In this article, we introduce an innovative declarative approach to describe AI solutions.
Data processing always made a big step forward with respective flexibility and speed, when software development was augmented by declarative approaches: Declarative business rules and SQL statements are prominent examples of this development.
SAISL, short for Structured AI Solution Language, marks the next big step for modern data processing. It introduces four semantic layers — cases, tasks, workflows and plugins — to enable a holistic declarative description of AI solutions:
From initiating business problems, drivers and desired outcomes to multiple associated business tasks down to AI-driven data workflows and its building blocks.
Tasks can be specified along the proven phases of analytic processes:
Understand, Prepare, Build, Evaluate and Operate.
This is inspired by the CRoss- Industry- Standard- Process for Data Mining, and the Analytics Solutions Unified Methods for Data M ining & Predictive Analytics.
SAISL supports two different semantic perspectives to categorize workflows:
- Analytics: This view distinguishes descriptive, diagnostic, declarative, predictive and prescriptive workflows — according to Gartner’s analytic continuum.
- Purpose: Workflows can be specified as discover, ingest, learn and predict workflows.
Plugins represent the basic technical vocabulary and describe the reusable building blocks of AI-driven workflows. A plugin is a (declarative) reference to a pre-built configurable software artifact and either specifies a software connector to a data source or sink, or an operator to transform data.
SAISL categorizes plugins as source or sink, or, as action, alert, analytics, condition, lookup and transform. Analytics plugins support
- Deep & Machine Learning,
- Business Rules & SQL Queries,
- Natural Language & Time Series Processing.
The entire architecture of complex AI solutions can be described in a single structured text document (in SAISL format) that is human understandable and machine-readable by a soft-ware code generator to transform AI workflows into executable software.
SAISL was inspired by the Common Workflow Language (CWL) that is used in data-intensive domains such as Bioinformatics, Medical Imaging, or High Energy Physics.
It goes way beyond CWL approach but starts from the same observation and ground truth: Any more or less complex data processing solution can be described as a structured set of related data workflows.
CWL and SAISL are both initiatives to make AI solutions reusable as a whole or in parts and, due to its declarative approach, provide additional business value:
AI solutions can be made accessible to search & recommendation technologies as any other structured or unstructured text document. The dream of a corporate AI knowledge management where every AI solution is human understandable, and reusable is near to become true.
SAISL is not restricted to facilitate cross-domain AI knowledge management: In most real-world environments, the underlying data changes constantly. Models need to be adjusted, or even entire AI solutions need to be rebuilt to tackle feature drift and ever-changing business and regulatory requirements.
SAISL makes two important contributions: (a) AI models become reproducible and (b) AI solution release cycles are accelerated.
Originally published at https://www.linkedin.com.