ag visuell - Fotolia

Implementing deep learning requires a creative approach

Using deep learning in an effective way requires creative problem-solving and a team approach that goes beyond simply hiring data scientists, experts say.

Implementing deep learning in enterprise settings requires a lot more than just downloading some open source algorithms, but with talent scarce, businesses are finding it takes creativity and an open-minded approach to achieve results.

"Established industries are largely missing out on the benefits of AI," said Ryan Kottenstette, co-founder and CEO of Silicon Valley geospatial data company Cape Analytics LLC. "If you're not in the tech sector, you might be waiting a bit longer for the benefits of AI to be realized."

In recent years, deep learning has taken huge strides. Algorithmic processes like neural networks, which historically lived more in the realm of mathematical theory, have moved into some enterprise use cases, like computer vision and process automation.

But adoption has been uneven. Digital-native companies have been quicker to implement deep learning than more traditional industries like manufacturing or retail. In a presentation at the O'Reilly Media Artificial Intelligence Conference in New York, Kottenstette said this is partly because large tech companies like Amazon, Google, IBM and Microsoft have bought up smaller AI companies and hired the leading talent.

Enterprises need to get in the deep learning game

But that doesn't mean enterprises in traditional industries should sit on the sidelines. Kottenstette said enterprises should look for niche use cases that have the potential to be impactful, but that can also scale. This strategy minimizes the need for large teams -- sidestepping the talent crunch -- and reduces the risk of high-stakes failures.

Established industries are largely missing out on the benefits of AI.
Ryan Kottenstetteco-founder and CEO, Cape Analytics

For example, Cape Analytics buys geospatial and satellite image data and applies neural networks to create curated data sets that identify specific objects. Rather than pitching the service as a general-purpose tool for anyone who might want it, the team started by pinpointing outdoor pool enclosures on single-family properties in Florida and selling the data set to insurers.

Kottenstette said the team knew that these pool enclosures are hard for insurers to identify when they write policies, leading to high unplanned for costs. The fact that there was a known customer base with a specific problem made it a strong test case for the service. Once the concept was proven, the team started looking to use the tool more broadly, selling more general data to insurers trying to identify other characteristics of properties.

Diagram of a common deep learning process
These steps are commonly part of the deep learning process.

"It's about as narrow as you can get, but it allowed us to test the thesis that these people are going to care about our product," he said.

There's been some debate about whether enterprises should start small when implementing deep learning, looking for quick wins, or go big, crafting their broader business strategy around AI. The Cape Analytics example suggests it's possible to do both. Start with small projects that have the potential to expand into more pervasive business opportunities.

It takes a village to raise AI

When implementing deep learning in enterprise settings, it's important to remember that it takes a team approach.

At the conference, Honeywell's chief scientist for artificial intelligence and machine learning, Chris Benson, said companies today tend to think that data scientists are the key to making AI and deep learning work, and many companies are trying to hire for the position. With the lack of data science skills available in hiring markets today, this mindset puts AI out of reach for most enterprises.

But, he said, data scientists aren't the only workers needed to implement deep learning models. In addition to data scientists, he hires data engineers, software developers and domain experts to create teams to implement AI. Some of these non-data scientist roles are typically easier to hire for.

Aside from overcoming the skills gap, the approach of hiring people with a diverse set of experiences also increases the chances of bringing in all the skills your team needs, Benson said. "Just because you have data scientists doesn't mean they are experts in deep learning," he said.

Dig Deeper on AI technologies

Business Analytics
Data Management