For Huawei, AI Development is a Job That Doesn’t End with the Final Product

B. Shimmin

Summary Bullets:

• At its annual Connect conference in Shanghai, Huawei formally rolled out a new AI portfolio aptly marketed as a full-stack, all-scenario proposition that spans the physical and virtual AI solution set, covering everything from chip to solution.

• Somewhat lost in the sheer size of this portfolio, however, is a hidden gem that seeks to solve one of the biggest challenges facing AI practitioners, namely how to manage the lifecycle of AI apps themselves.

When Huawei sets its sights on a new market opportunity, rarely does the vendor tentatively dip its toes into unknown waters. As has become customary for the globally ambitious technology Chinese provider, new challenges are met all at once with an all-encompassing, all-inclusive portfolio of products that emphasizes immediate availability over future roadmap potentiality. And so it was this week when Huawei introduced its hyphen-heavy full-stack, all-scenario AI portfolio, that begins with the company’s new round of AI Ascend chips and ends with pre-integrated industry solutions.

An attempt to boil the entire stack down into something reasonably suited to general discussion, there are four basic layers to consider here, running from bottom to top:

Infrastructure: Devices, edge equipment, servers, converged systems, and cloud services featuring various AI-specific chips such as Huawei Ascend.

Infrastructure Orchestration: Huawei’s Compute Architecture for Neural Networks (CANN) offering, which is where developers can assign specific tasks to specific chips (Ascend, GPU, CPU, etc.).

Model Execution: A unified training and inference framework, branded as Huawei MindSpore, which supports the usual suspects (TensorFlow, PyTorch, PaddlePaddle, etc.).

Model Construction: This is where the rubber meets the road for developers with a set of open APIs, pre-integrated solutions, a marketplace for those solutions, and a one-stop model development workshop, branded Huawei ModelArts.

Somewhat lost in the sheer breadth of these layers is a hidden gem that seeks to solve one of the biggest challenges facing AI practitioners, namely how to manage the lifecycle of AI apps themselves, namely Huawei ModelArts. Intended to help developers rapidly build out and deploy AI solutions across any of Huawei’s underlying architecture (phones, edge servers, cloud service, etc.), ModelArts pulls together the following functionality.

Scenario Definition: Identifying data sources, evaluation metrics, etc.

Data Acquisition: Basically data preparation featuring data quality assessment and data labeling.

Model Construction: A tool to visualize, instrument, and fine tune models (as well as algorithm matching for those models).

Model Management and Evaluation: Metrics selection, results analysis and model compression capabilities.

Model Deployment: An algorithm and model repository and batch processing facility.

Model Optimization: A facility for multi-algorithm integration, model compression, performance monitoring and feedback registration.

Hidden away in that last item there’s another hidden gem, an idea Huawei refers to as ExeML, which is basically a forthcoming tool that will ply AI routines themselves to automatically optimize production-side models. Think a traffic light management system that automatically and continuously re-evaluates and adjusts its underlying predictive model parameters in response to changes in data in real-time. That seems very futuristic, of course. But as is customary for Huawei, such ideas are already in the field. For example, Huawei customer Shenzhen Traffic Police Bureau is setting up a trial AI deployment that will do real-time traffic light management. This is a task that currently demands one human per 30 intersections.

Considered as a whole, then, what Huawei is envisioning with ModelArts is a full-pipeline, lifecycle-complete model production line that doesn’t end with a successful deployment but instead emphasizes continues adaptation and improvement of ever evolving systems such as traffic flow. Will ModelArts and the rest of Huawei’s new AI portfolio help the vendor compete directly and globally with existing hyper-scale AI platform providers Microsoft, Google, Amazon, and IBM? Certainly geographically, Huawei will still face the same challenge in reaching the North American market, but in terms of providing customers with a solution-complete entry point to the AI opportunity, the vendor has made a strong statement with this release — one that emphasizes the importance of continuous improvement.


What do you think?

Please log in using one of these methods to post your comment: Logo

You are commenting using your account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

This site uses Akismet to reduce spam. Learn how your comment data is processed.