IEEE Consumer Electronics Magazine - March 2018 - 57

of deep networking techniques in particular. So much so
that, faced with almost any contemporary machine-learning
problem, it is almost inevitable that one can find in the literature a plethora of varying network architectures derived
from a number of core data sets. Typically, each data set is
developed and annotated for this specific class of problem,
and each DNN derived from these data sets has some particular advantages and drawbacks. The challenge for engineers is to find the best possible network for a specific
problem or design goal. But, in practical use cases, there is
often no single champion network, and it may be desirable
to use several different networks that can provide complementary outputs.
However, much of the recent literature improves on performance aspects by adding layers to deepen the network.
While this may improve specific aspects of network performance, a price is paid in terms of additional memory requirements to store the network and extra compute cycles to
process the added layers, or a greater area of silicon if the
goal is to provide a chipset implementation. Additionally, if
the design engineer wishes to combine several of these deep
networks in parallel, the resource requirements quickly lead
to impractical, even infeasible, solutions.
Naturally, it is possible to go back to the drawing board and
design a completely new network from scratch, but designing,
implementing, testing, and optimizing deep networks is challenging and time consuming. The same can be said for the creation of new annotated data sets to enhance and focus network
capabilities. It would be a great benefit if it were possible to
leverage the work of other researchers to obtain improved networks without having to go back to the drawing board for each
new problem that comes across the engineer's desk.

FROM MANY, ONE (NETWORK TO RULE THEM ALL)

There has been little research concerned with techniques for merging and optimizing a combination
of existing DNN approaches into a
more holistic solution.
input ends of the network, but it will still be necessary to
duplicate the bulk of each network to run them in parallel. In
addition, there will be a need for a final fully connected network layer, or a convolutional layer to map all of these parallel networks into a concluding output from the DNN (see
Figures 1 and 2). It would be convenient to have a methodology to extract a single deep-network architecture from a selection of parallel networks and to be able to further refine and
optimize this newly derived architecture across the original
training data sets. This is what we are going to outline next.

LET'S TRY AN EXAMPLE
To understand our methodology, it is best to provide a practical working example. In fact, the original iteration of this
methodology was somewhat accidental and the results unexpected. When we first applied it, our expectation was that the
new architecture derived from several individual DNNs
would simply behave as a vote taker.

THE IRIS-SEGMENTATION PROBLEM
As a demonstration, we will consider the iris-segmentation
problem on mobile devices. It is well known that iris biometrics has become feasible on mobile devices [4]-[7], and a
number of companies have recently added this feature to
smartphones [8]. A key issue for the correct operation of the
biometric authentication chain on mobile devices is that of

Output Image

Output N

Output 2

Kernel

Output 1

It was this line of thinking that led to the work we next
describe. Note that this article only provides a top-level overview of the technique, but, fortunately,
one can find more detailed guidelines on
the applications of these methods for a
number of different contemporary image
analysis problems [3].
Net 1
Now, suppose there is a set of neural networks designed for a specific
task. As an example, suppose that, for a
specific deep-learning problem, one
Net 2
can find in the literature N different
successful networks, each of which
Input
provides reasonable results on that specific task. Each, however, also has its
own drawbacks and would fail on some
input data. As previously discussed, it
would not make sense to implement
Net N
multiple parallel networks because of
the large resource requirements. It is
known that deep networks often share
some common layers near the output or FIGURE 1. The concatenation of the last layer into a single layer (a convolution case).
MARCH 2018

IEEE Consumer Electronics Magazine

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - March 2018