generally deals with the automation of human intelligent behavior.

Artificial intelligence - the German term - includes machine learning, sub-areas such as knowledge-based (expert) systems, pattern recognition, robotics, the processing of natural language and machine translation. However, machine learning is currently one of the central and most successful artificial intelligence disciplines.

Why is machine learning soaring right now?

Machine learning is based on research in pattern recognition that was carried out as early as the 1980s. The area then stagnated for a long time due to technical restrictions. Only a few years ago, machine learning experienced a breakthrough with the ability to process data in parallel in graphics processors (GPUs) - which were actually developed for the video game industry. Graphics processors have thousands of computing units and are significantly faster compared to solutions with classic CPUs.

Further developments such as multi-core architectures, improved algorithms and super-fast in-memory databases such as SAP HANA make machine learning particularly attractive for the corporate sector. Another important factor is the increasing availability of large amounts of structured and unstructured data from a variety of sources, including sensors or digitized documents and images, with which the learning algorithms can be "trained".

Which methods are used in machine learning?

Machine learning uses mathematical and statistical models to learn from databases. In detail, there are dozens of different procedures. In principle, a distinction is made between two systems in machine learning: First, symbolic approaches such as propositional systems in which the knowledge - both the examples and the induced rules - is explicitly represented. Second, sub-symbolic systems such as Artificial Neural Networks, which function on the model of the human brain and in which knowledge is implicitly represented.

The algorithmic implementation of machine learning takes place with monitored or unsupervised learning. In supervised learning, the system learns from given pairs of inputs and outputs. A "teacher" provides the appropriate or correct value for an input while learning. The aim of supervised learning is that after several calculations with different inputs and outputs, the network is trained to be able to establish connections. In unsupervised learning, an algorithm creates a model that describes the inputs and enables predictions. The network then independently creates classifiers according to which it divides the input patterns.

How does an artificial neural network work?

Artificial neural networks simulate a network of interconnected neurons based on the model of the brain. They learn from experience by changing the connection strength of the simulated neuron connections. In this way, machines can acquire skills such as seeing, hearing, speaking, reading and writing. Supervised learning methods are used to train them for these skills.

The learning process works roughly as follows: First of all, the network learns in the training phase using the given material. The "trainer" gives the network a number of examples and repeats the whole process. For each example, it is known what the desired output should be. If the output of the network matches the desired output for an example, nothing further needs to be done. If the actual and desired output deviate from one another, the connection strengths or weights in the network must be changed in such a way that the error in the output is reduced.

The greater the amount of weight, the greater the influence of one neuron on another neuron. A positive weight exerts an exciting, reinforcing influence on another neuron, a negative weight an inhibitory one. The weight of the connection thus largely determines whether one neuron is influenced by another. It is one of the decisive factors for learning processes, one could say: The knowledge of a neural network is stored in the weights. Ideally, this training process continues until all examples are correctly calculated. The whole learning process is an iterative process in which a special algorithm sets the weights so that the output corresponds as closely as possible to the known result.

What is deep learning?

Deep learning is currently the most successful implementation of an artificial neural network. At the same time, deep learning is now the most widespread machine learning method and is used by large IT companies such as Google, Apple or Facebook. The speech recognition of the iPhone "Siri", for example, is based on deep learning. In addition to language processing, one of the most important areas of application for deep learning is the recognition of objects in images.

The process makes many work steps of classic neural networks superfluous because the computer takes over all intermediate steps. The researcher only has to present data such as images to the neural network; the network will then find out how to identify these on its own.

Deep learning uses the analogous mechanism that a toddler learns the term "dog" for example: First, training data is made available to the computer program, for example a series of images, each of which a human has with the meta tags "dog" or "not dog" "has marked. The program uses the information it receives from the training data to generate a feature set for dogs and build a predictive model.

The units of the first level only registered the brightness values ​​of the pixels. The next level would see that some of the pixels are connected in lines, whereupon the next one distinguishes between horizontal and vertical lines. This continues until finally a level is reached in which legs can be distinguished.

In another model, perhaps the computer would predict that everything in an image that has four legs is a dog until it is finally able to distinguish dogs from non-dogs. With each iteration, the predictive model the computer creates becomes more complex and accurate.

Can machine learning keep up with human learning?

Yes. This is proven time and again by human-machine competitions that require the highest level of cognitive skills. For example, IBM's cognitive learning-based system Watson clearly beat the human candidates in a TV knowledge quiz back in 2011. In the last sensational human-machine competition, Google's AlphaGo machine learning system defeated the reigning Go world champion in a game over five rounds with a clear 4: 1 at the beginning of 2016. AlphaGo used a variant of the deep learning process.

The Asian strategy game was previously considered too complicated for computers due to its complexity, because there is an almost unlimited number of possible positions. The players therefore mostly have to rely on their intuition. The AlphaGo algorithm that has been developed is now helping Google, among other things, to save electricity. With the help of the algorithm, energy consumption in Google's data centers could be reduced by 15 percent.

What are the most popular machine learning applications?

Machine learning can be found in the recommendation services of Amazon and Netflix as well as in the face recognition of Facebook. The ability to mark individual members with their names on pictures has led to the world's largest collection of faces in a database on Facebook. Facebook can use this data to specifically train machines in visual recognition.

Machine learning processes are also behind email applications that automatically detect spam. The computer analyzes the data contained in the email and categorizes it as spam or non-spam according to the patterns it recognizes. If a message is marked as junk, the computer learns and can thus identify junk messages even better. Learning methods are also used in the defense against computer attacks, the fight against Internet crime and search engine ranking.

How can machine learning be used commercially?