IBM AI OpenScale

2018

Release of the new version eliminating bias in artificial intelligence

On December 14, 2018 the IBM corporation submitted the new version of the platform of AI OpenScale capable to explain to users as AI models come to the conclusions and to reveal bias the decisions made by artificial intelligence.

According to the statement of corporation, AI OpenScale there is an important instrument of modern work with AI because a considerable part of the enterprises just does not trust artificial intelligence, when it comes to adoption of the most important solutions. The main problem is that users are not sure of impartiality of models and lack of system mistakes. Besides, many are confused by a problem of the hidden logic of AI – they would like to understand on the basis of what data the model came to this conclusion.

IBM issued the platform eliminating bias in artificial intelligence

The vice president and the director of laboratory of the research center IBM Research Almaden in San Jose explained that it is not always easy to recognize bias of AI. The platform of AI OpenScale is intended for elimination of a problem of system mistakes using algorithms which are capable to detect and correct them in all range of the AI applications, without interrupting their work.

The platform of AI OpenScale constantly traces the AI applications and reveals system mistakes which tend to collect in models of machine learning. The platform registers all forecasts, versions of model and all teach-pattern data, and then draws conclusions in language clear to the user.

AI OpenScale also helps to simplify deployment of AI models and allows the applications created using any model of machine learning open source or model of deep learning, to work in the normal environment. Own platforms IBM Watson and PowerAI, Wednesday with open source code Seldon, SageMaker from Amazon Web Services (AWS), Microsoft Azure and others concern them. Besides, the platform is supplied with the NeuNetS beta function - the mechanism of synthesis of neural networks which, according to IBM, will allow the enterprises to create quickly neural networks from scratch.^[1]

Announcement

In October, 2018 IBM provided the platform of AI OpenScale which, according to the company, allows to create artificial intelligence practically on any infrastructure.

AI OpenScale offers tools which allow to reveal among other things the systematic errors connected with identification of a certain outcome in researches.

The system of recording which writes the forecasts made by means of models of machine learning and their versions is provided. It is possible to use a data set for training and the related performance measures.

IBM issued the platform for creation of AI almost on any infrastructure

The platform is created for consolidation of several frameworks which are used for solving of tasks  using artificial intelligence. She constantly monitors perhaps negative decisions which are made by AI model, undertakes measures for mitigation and offers explanations to the recommendations issued by AI algorithms.

AI OpenScale supports the AI models developed on different open platforms, including Google TensorFlow Microsoft AzureML, SparkML, Keras, Seldon and Amazon SageMaker. Such compatibility facilitates implementation of these models in environments, it seems IBM Watson both Seldon, and other third-party platforms, notes the VentureBeat edition.

AI OpenScale will be available to clients of IBM Cloud and Cloud Private through online control panels.

 In addition to AI OpenScale the NeuNetS engine (Neural Network Synthesis Engine) which develops new personal AI models is created. This engine helps specialists with automation of process of creation  of AI, providing creation of complex and  deep neural networks  from scratch. Such technology is useful to the programmers who do not have experience with artificial intelligence and also for the accelerated implementation of such technologies.

IBM claims that during testing of the personalized AI models which are independently developed, trained and unrolled through NeuNetS the accuracy of results, "comparable to that which is issued by the neuronets created by people" was reached.^[2]

Notes