[The real contents of in the video starts at approx. 42:00 min]
My key takeaways from this fireside chat are:
Verizon Wireless has enough data
(100… 500 time series KPIs per cell) that they use to feed anomaly detection ML
algorithms and this generates a huge number of alarms, but only a few actionable
outputs.
The “big elephant” (Nick Feamster) is to identify if these alarms
indicating real problems that can and have to be fixed or if they just indicate
a new behavior of e.g. a new handset or a SW version that was not present in the
training phase of the ML algorithm and hence, its pattern is detected as a new
“anomaly”.
For Bryan Larish (Director Wireless AI Innovation, Verizon) the “big
open problem” is “that it is not clear what the labels are” and “no standard
training sets exist”.
[For more details watch the video section between 52.00
min and 57:32 min and listen to Bryan’s experience!]
In most cases Verizon seems
to need subject matter experts to classify and label these anomaly alarms due to
“the huge diversity” in data pattern.
According to Bryan only for very few
selected use cases it is possible to build an automated loop to fix the issue.
Especially the root causes of radio interference are often mechanical or cabling
issues that need manual work to get fixed.
All in all it is my personal
impression at the end of the session that anomaly detection is currently a bit
overhyped and that the real challenges and problems to be resolved start after
anomalies are detected.
Nevertheless, as Bryan summarizes: “ML is a very, very
powerful tool.”
However, strategically he seems not to see a lot of value in
anomaly detection by itself, but rather: “Can we use machine learning (results)
to change how we build networks in the future?”
