Deep learning use cases aren't limited to big tech companies

Industries that are not traditionally technology-driven are starting to find ways to use deep learning, proving the tools aren't just for large tech companies.

When digital advertising company Laughlin Constable was approached by the Wisconsin State Cranberry Growers Association about a project to use deep learning to identify pests in cranberry bogs, the idea was met with skepticism.

First of all, it's not the kind of project an advertising firm typically takes on. And it was also unclear if the growers had the technical foundation on which to build this kind of deep learning capability. But speaking at the Spark + AI Summit in San Francisco, the company's vice president of engineering, Sam Saha, said he felt it was worth trying as a deep learning use case.

"There are lots of businesses out there and they don't have the talent pool [to use AI]," he said. "It becomes a vicious cycle where the innovation isn't flowing downstream. At some point, you have to just get up and try."

Deep learning not just for big tech companies

Most of the deep learning use cases we've seen in recent years have come from large tech players like Amazon, Facebook and Google. This has created a perception that, while a powerful tool, deep learning is just for the leading companies, with limited applicability in more general enterprise settings. But some are out to prove this perception wrong.

Saha said his company approached the project as a proof of concept that would demonstrate how less tech-savvy companies can implement AI tools.

His team built an iPhone app that enables cranberry growers to take pictures of suspected pests. Initial images were labeled by employees at Ocean Spray, a major buyer of cranberries that has an interest in the health of bogs. Then, Laughlin Constable data engineers built a deep learning model using TensorFlow to identify specific varieties of pests in the farmers' pictures.

Saha said it's a good example of how AI tools like deep learning can be used by organizations outside the technological hubs of Silicon Valley and the I-95 corridor. It also shows the power of open source tools.

On the business side, people are always worried about whether we actually need AI. On the engineering side, they say we're not Google. We're not Amazon. Where do we start? Before they even attempt to solve a problem, they give up.
Sam SahaLaughlin Constable

The deep learning model used in the project was already built and available in TensorFlow; it just needed to be trained. But Saha said he doesn't see a lot of people in enterprises today taking this deep learning use case approach. Instead, he sees a belief that if you aren't writing models from scratch, you aren't doing real AI, which is self-defeating and impedes AI adoption at businesses that might benefit from it.

"On the business side, people are always worried about whether we actually need AI," Saha said. "On the engineering side, they say we're not Google. We're not Amazon. Where do we start? Before they even attempt to solve a problem, they give up."

Building AI in the construction industry

It's hard to get much farther from AI than pouring concrete and laying pipes, but at the summit, a senior data scientist at construction company Bechtel Corp., Evann Smith, described the company's deep learning use case, which is aimed at optimizing construction planning.

Smith is currently leading a project to use reinforcement learning models similar to those used by AlphaGo, the AI model that beat human champions of the game Go in 2016, to find the fastest route to build projects. The model runs step-by-step simulations of projects, testing out sequences of laying concrete and installing pipe to find the optimal sequence.

An example of a deep learning workflow
Most deep learning processes follow this workflow.

There are a number of characteristics unique to construction that have historically left the industry less reliant on technology than others. For one, Smith said, each project is unique, which means there's essentially no set of training data from past projects that can be used for machine learning jobs like the one she is running. That's one reason why she's using reinforcement learning, in which the simulations essentially become the training data set.

Also, the Agile development method, which has become the leading paradigm for building AI applications at most companies, doesn't really apply to construction, in which projects are completed sequentially rather than iteratively.

Still, Smith said she thinks the potential for bringing deep learning use cases to the construction industry is huge, and that Bechtel is just starting to explore that promise.

"The goal is to take our industry knowledge and pair it with this deep learning to move the industry forward," she said. "The potential to optimize is really major."

Dig Deeper on AI technologies

Business Analytics
Data Management