₪ State categorization of high-dimensional amorphous parameter spaces using novel neural network architectures and learning methods

Classic as well as deep learning methods quickly reach their limits, especially with high-dimensional and structurally heterogeneous data sets. The reason for this is, among other things, that the most extensive learning data sets of the neural network structure have to be presented for these tasks in order to achieve a suitable representation structure for successful categorization/identification to adequate classifiers using the "trick" with many neural layers and a high learning rate. But if you look into the structures of the central nervous systems of vertebrates, it quickly becomes clear, that Mother Nature did not go down this path. So for example the pyramidal cells - as essential classifiers of the central nervous system - are by no means laid out in high-level structures. Taking this fact into account and following the principle of "Computing with Activities", the employees of IngB RT&S have created new situation evaluation systems in recent years - especially in the context of European R&D projects - and put them on the Internet.

Two examples, here from the area of ecological situation assessment and personal health categorization, are listed here.

• General assessment of the situation, here ecological, economic, industrial, tourism and shipping-related assessment of old ammunition stocks in the Baltic Sea (developed as part of the EU joint project DAIMON).    

The task was to build a CI-based system using data sets from the above-mentioned areas, which categorizes the influence of old ammunition stocks in terms of whether and to what extent the condition affects the different "protected goods", which (known) measures are recommended for evaluating and clearing the situation and which legal restrictions in the respective national and international areas have to be taken into account. The procedure of the situation support system (availiblr in online operation) built by our current employees is shown on the right as an example.

A user selects an object based on a general object map.

Next the object data (mostly chemical, physical, biological, site-specific and exploration-related) are taken from a database, the single-layer neural network creates a categorization and selects the appropriate legal and procedural situation support information. All relevant information relating to the selection of protected goods in focus, together with a general categorization into the areas "green", "yellow", "orange" and "red" of the effects of old ammunition, are immediately made available in the form of a report.

• Position assessment and position support of diabetes patients using a 360 degree analysis using continuous data collection and a neuronal-based personalized "fingerprint"

There are now many online-based diabetes assessments based on physiological parameters, but fewer online-based diabetes assessments that take into account the social, seasonal, professional and psychological environment.

In addition, most commercial systems are not able to monitor the trend in behavior, feelings and relevant changes in the social and professional environment of the diabetic using a personalized and individual neuronal "companion" and to predictively create a prognosis about their state of health (here). . However, this is easily possible using our neural networks, since their special structure firstly allows time trajectories to be integrated into their categorization structure and secondly, their categorization structure can be continuously expanded - without losing their old knowledge base. Of course, the categorizers are encrypted in such a way that third parties cannot read their knowledge base without a special key. 

Publication

..and how do our single-layer networks work? Well, a little like rotating salad bowls!