Hitachi AI Application Framework for Development and Operation of AI Applications

Determining Benefits through Trial Development

The use of the agile development approach for developing enterprise systems is gradually starting to gain ground. Meanwhile, waterfall development is more commonly used. This approach uses the defined requirements to progressively subdivide functions, adding more and more detail as development progresses. However, defining the requirements ahead of time is very difficult for AI application development as the type of learning results returned by the AI engine will depend on the input data and parameters.

So when developing AI applications, the requirements are often set by first holding a trial development phase to verify feasibility. This phase consists of repeated cycles of learning and verification to determine the benefits of using AI. System development is then expected to begin once these benefits are understood and the requirements have been redefined.

The use of this different approach is what makes AI applications often take longer to develop than standard applications.

System Integrity Requirements

Since enterprise-domain mission-critical systems have a major impact on companies and consumers, they demand a higher standard of integrity than is demanded for information systems or systems in other domains. Ensuring system reliability and stability is difficult on a system-by-system basis. The Hitachi Application Framework has improved system-building efficiency by providing Java control functions and common components used to ensure non-functional requirements such as reliability and stability.

Python^*2 is often used for AI application development as it has a number of different mechanisms for making the work easier (such as more efficient use of AI engines and data). However, Python does not yet have sufficient functionality for implementing the non-functional requirements.

Accommodating AI-specific Characteristics

AI, and especially machine learning are made to learn by using an AI engine on a specific data set provided as input. The learned model obtained as the result is then applied to unknown data to make predictions. Programming an AI application usually means creating the application used to generate these predictions from the learned model.

Learned models can sometimes return impermissible results when the data entered in them has different characteristics from the data that was used during learning (such as a different distribution or inclusion of invalid values). Instead of just being a one-time occurrence, this different data can sometimes be the result of a change in preconditions (such as a change of seasons). So when this data is entered, learning may need to be re-done with a new data set.

Therefore, when integrating AI into a system, these AI characteristics need to be accommodated by having the system operation incorporate mechanisms for continuously monitoring and managing learned models and input/output data beforehand. This requirement is a major difference from conventional systems.

*2

Python is a registered trademark of the Python Software Foundation.

Providing AI Application Templates

The approach to increasing application development efficiency used by the Product’s forerunner (Hitachi Application Framework) was to automatically generate source code and test code from design sheets.

However, Hitachi’s developers realized they would need to use a different approach from this automatic code generation since AI application development often includes a trial development phase before the start of design to generate results for use in the production development phase.

So the approach they opted for instead was to provide general-purpose features in the form of templates covering all the elements except for those specific to each system. These templates were derived from previous successful proof-of-concept and production AI applications (see Figure 2).

Figure 2 — Mechanisms that Improve Application Development EfficiencyThe Product’s forerunner (Hitachi Application Framework) was designed to implement the original requirements by automatically generating source code templates from design sheets. The Product alters this approach by instead providing templates created from past examples and enabling them to be customized to meet original requirements.

By using these templates starting from the trial development phase and adding customizations to meet the requirements, the developers thought they would help enable faster AI application development.

To help boost AI application development efficiency, the Product’s initial release provided the following three templates:

1. Regression model examples (call center workload prediction)
2. Classification model examples (employee data-based behavior prediction)
3. Classification model examples using AI engines expanded by Hitachi (repeat customers prediction for marketing purposes)

Providing Common Components and Control Functions for Enterprises

Although there are some differences between Java and Python as development languages, they are not considered to have any major differences in terms of the common components and control functions needed for developing applications to be embedded in enterprise systems.

Therefore, the existing common components and control functions of Hitachi Application Framework were rebuilt for Python to efficiently fulfill the requirements for a high degree of integrity (The common components are components for areas such as log output and message management. The control functions are used to control elements such as application execution sequences and transactions). Since the data to be used for input resides in the existing system, an interface to connect the existing system with the AI system has also been provided to enable easier use of previously gathered data (see Table 1).

Table 1 — Common Components and Control Functions for EnterprisesThe Java common components/control functions that were provided by Hitachi Application Framework have been used to extract the essential items needed for AI application development. They are provided as rebuilt component sets for Python.

Providing Mechanisms for Managing and Monitoring Learning and Prediction

As mentioned, input data that trends differently from the learned data can sometimes yield impermissible prediction results since AI applications use learned results (learned models) to run daily predictions on unknown data sets.

This characteristic led Hitachi’s developers to believe that operating a system with embedded AI applications would require the input/output data processed during learning and inference to be constantly monitored during execution, along with ongoing version management for learned models. They decided to provide these features as common mechanisms.

While monitoring requirements vary by system, it is possible to standardize the items to monitor and information to check to some extent, considering the characteristics of AI. The developers therefore decided to provide a mechanism that enables automatic error detection where users register sets of items to monitor and information to check as monitoring rules in the Product beforehand, and the Product automatically checks the input/output data processed during learning and inference (see Figure 3).

Figure 3 — AI Behavior MonitoringThe use of preset rules to automatically monitor input data and output results allows the user to detect data changes and unexpected output.

This mechanism was intended to enable system operation using suitable learned models without requiring any model awareness from the user. This operation is possible because the mechanism enables the Product to automate the version management of learned models and the selection of learned models used during inference.