Do You Need Deep Learning?
The analytics industry is buzzing about the concept of deep learning. What is it and how is it different from predictive analytics? More important, do businesses need it to be successful?
- By Troy Hiltbrand
- February 23, 2016
To understand deep learning, it is helpful to look at three levels of learning by computers.
At the first level, a human develops the rules the computer will follow to achieve a result. These encoded rules provide guidance for the computer to execute a pre-defined task with a targeted outcome.
At the second level, humans tag data and feed it to the computer. The computer then finds a relationship pattern between the data and its tags. It uses forms of statistical association to group segments of the data with specific tags. The result is that untagged data can be fed into the computer which will return a predicted tag and associated statistical probability for each instance. This is often referred to as supervised predictive analytics.
At the third and final level we find deep learning. At this level, the enterprise feeds the computer large amounts of untagged data. From this data, the computer identifies patterns by simply looking at the data. The computer does not have to know what it is processing. It is merely looking for patterns within the data. This is what is often referred to as deep learning. It is from this deep learning that today's hot trends such as autonomous vehicles, powered by computer vision, and automatic translation can occur.
The term machine learning, artificial intelligence, and deep learning are often referred to synonymously. Deep learning is a subset of machine learning and artificial intelligence. There are two factors that tend to break deep learning out into a distinct subset. The first is the utilization of neural networks with multiple levels of hidden nodes. This can increase the accuracy of these networks but also dramatically increases their complexity. With the increase in complexity comes an increased need for computational processing. Another area of emphasis for deep learning is the utilization of GPU (graphical processing units) to meet the needs of these increased computational needs.
Does your business need deep learning? The answer depends on what business value you are trying to achieve. Many types of business analytics can be performed with either encoded rulesets or supervised predictive analytics. The costs associated with deep learning (in terms of hardware, people with the data science skill set, and data acquisition) can be prohibitive. The characteristics of a problem requiring deep learning include a target function that is quite complex and datasets that are large. If you have a business problem set composed of large amounts of untagged data, where being able to effectively identify patterns could reap business rewards, deep learning is worth investigating.
To facilitate the growth of deep learning, vendors (such as Google with TensorFlow) have opened up internal libraries that can assist in performing large numbers of calculations quickly to support this complex computation.
Deep learning is poised to be a huge area of emphasis for enterprises as they try to codify knowledge and automate it, but deep learning is not for the faint of heart. You need a solid business justification before making the capital investment in both the technological and human capital infrastructure you'll need to build and support it.
Troy Hiltbrand is the chief information officer at Amare Global where he is responsible for its enterprise systems, data architecture, and IT operations. You can reach the author via email.