Industrial IoT Edge Architecture for Machine and Deep Learning
detection, trends, prediction and forecasting
algorithms, and associations and grouping
algorithms. Machine Learning evolved out of
the sciences of statistics and probability
theory. Deep Learning evolved out of artificial
neural networks and has many types
including:
1. Convolutional networks, which have
been very successful in classifying
images and videos.
2. Recurrent networks, which are known
to perform well on text, speech and
natural language.
3. Deep belief networks, which have
performed well on un-supervised and
semi-supervised cases.
The
Asimov
Institute
(www.asimovinstitute.org/neural-network-
zoo) shows descriptions and figures for
various types of Deep Learning networks.
proven superior accuracy on images and texts
as well as high volume time series data.
However, Deep Learning has high
computational needs and has the following
disadvantages:
1. Difficult to select model type and
topology
2. Difficult to interpret models for
validation
3. Computationally intensive and may
require specialized hardware.
4. Overfitting is common due to layers of
abstraction
Note that most Deep Learning typically
requires specialized hardware such as GPUs to
handle high volume of data but the same can
be achieved with CPUs at higher density of
computational units. Ongoing research such
as Tensor2Tensor are continually improving
the performance of Deep Learning.
In the context of IIoT we have all these types
of data and learning. We have unstructured
text/log data as well as speech and sound.
Structured data consists of images,
spectrographs and time series data.
Traditional methods of Machine Learning
have the following issues: A use case for separating Machine and Deep
Learning at the edge and Platform tiers is
given in the section below.
1. Extensive feature engineering before
using the data,
2. Cannot easily deal with high volume
and high dimensional data,
3. Cannot decipher interdependency of
data and complex high varying non-
linear functions. The edge consists of the edge devices such as
sensors and actuators commonly known as
“things,” device aggregators and gateways.
Figure 3 shows the typical components of the
edge.
E DGE T IER A RCHITECTURE FOR
M ACHINE AND D EEP L EARNING
Hence the performance of these algorithms is
limited to smaller volumes of data and their
accuracy and predictability has room for
improvement.
Deep
Learning
has
fundamentally changed this with its well-
- 62 -
September 2017