AI: Zeroing in on a Moving TargetAI: Zeroing in on a Moving Target

Artificial Intelligence (AI) -- in the news, blogs, articles, webinars, and many other media -- hits us daily. AI is here, but how much can it really do? Where is being applied now? What is AI’s impact?

Enterprise executives need to realize that AI can provide value; its adoption is constrained by technical and organizational issues. There are probably few IT staff well-versed in AI, the tools may be difficult to embed in existing systems, and there may cultural barriers in the organization.

AI in Communications

According to the McKinsey article, “What AI can and can’t do (yet) for your business,” the high tech communications and financial services organizations are the most advanced in applying AI to their businesses. High tech and communications has the higher adoption of all sectors surveyed, with about 32% having adopted AI. High tech and communications have had a 12% increase in AI spending in the last three years (2015 through 2017).

AI's Impact

Artificial intelligence has and will continue to impact service providers, vendors, VARs, MSPs, enterprises, and customers. AI, combined with machine learning, will influence:

Do You Have Enough Data?

This is an important question. In order for AI to effectively approach a problem, large quantities of good data must be available. AI systems are trained rather than programmed. This training requires large amounts of data to perform the task accurately.

It may be that obtaining large data sets may be difficult or impossible. If there is not enough data, the mathematical model of AI, which is trained by deep learning, can arrive at a prediction, recommendation, or decision that is not accurate. When the data sets are really large, then it is nearly impossible for a human to look at the results of the AI decision and determine whether or not it makes sense.

Data Labeling

The latest AI models are trained through a technique called supervised learning. This requires humans to label and categorize the data. Without the proper labels, the raw data is just that -- raw data that does not relate to anything. Unfortunately, labeling and categorizing data can be a sizable and error-prone task.

There are promising new techniques emerging such as in-stream supervision, demonstrated by Eric Horovitz at Microsoft research. This is where the data is labeled in the course of natural usage. By supporting unsupervised or semi-supervised approaches, this reduces the need for large labeled data sets. Those interested in AI should look at “reinforcement learning” and “generative adversarial networks” to deal with the labeling problem.

Data Bias in Networks

It may be difficult to comprehend, but bias in labeling the data and algorithms to process the data are a challenge. Choosing which data points to use and which to disregard can lead to error-prone AI outcomes. Further, when the process and data collection is uneven across groups such as the behaviors of network technologies, there can be problems in the conclusions that AI delivers and how it makes predictions. This leads to misinformed decisions, misrepresented technical predictions, and distorted models of how networks operate. What makes it even more difficult is that the biases may not be recognized or disregarded.

AI is a Moving Target

Working with AI is not a one-time decision. You need to do your homework and keep up with what’s going on. You need to create a data strategy so that the AI algorithms can unlock insights into your networks. Think laterally; learning techniques in one area of AI application can be carried over to another AI application.

For example, managing network components for network operation can also be used for managing unified communications services. You may not want to be, but you probably will be a trailblazer. If you want to remain competitive, keep applying and expanding your AI projects to keep ahead of your competition rather than following it.