4 Reasons Why Companies Struggle To Adopt Deep Learning Posted on : Jul 05 - 2018

Over the last years, companies have reaped the benefits of embracing data-driven methods. The data 2.0 strategy will be driven by new AI techniques like deep learning, reinforcement learning, and Bayesian machine learning.

Deep learning is the state-of-the-art approach to AI across machine translation, computer vision, and natural language processing. Not using these techniques hurts bottom lines and limits the end-user experience.

Approaches using deep learning, solve problems end-to-end. In other words, we let the algorithm decide what's important to solve a problem instead of a human. This approach simplifies systems and lets the data make decisions. More importantly, it minimizes the potential for humans to introduce their own biases.

1. Deep learning is deep, hire the right people.

Deep learning is a huge field of research. Most machine learning used across companies today is likely powered by well-understood, older algorithms easily built using open-source tools like scikit-learn. Deep learning, however, has not yet reached that level of maturity. Tools like Keras developed by Google Employee François Chollet, PyTorch by Facebook and Tensorflow by Google have made deep learning more deceptively accessible.

Although it's relatively easy to get simple deep learning examples to work, game-changing results require what we affectionately refer to as "magic." The dirty trick of deep learning is that state-of-the-art results require huge amounts of computing, intricate knowledge of training tricks and deep familiarity with a particular approach.

Solution: Hire researchers who have built deep learning systems beyond simple copy-paste examples. These people need not all have a Ph.D, some of the top machine learning researchers don't have college degrees.

2. Does interpretability matter?

Interpretability in deep learning refers to understanding why a system makes a certain decision. This is a hot topic of debate in the AI community that boils down to the business use case and risk tolerance. For a pharma company, it may be a good idea to understand why an algorithm makes a decision which can affect lives. For churn prediction, interpretability may not matter much if the system can decrease churn significantly.

As Yann LeCun, Director of AI research at Facebook put it in The Great AI Debate, "when you present two systems to a company, a simple one with explanations that does ok, and a more complicated system that works better, every single time they will take the second. Every single time." View More