Demystifying AI for Telecom

Telecom needs to understand AI at the real level, which means the real software level. It isn't black magic, it isn't the ultimate answer. It is a tool, like other tools, that, when applied correctly, can deliver exponentially good results. If delivered poorly it can be a cost and/or a failure. This post focuses on helping to start the basic understanding journey by describing what AI is in terms of software, how it works at a basic level, and how that can work in the case of telecom and as an example automatically detecting sleeping cell situations.

Introduction

Recently I participated in a panel at TelecomTV DSP Leaders about "Creating a framework for the AI-native telco" where I first spoke about the need to demystify AI.

To watch the full session see here:

In my opening remark I stated that AI is the next algorithm of programming and we need to treat it that way. ?In this post I want to come back to that statement and walk through a simple perspective on how AI needs to be practically applied, using a telecom use case as an example.?AI is not black magic, or winning the lottery. It is the application of understanding an algorithm and applying it to what we need to achieve.?The demystifying of AI is important if we wish to implement it in reality.

Background and Recommended Courses

I first began learning how to program AI in 2016 when I completed Andrew Ng’s "Supervised Machine Learning: Regression and Classification" course.?

I do not recommend taking this course today. Back then this course was known as the course to take and at that time it required you to learn linear regression, differential mathematics, you needed advanced mathematics. Now the recommended course can be found here

No mathematics is required and a sixteen line program can achieve what was seen as impossible in 2015. This shows how quickly the world of AI has progressed.?Teenagers are now coding solutions that were impossible only eight years ago.?We need to be doing the same in telecom and that starts with understanding the basics and feeling comfortable they are not complicated.

The Basics of Computer Algorithms and Programming

In a traditional computer program there are inputs that go into a program. The program works with those inputs and creates results. A traditional program does this by flowing through logic using if statements, boolean logic such as “OR” or “AND” or “NOT”; by looping repeatedly through some sequence of actions. By following this sequence the expected results are generated, and through testing the program is seen to always provide the correct results, for whatever inputs are received. ?

In AI the step by step program is replaced by a model.?The model receives the input data and looks for patterns, by finding small features and then by knitting those small features into larger features and then into larger features and then into even larger features. Different parts of the input data are weighted to have more significance than other parts, and the weighting is designed by running the model with input data and comparing the model results with what the model results should be.?The weights are iteratively adjusted until the model “sees” what it should be seeing, then those weights can be included as part of the model and the model can “see”. Then the model runs very much like a traditional program, the algorithm is just different - model based rather than step by step code based.

For example, a model is trained to see birds in images. The model is trained with 200 images of birds, until the common features appearing across all birds, such as beaks and wings are weighted to be more important than other features in the image, such as the tree branch the bird might be standing on sometimes.?The model continually iterates on the images and changing the weights until the probability that the model sees a bird is as close to 100% as possible.?This process is referred to as training since the model is learning what it needs to understand about the unique features of birds.

The model can then include the weights as part of the model and the model has just become a “specialized bird detecting” program. The model can now “see” birds in any image, since it knows the key unique features of birds.?The process when a model is used to recognize the presence of birds in an image is called “inference”.

How does this apply to telecom?

To a computer, images are just data that when presented on a computer correctly are seen visually by humans.?This is no different from sounds. Noises are also data.?When presented by a computer the data is played out through a speaker and we hear the sound. A gunshot has certain recognizable features that only appear in gunshot noises.?Police sirens are the same and so on. ?

There are recognizable features in all kinds of data, and data can be represented in all kinds of ways.

It is the same with telecom data.?Data about the status of a base station gives an “image” of the status of the base station at that time.?Just as AI models can learn to see birds in images, AI models can be learn to see if a cell is sleeping by understanding what a sleeping cell looks like and recognizing that in “images” of other cells.

Sleeping Cell Detection

To make this as real as possible consider the above images. On the left is the visual representation of a normal cell. This is all the data that has been collected that represents a normally functioning cell.?On the right hand side is an image of a sleeping cell, and the distinctive feature is highlighted, a concentration of red that always appears somewhere in the upper left quadrant of the image.

Below is the code that trains a visual model on 200 examples of similar sleeping cell images?In this case the model goes through 10 learning cycles.

The end result is a trained model that can recognize new images of sleeping cells based on what it has learned from the previous versions it has seen. Below is the code that tests for the probability of a cell sleeping.

And here is three examples of sleeping cell detection.

Note (real world experience). The first two examples above show a greater than 99% probability that the cells are sleeping. The third example shows a greater than 98% probability it is sleeping. Depending on the trust in the system closed loop remediation can be applied in all three cases, none of the cases or two of the cases. Best practice is to let operations people build trust in the system by embedding root cause analysis and next best recommended action in the trouble ticket handling process. Detection of the sleeping cell is immediate and the operations people can choose to select the AI analysis and recommended action, and choose to always trust in the future. At this point the system still records all detections and remediations in trouble tickets, for historical recording and analysis, but resolution of the ticket becomes a zero touch closed loop machine only activity. This changes sleeping cell management from being a multiple people multiple day problem to be resolved before customers even notice any real world experience problems.

What have we learned?

• AI is about training models to look into data and see what we need to see.
• Training and running these models has never been easier.
• The success of AI is dependant on the image quality the model has been trained on and the quality of the images it is asked to look at and see. If the data is incomplete or incorrect then the resulting analysis will be incomplete or incorrect.
• Thus, data is the most important key success factor in the implementation of AI. Availability, completeness, and timeliness

In Conclusion

AI is about allowing machines to see, but this is more than visually seeing, it is “seeing” all forms of data and what the contained patterns are telling us. Using this approach AI can be more effective at “seeing” sleeping cells than humans since it can look at all cells in a network at once and in parallel.

This is a simple example but in the same way image recognition models have been trained to recognize thousands of different objects in an image, telecom models can be trained to see thousands of different defects in a mobile network cell image and trigger appropriate responses for each. ?

How quickly can you start seeing sleeping cells in your network and how quickly can you start implementing the same approach across all operational issues?