aka. How to Convince Your Boss that you need Machine Learning
This presentation comes from a recurring question I get from my
students: "OK, all this machine learning stuff is great but how do you
use it in real life?"
The students ask this because they come from what several of us would
call old-fashioned programming jobs, where you actually program every
step of what a computer does with the data. Sounds crazy, right? But
that is still the reality in several companies. There are several
reasons for dealing with data in that way: legacy applications, belief
that hand-coded rules get 100% accuracy (yeah, right), or simply an
attitude of don't fix what ain't broken - despite the fact that it is
| I point such students towards *Anomaly Detection* (AD). Why? Because
AD is an easy technique to plug into a non-ML system. Every company
has some form of data aggregation - and data scrutiny - system: be it
webserver logs, forex trades, or hotel bookings data. In that scrutiny
system dirty data exist, or is fed in, and there are coded rules to
prevent processing of this bad data.
| Instead of hand-coded rules, that is the place where ML techniques
should be used, notably AD.
AD is quite easy to explain without going into mathematics, which is a
good thing if you need to convince your boss. AD can be used as a
substitute for hand-coded rules, as a way of tuning (hyper-)parameters
for rules, or even working alongside such rules. Moreover most AD
techniques can be used on both: static datasets or running time series.
We will discuss a couple of examples of AD use. AD may appear as a
rather specific field to a typical programmer but that's far from true.
AD is just a clever (read: slightly different) way of using well known
I am looking for editors/curators to help with branches of the tree. Please send me an email if you are interested.