The Vikings ruled the seas from 793–1066 AD with their now-famous long and durable ships. Their exceptional sailing and navigational skills made them a dominant force that changed the course of European history. But it was their technology and ship building mastery that proved to be the true differentiator. At the core of this knowledge was their deep understanding of the nature of the trees and wood.
In fact, it was a common practice for Vikings to personally build the ships they sailed in, creating a special continuity and integrated understanding that assured a ship’s performance. These ships were built to handle the roughest seas, with structural designs that would bend and flex with the wind and waves, but never break.
By carefully analyzing the grain of the logs, boards were split from logs with the grain running the entire length, which was a time-intensive process but delivered superior results. These masters knew exactly where to split them and what part of the ship each best served. For example, straight-grained logs planked the hull, while curved-grained logs were used for the ribs.
As a math geek and an artificial intelligence (AI) and machine learning (ML) veteran, this ship building analogy really intrigues me. Though each tree and log had different qualities and each group of sailors built their own vessels, there was a shared technology and contextual knowledge that allowed the Viking’s to consistently build the strongest ships in the world. Each part intelligently served to make the whole stronger. Ultimately, it wasn’t about building boats. It was about building a healthy, sustainable, and dominant community.
In contrast, what disappoints me today is to see so many companies building AI models with no end game in mind. This is especially true in the nuanced and challenging world of natural language processing (NLP). Too many teams are relegated to building “stovepipe models” purposefully built for specific use cases, which cannot be leveraged by other teams across the enterprise for a variety of reasons, including:
· No access to the corpus to change the model functionality
· Different algorithms used to create the model
· No common runtime engine to execute the model
· No standards to share or deploy the models
· Inability to rectify competing models.
In most environments, every new NLP project starts from scratch and is designed to operate on its own. Too often I hear engineering groups say, “Oh that’s easy. We’ll get some training data and build some models”. They are chopping wood, without any plan of how to integrate it into a sea-worthy vessel.
NLP is all about language, which is an infinitely complex challenge. To meet this challenge, organizations need to take a page out of the Vikings shipbuilding handbook, and create a common model — building resources where learning is additive, dynamic, and incremental. The models need to be aware of each other and work together to establish an aggregated intelligence, and the entire system needs to flex and bend with turbulent business environments.
It’s not about building models. It’s about incrementally building intelligence that incorporates all the work from all of the projects across the organization. AI/ML is the new programming language. Just as object-oriented programming and reusable components became an accelerator to software development — companies must have the same vision for NLP. Models created by one group should be shared, added-to, and deployed by another group.
And as new models are created, the infrastructure should become more powerful and intelligent. At the risk of going against the grain, maybe we should call it NLP: Norse-like Language Processing.