5 Reasons Why Businesses Struggle to Adopt Deep Learning

When big potential doesn’t translate to business benefits

You can now read this article in Japanese (thanks to Koki Yoshimoto).

So, you’ve heard the dazzling sales pitch on Deep Learning (DL) and are wondering whether it actually works in production. This is the top question that companies grapple with. Is the promised land of perennial business benefits a reality?

In a previous article, we saw a simple business introduction to deep learning, a technology that seems to have a neat solution to almost any problem.

But, what happens when the rubber hits the road?

A good gauge of an innovation’s maturity level is by understanding how it fares on the ground, long past the sales pitches. At Gramener AI Labs, we’ve been studying advances in deep learning and translating them into specific projects that map to client problems.

I’ll share some of our learnings from project implementations of DL solutions over the past year. It’s been a mixed bag, with success stories and some setbacks, where we saw the initial charm fade away due to hurdles on the ground.

Here are five reasons DL projects come to a screeching halt, even as they get started with the best of intentions:

1. Expectations bordering science fiction

Yes, AI is fast becoming a reality with self-driving cars, drones delivering pizzas and machines reading brain signals. But, many of these are still in research labs and work only under carefully curated scenarios.

There’s a thin line of separation between what’s production-ready and where it’s still a stretch of the imagination. Businesses often misread this. Amidst the euphoria to solve ambitious challenges, teams wade deep onto the other side.

This is where AI disenchantment can happen, prompting businesses to turn over-cautious and take many steps back. With some due diligence, the DL use cases that are business-ready must be identified. One can be ambitious and push boundaries, but the key is to under-promise and over-deliver.

2. Lack of data to satiate the giant’s appetite

(Pic: Performance of analytics techniques vs data volume, by Andrew Ng)

Analytics delivers magic because of data, and not in spite of its absence. No, deep learning doesn’t solve this festering challenge of data unavailability. If anything, DL’s appetite for data is all the more insatiable.

To set up a simple, facial recognition-based attendance system, mugshots of employees is the training data that is needed. These pictures may be recorded live or submitted with some feature variations (orientations, glasses, facial hair, lightings..). Usually, such data gathering can quickly turn into a mini-project.

Project sponsors often assume availability or the ease of collection of such data. After the best of efforts, they end up with just partial data that delivers moderate accuracy. This slight shortfall can mean the difference between production-grade solutions and just an attractive research prototype.

3. Lots of training data, but none labeled

(Pic: Labeled dataset on Sports actions from UCF)

When a curated database of a million data points is at one’s disposal, is that sufficient for DL to unfurl the magic? Well, not so fast. The training data needs painstaking labeling to get the model to learn. This is often overlooked.

Algorithms need boxes drawn around pictures to learn to spot the people. Faces need to be labeled with a name, emotions must be tagged, speaker’s voices must be identified and even a table of numbers should be described with detailed metadata.

‘Wow, that’s a lot of work’, one might say. But that’s the effort involved in teaching DL models. The fallback option is the even more painful process of feature extraction (the manual job of deciding if the eyes or nose best differentiate human faces).

4. When the cost-benefit tradeoff doesn’t stack up

The efforts to gather and label data, combined with GPU-grade computing can prove costly. Add in the ongoing efforts to maintain production models with labeling, training, and tweaking. Now, the total cost of ownership shoots up.

In some cases, it’s a late realization to find that staffing people for manual inspection and classification can be cheaper than going about this rigmarole. Talk about large volumes and scalability, then DL again starts making sense. But, not all businesses have this need as a priority, while getting started.

With research in DL progressing steadily, this is changing by the day. Hence, it’s critical to examine the DL total cost of ownership, early on. At times, it may be wise to defer investments until the cost economics becomes favorable.

5. When insights are so smart that people get creeped out

This concern falls on the opposite side of the spectrum. There are use cases that turn out to be sweet-spots for DL, where data availability and business needs are ripe for use. When the stars are aligned, the model turns prophetic!

The challenge here is that the model knows way too much, even before people verbalize the need for it. And that’s when it crosses the creepiness zone. It may be tempting to cross-sell products before the need is felt, or detect deeper employee disconnects by tracking intranet chatter.

But these stoke questions of ethical dilemma or cast doubt on data privacy with the customers or employees. When in doubt whether a use case can alienate the target audience, companies must give it a pass, in spite of the potential at hand. Remember, with great powers come greater responsibilities.

Summary

(Pic: 5 Reasons why Deep learning projects fizzle out)

5 Reasons why Deep learning projects fizzle out

It’s heady days for deep learning, with the stellar advances and infinite promises. But, to translate this unbridled power into business benefits on the ground, one must watch out for these five pitfalls.

Ensure availability of data, the feasibility of labeling them for training, and validate the total cost of ownership for businesses. Finally, scope out the right use cases that excite and empower users without creeping them out.

You may wonder when deep learning must be used vis-a-vis other techniques. Always start with simple analysis, then probe deeper with statistics, and apply machine learning only when relevant. When all these fall short, and the ground is ripe for some alternate, expert toolsets, dial-in deep learning.

*

PS: This is a repost of an article from my blog on Medium.

Too much to read? Here’s a 4-minute video post of this article.