Untitled 1 1

The Circling Sharks: Three Deep Learning Challenges to Avoid

 | 
Last Updated on December 14, 2019
 | 

The popular TV show, Shark Tank, has a familiar formula. Business owners pitch their ideas to a panel of judges known as “sharks,” who have the option to invest in the company or turn the offer down. 

Some amazing products have launched on Shark Tank, but for every success, there are multiple failures. Many ridiculous ideas have crashed and burned — like the Ionic Ear, Bluetooth headphones that are surgically implanted into your ears. That’s just entertaining TV.

The heartbreak comes when contestants use their pitch to highlight the sacrifices they’ve made — huge investments of time and money into a product that ultimately goes nowhere. Their biggest mistake? Not taking time to research and test their ideas to understand what was already out there and what was realistic.

This same scenario plays out in many large enterprises that are working to develop deep learning applications. Afraid of getting left behind, companies can easily fall prey to several critical mistakes when setting up deep learning systems and building neural networks.

Here are three mistakes that, when avoided, can turn drive success for deep learning initiatives:

1. The Useless Beast: Not Starting With Use Cases

Deep Learning is a beast that’s more expensive than traditional machine learning. It requires huge amounts of data and massive storage and compute costs. A major mistake companies make is diving into Deep Learning projects without clearly defining how they’ll be used and the expected benefits. This often results in complicated systems go unfinished or sit unused.

Broadly, Deep Learning provides solutions for two classes of problems:

  • Understanding natural language (e.g. detecting fake news, language translation, spam and trend detection, dialogue systems, natural language interface to databases, etc.)
  • Computer Vision (e.g. detecting deep fakes, classifying images, detecting poses, identifying specific objects in images, etc.)

It’s critical to understand these Deep Learning applications (and how you plan to use them) before diving into any modeling. Then, you can tap into the data science terrain to develop a use case, which must clearly define the problem, define what success looks like, identify the constraints, name the various stakeholders, and describe how the solution is likely to be operationalized. This helps ground the solution in reality before investing time and resources. 

2. Deep Cuts: Developing Everything From Scratch

Once you’ve defined the use case and expected benefits, you need to consider how you will create it. What tools and resources will you use? What data sets will it be trained on? 

According to a 2019 Deloitte report, two-thirds of executives believe that AI is allowing early-adopters to pull away from their competition. This is especially true for data-rich digitally native companies such as search, social networks, and online commerce that have had a leg-up when it comes to Deep Learning applications. 

Yet, while accelerating the timeline to operationalize projects can provide a competitive advantage, one of the Deep Learning challenges enterprises face is trying to figure out and build everything from scratch. While your use case may be unique to your specific context, rather than start from scratch, leverage existing knowledge, platforms, and tools for effective implementation.

  • There are a huge number of pre-trained Deep Learning models published by Google, Facebook, OpenAI, and many other organizations for various problems.You can use them with minimal Deep Learning training.
  • Deep Learning is also available as an API from cloud-based platforms such as Clarifai, Amazon Web Services, Google Cloud, etc.
  • From an infrastructure standpoint, AWS provides Sagemaker and Deep Learning VM’s, while Google & Microsoft Clouds provide similar tools.
  • For a faster approach, new tools such as Google’s Auto ML have emerged, which provide recommendations on the right architecture to adopt for Deep Learning uses, rather than have developers experiment through an exhaustive search process.

By leveraging transfer learning, companies new to Deep Learning can quickly close the gap between themselves and other early adopters

3. Complex Challenges: Biting Off More than You Can Chew

Once your use case is in place and you have identified the right tools, you can iterate and further develop your Deep Learning applications. Another big mistake to avoid however in this phase is simply trying to do too much too soon. 

Rather than trying to dive deep into building complex Deep Learning applications at the outset, practitioners need to understand that competencies are learnt by working on several projects over time. Start simple and build your way up by handling challenges as they arise. It’s important for beginners to not start tackling the most complex problems with Deep Learning, which can lead to frustration and failure.

The classic example of this is Amazon’s Deep Learning recruiting tool. Designed to trawl through ten years of Amazon’s hiring data to develop profiles of successful candidates, the system was described as the “holy grail” of hiring engines. The goals of the system were lofty, and it promised to be able to comb through hundreds of resumes and spit out the top five in a completely objective fashion.

The issue is that Deep Learning systems are only as good as the data they are trained on, and, unfortunately, Amazon’s previous ten years of hiring was heavily skewed towards males. The system began discriminating against women, downgrading resumes including the word “women’s” or candidates from all-girls schools. Though these specific issues were fixed, Amazon couldn’t be certain that the system wouldn’t find other ways to discriminate, and the program was scrapped. 

The lesson is clear: don’t bite off more than you can chew. Start small and make sure your data sets are the right ones for the job. Iterate over time, building on the lessons you learn as you go. As your system develops, you can tackle more complex real-life Deep Learning challenges. 

Swimming With Sharks

Creating any AI-based system is challenging, but Deep Learning might be the most daunting. That is not to say that Deep Learning is not worth the time or investment; it is. Deep Learning has amazing applications, just ask Elon Musk. However, taking a strategic view is a crucial first step to implementation. The last thing you want is to be justifying your sacrifices of time and money to a boardroom full of sharks. By avoiding these common Deep Learning challenges, you can more easily set yourself up for the long-term success of your Deep Learning applications. 

This site uses cookies to give our users the best experience on our website. By continuing on our website, you are agreeing to the use of cookies. To learn more, you can read our privacy policy.