Machine Learning

Machine Learning (ML) involves training a mathematical model with training data and then using it to automatically evaluate data from the usage context. To ensure that the model only learns the relevant aspects of the input data, the data is usually subjected to feature extraction beforehand. For image data, the extraction of the most frequent color or the identification of geometric shapes would be such a feature extraction.

There are various methods, which differ in terms of the learning process: supervised learning, unsupervised learning, semi-supervised learning, and reinforcement learning. The models used include decision trees, regression analysis, cluster analysis, or artificial neural networks.

Machine Learning methods

Supervised Learning: Predictions and forecasts according to a target pattern (from existing “labeled” data) for new data, e.g., traffic sign recognition.

Unsupervised Learning: Recognizing patterns and correlations in data independently, e.g., for clusters of shared interests of buyers of wine and cheese.

Semi-Supervised Learning: Mixture of supervised and unsupervised learning, e.g., to make do with a small amount of existing labeled data, which usually has to be classified by humans at high cost.

Reinforcement Learning: A reward system guides learning through additional data. Example: the distance traveled to the destination to find the shortest path through a maze for a robot.

Machine Learning vs. Deep Learning

Deep Learning is a method that uses very large neural networks to learn feature extraction during unsupervised learning, for example, and then manages without specially labeled data. However, the neural network used is significantly more complex, as more stages are required and significantly more training data must be used to train such a network.

 View comparison table

Unlike in the case of software programming, Machine Learning models are “taught” with the help of software tools. This training phase is followed by a validation phase before the model can be integrated into a software system and is ready for use.

Even though the training or learning phase requires a lot of effort, the resulting models are significantly simpler and therefore also able to analyze data quickly. This means that, up to a certain level of complexity, ML applications can also be used in embedded devices or in apps on mobile phones.

Machine Learning requires at least 50-100 data sets per feature to be considered if the features in the data are known or can be extracted. Methods such as Deep Learning, which manage without these features, require several thousand data sets per feature. This means that very large amounts of data are accumulated very quickly.

If large amounts of data of the same type are to be analyzed with regard to patterns (e.g., in the case of image or frequency analyses) or if correlations in data need to be identified, solutions can be obtained much more quickly and easily by using Machine Learning methods rather than algorithmic methods.

With the help pf Machine Learning methods, astonishing results can often be achieved in a wide variety of application areas.

Advantages of Machine Learning methods are:

• Automated solution without “programming”
• Complex, even unknown data correlations can be resolved under the right conditions

Disadvantages of Machine Learning methods are:

• Need for extensive, possibly labeled training data
• Difficult to impossible traceability/comprehensibility
• Effort-intensive validation with extensive test data
• Remaining validation gap of non-predictable “false” statements or predictions in most methods

View use case examples

The decision criteria for selecting the appropriate learning method include:

• Number of existing data
• Number of classified/labeled existing data
• Characteristics of the data (homogeneity, quality (completeness, correctness, accuracy), number of data types, etc.)
• Aim of the evaluation
• Available computing power (embedded device or cloud environment)

There are many tools, libraries, and even freely available open-source software options for AI solutions. Initial results can usually be achieved faster than initially expected.

We are happy to support you in the development of your specific solution and make our AI Innovation Lab available to you as an infrastructure for evaluation. 

View service portfolio