Previously:

Artikel 1 Was ist künstliche Intelligenz? | Banian AG

Artikel 2 Was ist künstliche Intelligenz? | Banian AG

After a general introduction to AI and a short insight into the training of a neural network, the focus is now on the application. If a neural network has been produced by Data Scientists, the software "Alteryx" enables other business users to use such prefabricated neural networks themselves in an elegant way without having to delve into Python or R code.  An exemplary application illustrates the limitations of AI. A hypothetical use case serves here as an example for the AI application.

"A supermarket has decided to save customers from memorizing fruit and vegetable identification numbers and wants to develop a supermarket scale with integrated image recognition. This would involve holding the fruit in front of a camera after weighing it, which would then recognize the fruit and determine the correct price."

To learn the limits of the neural network, it is particularly interesting to present the network with different limiting cases.  For this purpose, the following new, and unseen data points have been selected.

Another Alteryx pipeline provides the business user to test and apply the model.

Legend:

New test images that have never been 'seen' by the network before have been saved in their own folder. The Directory Input Tool reads all items from the respective folder. The important information here is the "FullPath", which precisely describes the file location of each image. To add a solution (= label) to each with the small amount of test data, a text-input tool is used to simply enter the corresponding label. In the self-createdFruit Prediction Tool, the neural network is loaded and applied. The predictions are output behind it. Afterwards, the predictions are combined with the true labels using the Join tool.

With the browse tool the output of the neural network can be viewed and are shown here as a table:

Using these test examples, a few observations can be made that apply in principle to all neural networks.

If the network has seen enough training examples on a category of data, it can also make reliable predictions.

It becomes more difficult when background factors are added that were not seen during training. Even a gray shadow is enough to confuse the neural network, although a very similar apple image without background was correctly predicted. A good method to combat this problem is to use different backgrounds and lighting situations in the training data. This way the model can learn that backgrounds are not crucial for the prediction result.

So, the more diverse training data the better!

If the network has only seen apples and oranges in training so far, it will only know apples and oranges. All new categories (such as grapes or bottles) will be sorted into the known training categories. It is important to understand that the 'new' sorting does not follow human logic. Thus, one might assume that a green grape has a color that sometimes occurs in apples but never in oranges and therefore is recognized as an apple. However, the neural network analyzes the images by very different methods and can therefore make very surprising and unpredictable decisions with new data categories.

If the network has been trained to predict only one category, it can predict only one category at a time. Combined solutions and many other special features can be learned in principle, but they must occur exactly in this form in the training. Since only one category was trained as a solution here, only one can be predicted. Neural networks cannot reason logically like we humans do, and only predict what they have learned before.

So what can a neural network do?

Exactly what it learned in training. The quality of the prediction rises and falls with the training data that is provided. If you provide a lot of and different (background etc.) training data per data category, the neural network can even become impressively good. But beware: If there is data that was not learned in the training, neural networks can make very unpredictable errors. And even errors that we as humans would not necessarily consider logical, since the neural network does not analyze data the way we humans do. So all in all, neural networks are very impressive tools when properly trained and applied! Of course, the supermarket scale is just one example of the many fields where AI in the form of neural networks can be applied. A business benefit can be achieved by analyzing or predicting with neural networks with a wide variety of use cases. In the next article, some of these use cases will be presented.

Outlook:

In Article 4 follows: So in what areas can the strengths of AI be leveraged in SMEs? A typical implementation process flow is presented.