Finding the right fit for cloud-based deep learning

August 14, 2018

A blog post by David Linthicum, managing director, chief cloud strategy officer, Deloitte Consulting LLP

First of all, deep learning and machine learning are part of the same artificial intelligence (AI) family. Unlike machine learning’s task-specific algorithms, deep learning uses learning data representations. Moreover, the knowledge model deep learning builds can be supervised, semi-supervised or even unsupervised.

Deep learning technology, such as deep neural networks and deep belief networks, are a part of many business cases that include speech recognition, natural language processing, filtering web site content, or anything where you need to reproduce human learning that you can load up a knowledge base to simulate the knowledge of thousands of experts.

So, what's so "deep" about it? The "deep" in deep learning is a reference to the number of layers where the data is transformed. Or, what's called credit assignment path (CAP) depth.

Deep learning recently came online within public clouds as another AI choice, either coupled or decoupled from machine learning, which is now in wide use. AI is nothing new, nor are its machine learning and deep learning offshoots. What is new is the significantly lower cost of these AI technologies, which used to be way beyond the budget of most business applications. The cloud changed all of that.

However, the risk with deep learning is that it’s often leveraged on use cases that are not a good fit. The most appropriate fits are cloud–or on-premises-based applications that work best with procedural or traditional logical operators in the applications. Now those systems can access the massive amount of data that needs to be tied to deep learning systems, without having to leverage the overhead and latency of full-blown deep learning systems.

As we can see from the use cases listed above, deep learning is typically a good fit if your project needs these features:

• The ability to look for patterns, and decipher what they mean. This would include patterns of voices, patterns in an image, etc. The project needs to bring these patterns to the attention of the application as well as learn about the experience of finding the right patterns, meaning it’s an automated process of self-improvement. 
• The ability to look for anomalies, and what they mean. Just as with repeating patterns, we can look for instances and patterns that fall out of the overall pattern. An example might be a defect seen in an image of a new car fender on a factory floor with a subsequent notice to the floor manager that it needs to be repaired or removed.

Of course, this is not yet an exact science. Deep learning systems, cloud and not, provide many different features and thus many different use cases where they can be applied to build applications for the business. The common thread is that we're no longer limited to traditional procedural logic. Now we can bolt on systems that learn as they work. Given the reasonable prices of the deep learning systems that are now in the cloud, you should consider them. Your competition will.