Neural networks

Neural networks

Machine learning

Machine learning enables the artificial intelligence of open bi to develop new content-related connections between data through training. If a search on the web finds a >85% result, new neurons are formed independently. In future, this question will lead to a 100% hit, which can then also be found in a shorter time. So when open bi is asked new questions that are similar to questions known to the system, open bi learns.

In majaAI's cloud solution, all customers use a common neural network. The output of the answers remains customer-specific, so no one has access to the data of another customer. Thus, however, all users train a common neural network, which maximises the learning effect.

Network training

In artificial intelligence, a distinction is made between supervised learning and unsupervised learning. The Artificial Neural Networks (ANN) of open bi learns with the help of unsupervised learning: This means that the ANN learns without a control instance, but independently through application by the user alone.

The training of the ANN takes place in 2 phases in open bi:

  1. In the first phase, the exact phrase is trained, i.e. a formulated question is entered into the system. The NLP breaks this phrase down into its components and forms the first neurons and connections.
  2. In the second phase, open bi independently derives different formulations of this phrase. This happens, for example, through the use of synonyms. Or morphology (= derivation of different tenses)

Through this automatism, the system automatically forms a variety of answerable questions from the one learned question.

Searching and finding in the ANN


A successful search in the ANN should be both performant (=available in a reasonable time) and correct in terms of content. To ensure maximum performance, the first step in searching for a result in the ANN is to clean up the sentence, i.e. the question posed. For example, punctuation marks and breaks are removed, as well as irrelevant filler words (e.g. "actually"). Using a specific algorithm, the network is now searched in parallel for a suitable result. A hit occurs when a word in the network corresponds to the word searched for. This match can be complete if the word corresponds 100% to the searched word. However, partial matches can also occur if, for example, a word was misspelled. By noting these partial matches, typing errors that occurred when entering the question can be corrected, for example. If no match is found, so-called skipping mechanisms can be used to continue the search. Both by correcting typing errors and by skipping, the answer found is rated as worse. Each individual search can therefore

  • return a 100% result: if the question was found exactly on the web
  • provide a result between 80% and 100%: if the question was found partially on the web
  • provide no result: if the network was sufficiently searched and no result had more than 80% match with the question

Once a 100% result has been found, the search is terminated and the user immediately receives his result.

Creating neural networks

open bi offers all programmatic means to create, train and apply neural networks. For example, "Conversational AI" is just one use case that has been implemented using open bi's artificial intelligence. For example, you can read report or sensor data into a neural network from open bi to identify correlations and find relevant data using Artificial Intelligence.

The Conversational AI of open bi is suitable in this context as a "mouthpiece" for providing the results to the user: the user asks a specific question, thus starts the neural network and receives an answer based on the result of the network search.