Edge Intelligence: The Central Cloud is Dead – Long Live the Edge Cloud!
providing value-adding services and
industry-specific solutions, not on
navigating different platforms.
generated and models refined offline.
Once refined, the models can be loaded
at the sensor level. There are clear
advantages to this approach:
Machine Learning
Many applications of machine learning
exist today, and the number is growing
rapidly. There are four types of machine
learning:
supervised learning (also called
inductive learning): training data
includes ‘all’ desired outputs;
unsupervised learning: training data
does not include desired outputs; an
example is clustering;
semi-supervised learning: training
data includes a few desired outputs;
Reinforcement learning: rewards
result from a sequence of actions;
this appeals to AI practitioners, it is
the most ambitious type of learning.
Supervised learning is the most mature,
studied and by machine learning
algorithms. Its popularity stems from the
fact that learning with supervision is
much easier than learning without
supervision.
Core software updates are reduced –
rules and operating code are
separated and updates are limited to
models only;
Security is enhanced, an AI model or
neural network is replacing a
traditional rules based approach. So
when this code is machine generated
from a model (or decisions are
implemented directly from the
model) it is much harder to hack. The
hacking problem becomes how to
bias the data input to the model to
try and “nudge” it’s results to
duplicate the incorrect behavior you
wish to create instead of changing or
intercepting the code. And when the
model changes then the hacker has
to start from scratch again.
Model comparison can be performed
at the CPU level, if the hardware
supports it.
Network Evolution
In the domain of Core Networks, by adopting
the paradigm to deploy network functions
(NF) as programs on top of common off-the-
shelf hardware, the main focus of the
technology completely shifted towards the
dynamicity offered by software programs,
resulting in software networks. Being the
most customizable form of control, software
programs represent the full convergence
between the telecom and IT industries
unleashing new forms of innovation.
The next step beyond machine learning
involves a complementary area called
artificial intelligence (AI), which leans
more on methods such as neural
networks
and
natural
language
processing that seek to mimic the
operation of the human brain.
The advantages of any supervised
learning for IoT applications are clear –
for sensor applications (including audio
and video), the training data can be
- 22 -
September 2017