3 Essential Steps to Achieving Optimal Deep Learning Results

BrandPostBy Roland Kunz Ph.D.
Oct 25, 2019
Analytics Big Data Hadoop

When executing a deep learning environment, a thoroughly deliberated architecture is crucial for success.

shutterstock 272663177 retouched 1280x1280
Credit: Dell EMC

Adoption of deep learning has gained a lot of traction during the last two or three years across a wide variety of use cases and has become a top area of interest for many enterprises around the world. Yet, these enterprises still need to determine how best to spend their investment to yield meaningful results for their business. Once an organization has identified the most advantageous use case with which to begin, it is important for CIOs to consider a thoroughly deliberated architectural design.

Here is why…

On the one hand, the toolchain for deep learning environments is diverse. There are a wide variety of development toolkits, frameworks and libraries from which to choose. Even the choice of hardware to run the deep learning workload can have a significant impact on an organization’s results.

On the other hand, data scientists and departments begin their evaluations based on their first-hand knowledge and ideas. Unfortunately, this approach may not always provide an ideal fit for later scalability and production-readiness. From first evaluations running on a data scientist’s laptop, or from a set of experiments using virtual machines in any given public cloud, the initial ― perhaps randomly set ― starting points carry over to later production.

Bearing this in mind, parallel planning for the underlying architecture, including careful consideration of vital architectural ingredients, is essential to ensuring optimal deep leaning results. Three key steps should be incorporated as early as possible into every design process.

  1. Pin down the architectural ingredients

It is crucial to consider the following four architectural ingredients as early as possible in the process:

  • Maturity of tools, libraries and frameworks used
  • Scalability from one data scientist to mass production
  • Data locality, amount and availability of meaningful data
  • Ease of use for data scientists to deploy their workloads

These are critical success factors. As such, it is important to bring data scientists and IT architects together at the very earliest possible stage to form a team that will be capable of describing an optimal solution for the organization.

  1. Answer key data questions

Initial discussions should be held to answer key questions about the data, including

  • Where will the data be collected and cleaned?
  • What amount of data will be required?
  • Where will the data be placed in relation to compute resources?

Further investigations should occur on a mature software stack, where the deep learning can be best performed. Be sure to consider the skills and experience of your own workforce.

  1. Determine compute resources

Finally, a choice of compute resources (on premise / off premise) needs to be determined and should provide enough network bandwidth to bring the data to be trained in and out of the system.

One starting point for such a scalable environment might be a predefined solution, such as those described at Dell EMC Ready Solutions for AI, where the architectures allow data lake-like storage of all required training and inferencing data, as well as enough compute resources to perform various deep learning experiments at the same time. A valid software and toolchain stack is integrated, which relieves the data scientist from ever-changing software dependencies. Finally, the necessary automation to scale and expand the system is integrated ― even for hundreds of data scientists working in parallel. Just like a good menu, it is created based on standardized recipes leading to a good state.


To summarize, along with defining the deep learning use case and starting research, it is equally important for CIOs to plan in parallel for the underlying architecture and to consider crucial architectural ingredients as early as possible in the process. This approach, including the three key steps outlined above, will help to ensure meaningful deep leaning results for your business.

To learn more

For more perspectives on tapping the value of data with deep learning and artificial intelligence systems, explore Dell Technologies AI Solutions and Dell EMC Ready Solutions for AI.