Embedded AI: Challenges & Solutions for Business Managers

AI-computing is different from conventional computing! This is now a widely accepted knowledge, though the details are not understood by most. (I intend to publish lessons in difference between AI-computing and conventional-computing as soon as I get time..) Additionally, computing for Embedded-AI is very different from "Server-based-AI". Unfortunately, the differences are not understood by most engineers. Business decision makers may often hear conflicting views from engg team, which makes business decisions challenging.

The good news- enormous enthusiasm. Everyone, absolutely everyone, wants to adopt AI in their products

Expertise

Lack of deep AI understanding is the biggest challenge. But it can be solved with more open and frank dialogues with AI-component vendors. By definition, they have prepared methods & tools to build applications. Your engg team may need to adapt. Some adjustments in attitude, workflow, legal and business procedures may be needed from existing push-button, quick-and-easy mechanics. But, overall, a solvable problem with just a small will power!

Embedded-AI is NOT datacenter-AI

Embedded AI is harder because of significant resource, power and price constraints. Models and applications that run on servers and Cloud fail in Embedded world because of resource constraints. We have found it easier to train an embedded engineer to build AI algorithm than to train a server-AI-engineering to become embedded AI-engineer.

The Pretenders

AI is such a selling word that too many, who are not, tag themselves as AI expert. HW & SW components selling as something else in past have been tagged as AI or AI-enabled without a material change. If your HW & SW vendors cannot explain their AI specialization in excruciating technical details, then you might be talking to a pretender. Best to be careful!

Quest for data and Embedded-AI-models

here is abundance of reference AI-models, for free & sale, for server and cloud AI applications. But, Embedded doesn't have the same eco-system. This can be a problem or opportunity. Opportunity, because you can own a differentiation.

But what about TTM? Embedded-AI market is here to stay. The problem exists for you exists for others as well. As long as you can use your domain expertise to create a unique model and create your differentiation - you are fine. The worst you can do - Is to do nothing! Remember: In your area - your domain expertise is the best tool to create those models. Not a third party. At least not yet!

Benchmarks?: I am not a personal fan of benchmarks. Every time I have gone digging deep in performance benchmarks, I have found them insufficient and error prone. But my personal bias aside, benchmarks for embedded-AI are still too pre-mature. I would just focus on end-application parameters.