The Cloud Experience Everywhere
cancel
Showing results for 
Search instead for 
Did you mean: 

Operationalizing machine learning: 5 challenges, 1 solution

ML can unlock valuable insights from data, but many companies struggle to implement effective, consistent workflows. A new HPE GreenLake offering breaks through the barriers to ML success.

MLOps_Blog_Graphic.png

Machine learning (ML) is enabling enterprises to make data-driven business decisions by using sophisticated models to deliver insights from large datasets. But in order to be able to fully realize the value of ML — including new revenue streams and improved customer experiences—enterprises need to implement fully operational ML workflows.

According to a recent study conducted by Forrester Consulting, 41% of companies say they have struggled to operationalize any ML models and lack the process to do so.1

Organizations in every industry are looking at ways to leverage ML to harness the power of their data and deliver business innovation through data science. But even when they achieve some measure of success with ML pilot programs, many organizations face challenges when they seek to scale these programs to production, such as security concerns, legacy hardware, siloed data and workflows, inefficient processes, and daunting costs.

Facing challenges along the way

Enterprises must capitalize on ML’s value – or risk getting left behind. But this landscape is complex, and enterprises must quickly address the barriers to speed. No business wants their data science projects to be held up for months waiting for the hardware, software, and tools needed to run ML models. Across the ML lifecycle, you may be experiencing some of these common challenges:

  1. Data Preparation. Data has changed – the nature of the data, the volume of the data, the velocity, frequency, format – everything is different now. The scale is much larger than it was before. And that influences some of the build environments and the requirements for those environments. This is producing a lot of open source technologies along with proprietary technologies from vendors to accommodate or to facilitate the changes that are happening in the actual data itself. This creates an ever changing, expanding, open source ecosystem and customer-driven or vendor-driven ecosystems.
  2. Build. Data scientists tend to work in silos and oftentimes are building their models on local machines. This creates variability in development environments, making it hard to share code and collaborate. Localized development is the result of data scientists having limited access to infrastructure with on-demand provisioning of a containerized sandbox environment providing necessary ML and Deep Learning tools, interfaces and frameworks, as well as secure access to shared data.
  3. Train. Having access to a training environment is key. You need one that is dynamic enough to be easily provisioned and also can scale – to the size needed to support the vast amount of historical data that the model is trained against, and to the location or geography where your training is going to happen. Many companies don’t have the right infrastructure in place to support this.
  4. Model Deployment. Once the model training is complete, it is handed off to the development team for deployment. The infrastructure needs to be able to leverage the runtime environment and applications that were used to train the model. And it must be capable of being built up to meet the requirements of production, with rules in place to address load increases with autoscaling.
  5. Monitor. Monitoring and retraining are ad-hoc and opportunistic at best, increasing the risk of an inaccurate prediction. Often, the operations team is monitoring the environment for scalability, network saturation, and overall performance while at the same time, the data analytics team is monitoring the models to determine their efficacy and need for retraining.

None of these areas in the enterprise can work in isolation. How are you making sure that the team that’s developing is looking at the data from the same sources? And the team that is providing the scalable infrastructure is making sure that that’s available to the end users? Etc., etc., etc.

In our experience working with customers, we have seen them encounter these challenges every step of the way. But a lot of those challenges are around deploying the model at an enterprise scale. By operationalizing ML pipelines, businesses can rapidly turn data into business intelligence, generating new opportunities and value for the organization. The faster you can deploy and scale infrastructure to operationalize ML models, the faster you recognize your ROI.

HPE GreenLake: Delivering end-to-end value

To deliver the value of ML and data science to your enterprise, HPE GreenLake offers an enterprise-grade ML Ops cloud service that makes it easier and faster to get started with ML/AI projects and seamlessly scale them to production deployments. Designed to address all aspects of the ML lifecycle, from data preparation to model building, training, deployment, monitoring, and collaboration, you can quickly deploy ML/AI workloads within your datacenter or colocation facility on HPE's ML-optimized cloud service infrastructure featuring HPE Apollo hardware and powered by HPE Ezmeral ML Ops software.

HPE GreenLake offers consumption-based pricing, allowing you to consume these resources on premises with a cloud experience. It provides data scientists with self-service access to a sandbox environment for prototyping and testing, enabling them to eliminate IT provisioning delays, ensure repeatability, and accelerate time-to-value. And as a fully managed solution, the HPE GreenLake offering frees IT from routine infrastructure management tasks.

Organizations need an end-to-end solution that brings DevOps-like speed and agility throughout the ML lifecycle. HPE GreenLake now makes it possible to realize faster time to value with operationalized ML across your business. Reducing the typical six-to-nine month infrastructure procurement cycle down to mere weeks, the integrated solution is delivered to your on-premises location and monitored and managed to reduce risks, provide a simplified data science experience and accelerate insights from your data.

Want to learn more?

Come see us at NVIDIA GTC virtual event October 5-9, 2020.

Featured articles:

1. "Operationalize Machine Learning", a commissioned study conducted by Forrester Consulting on behalf of HPE and Intel®, June 2020

0 Kudos
About the Author

katedavis

I have been working in the tech industry for over 14 years marketing hot topics including storage, software-defined, big data, hybrid cloud and as-a-service.