Security solutions that autonomously detect and ward off attacks from cyberspace. Chatbots that answer customer queries in no time at all. Manufacturing robots that work hand in hand with their human colleagues to perform complex tasks: Artificial intelligence (AI) today has many faces. But smart products and software solutions have one thing in common: they perform their tasks without human intervention – and they also solve any problems that arise independently.
Machine Learning needs concrete instructions for action
Machine learning (ML) methods provide the basis for this. They enable machines to extract, classify and analyze texts and images from electronic documents. Sounds simple, but it is not. This is the result of a new study by the Massachusetts Institute of Technology (MIT). According to this, AI can recognize stereotypical images very well. However, if the motif deviates only slightly from the norm, the errors become more frequent. This is because intelligent software has to learn from scratch how to classify such entropies correctly. The human brain, on the other hand, does this by itself. Accordingly, training plays an immensely important role in the development of AI applications in addition to the database.
The first step in the development of AI solutions, however, is to define the task as concretely as possible and to create initial solution options. For example, if a certain piece of information is to be filtered out of a document, the artificial intelligence has to know where it can be found in the document and which components it contains. An example: In a Swiss vehicle registration document, the chassis number is usually always found in the same place and always contains 17 digits. This is what the AI needs to know in order to identify the desired information (i.e. the chassis number). Accordingly, it is necessary to program the data model and algorithms. With the help of rapid prototyping, it is possible to create corresponding prototypes quickly and easily.
Afterwards, the selection of the training tool is on the agenda, as there are a number of different machine learning methods. For processing image data, Convolutional Neural Networks (CNN) are the tool of choice. This deep learning architecture uses different filters (Convolutional Layer) and aggregation layers (Pooling Layer). This enables CNNs to use the acquired image data to generate a model that can predict the chassis number.
AI: Iterative training leads to success
Once the ML model and method are in place, the training begins: The solution should learn to perform its task – for example, the extraction of chassis numbers from a document -– as error-free as possible. For the AI this means: Practice, practice, practice. And for the developer: continuous fine-tuning. Since the continuous iterations not only help to improve the AI’s skills, they also show optimization possibilities of the algorithm and thus pave the way to sharpen the AI’s learning ability.
While traditional software development uses iterations mainly to test new functionalities, ML applications are about teaching AI to recognize different entropies. Ultimately, the intelligent solution should also be able to identify and filter out the fuzziness that it does not yet know. To this end, the developers continuously check the delivered results during the training phase and continually adapt the individual neurons. In this way, the self-learning algorithm can be improved with each iteration.
The goal of the training loops is thus to reduce the error rate as much as possible. Only when artificial intelligence can compete with human intelligence in the particular task will its use pay off. Who wants a chatbot that keeps answering customer questions with “Sorry, I don’t know”? To prevent this from happening, the AI keeps learning – even when it is already in use.
And also the programmers continue to stay in charge. Because continuous monitoring not only enables them to close potential knowledge gaps of the AI in productive use and use corrections for bias drift due to changes in input signals and (for online models) changes in the “transformation”, but also to identify new requirements. Lifelong learning is therefore also and especially indispensable for smart products and processes.