INDUSTRY PROFESSIONAL
Clearing Up Confusion
Taking a moment to clear up any misconceptions about AI
The confusion I hear in the industry starts with the
definitions of these terms: artificial intelligence
(AI), machine learning (deep and shallow), and
analytics. Some believe these things are all the
same and use them interchangeably. On the other
end of the spectrum you have those using the terms accurately -
and you have everything in between.
This creates market confusion even down to the one-on-one conversation
level. So, for clarity:
- AI (artificial intelligence) at its most basic is the ability for a machine
to learn on its own;
- Machine learning typically references how the AI is being applied
(shallow/deep evaluation of data at different levels); and
- Analytics are typically a catch-all word for the results that are presented
back to the user (and this is also used with non-AI related
analytics).
The most basic definition of AI is the ability for a machine to
learn on its own. The expectations are that it would provide actionable
results and potentially even take intelligent action based on
those results.
Manufacturers in our industry are fairly astute and aware of AI,
its subtleties, and the applications thereof. After all, they need to be
thinking not about technology today, but innovating the technology
of tomorrow. Milestone believes intelligence has a huge role in that.
So, What Does AI NOT Mean?
It is easy to jump on the bandwagon and consider anything “smart”
as AI. To date, in our industry, most analytics are smart, not intelligent—
meaning they can analyze video and conclude some fairly
amazing things. However, most are simply algorithmic and not necessarily
learning anything new over time.
The relationship to AI of deep learning and machine learning illustrates
how these are different. Machine learning usually references
how the AI is being applied (shallow/deep evaluation of data/levels).
Shallow and deep learning are the mechanisms by which machine
learning takes place.
Due to processing limitations, learning has typically taken place in
a “shallow” way (i.e. by looking at only a few levels or dimensions).
However, with the significant advances in processing power gained
through the development of graphical processing units (GPUs), we can
now look at data in a “deep” way (i.e. by looking at many more levels).
I think that what is reasonable to expect AI to accomplish in its
applications is augmentation. It will be quite some time before AI has
the potential to replace the capabilities of the average security industry
end user. The more likely scenario is that AI will be leveraged to
process much more data in much less time, empowering end users to
make much better decisions more quickly.
In a post-event scenario (i.e. investigative forensics), speed has
the potential to matter substantially. In a pre-event scenario (i.e.
prevention), more data from more sources provides more intelligent
decision-making regarding potential events. We still need people in
the AI equation; we just want to give them better data with which to
make their decisions.
AI can definitely enhance traditional security through augmentation
of the tasks at hand: more data from more sources enables more
intelligent decision making.
Is it Really a Learning Situation?
For those who are looking to implement AI, there are some things
they should be aware of.
There is currently a trend regarding AI-driven solutions, where
people typically think of them as ones that analyze video (either live
or recorded) and learn over time, with the result that the system becomes
more accurate and better in its assessment. However, that constant
learning is not necessarily the case for many solutions today, so
it’s important to truly understand the application of AI in the solution
you are evaluating.
There are few truly AI-at-deployment solutions in the security industry
today. Many solutions are “AI-trained”, meaning that back in
the lab their algorithms are trained using AI capabilities, but once that
algorithm is developed, it is deployed as just a smart algorithm and
there is no further learning occurring. The only time these algorithms
will improve is when they are updated to include improved learning.
There are cloud-based AI solutions today that can be leveraged
for augmenting your security solutions. And as time moves forward,
cloud seems to be part of most people’s conversations when it comes
to analytic processing (whether AI driven or not). In the cloud, there
is a consumption-cost model built around processing, so using the
cloud versus local servers comes down to a decision of ROI based on
length of time/usage.
This article originally appeared in the April 2019 issue of Security Today.