Technical info about ElinarAI™
ElinarAI™ consists of two major components. There are ElinarAI™ Miner (Docker Container) and actual AI instances that are run on Tensorflow Serving containers. The ElinarAI™ Miner instance is commonly hosted by a customer data center. AI Instances run on the Elinar data center on IBM Power8/9 HPC appliances. We also fully support a customer-hosted option at an additional cost. We carry out AI model training at Elinar’s data center using IBM Power8/9 HPC appliances.
With ElinarAI™, you will receive a full REST API for business processes and external systems to query AI and provide training data. It also provides a highly optimized user experience for manual training data entry with an intuitive approach for both tagging and advanced document classification.
ElinarNER (Elinar Named Entity Recognizer) is one major part of ElinarAI™ Miner.
ElinarAI™ relies on text analytics to do what we call “Glorified Named Entity Recognition” or GNER. This is the first part of the ElinarAI™ processing pipeline. In practice, it prepares and anonymizes data for both AI training and inferencing. Elinar has developed a high-performance ElinarNER module that preprocesses text with high efficiency and speed so that Deep Learning models can create more accurate understanding. Elinar has currently GNER support for several business areas including Financial Documents, Healthcare, and Privacy Data.
Some use cases require more flexible GNER and for this, we use IBM Watson® NLU. Watson NLU can be very useful in cases where there is a need for a large number of variations on the underlying text analytics pipeline. For example, if working within a large corporate environment, where there is a need to map slightly different business domain data into a single AI model, Watson NLU can be an excellent choice for GNER functionality.
Watson Knowledge Studio allows multiple corporate developers or data scientists to develop underlying text analytics models concurrently where each developer has specific domain knowledge and they use this to map interesting entities into a single ontology that feeds AI for common corporate analytics. By using Watson NLU, customers can easily make a single AI model to serve multiple business processes or data sources within the analytical domain. ElinarNER is a high performance tool but it does not always shine on this type of agile variations that are needed within a complex corporate environment.
Model training takes place 100% anonymously as data is pseudonymized and significant portions of content are heavily changed. We are training models to recognize meaningful patterns in unstructured data. This is important because we can then remove all identifying information and business secrets from the training data. This ensures that even the strictest data security and privacy protection requirements can be covered with ease. In practice, it means that when we train the model in sales orders (SO), we are unable to deduce the following from training data: For example, which product was ordered, the quantity of the product, or who placed the order.
Deep learning models
The TensorFlow models of ElinarAI™ vary based on the complexity of the business problem and the volume of training data. They range from simple classification to sophisticated deep models for data extraction. And in this case, simple classification means a fairly narrow construct with a few complex layers and 3-5 recurrent layers.
We have also developed a large “cookbook” of deep learning models for various text and content-centric problems. Naturally, Elinar specialists have extensive experience with these models. They fine-tune the models with the client architects for optimum performance and accuracy based on business requirements. For example, AI that creates a security classification uses rather a different topology from AI that needs to extract 40 different topics (like the reason for a complaint) from a document. These classification levels are secret, confidential, company confidential and public.
Training data management and workflows
All customer data is managed by ElinarAI™. You can, of course, manage training data on a project or topic-specific basis. We’ve secured training data so that individual trainers will only see data assigned to them. This ensures that training data is created securely. Administrative users can assign training data to individuals in bulk. They can also access training data created by each individual to ensure the quality and accuracy of data. The use of ElinarAI™ allows a business to utilize crowdsourcing and other innovative ways to mitigate the cost of AI implementation.
- A comprehensive solution for understanding unstructured data.
- Utilizes state-of-the-art text analytics (Watson NLU or ElinarNER) together with deep learning.
- Includes IBM Watson and AI solutions developed by Elinar.
- Runs on top of the IBM Watson Machine Learning Accelerator (formerly known as IBM PowerAI) and utilizes Tensorflow, Tensorflow Serving, and TensorRT.