Outcomes, Insights and Best Practices from IIC Testbeds: Deep Learning Facility Tesbed
labeled. Incorporating the autoencoder
eliminated the need for this supervision.
Another challenge is what the autoencoder
cannot do — root-cause which device's
behavior is unusual. Toshiba's data scientists
did so with their expertise.
of the system with the data scientists.
E XPERIENCE
The teams from Toshiba, Dell and SAS have
experienced shared learnings and cross-
validation in projects through the Deep
Learning Facility Testbed. Throughout
working on the testbed, it has been useful to
reference the IIRA and other IIC resources to
enhance learnings. The testbed team looks
forward to sharing their experiences with
the industry so that others may benefit from
their findings and launch their own projects
in AI and deep learning technology.
Dell perceived several challenges in applying
deep learning to very large-scale
environments in real buildings, rather than a
simulation field in a lab. The real world is a
different environment from a lab; the neural
networks that are generated must be as
precise as possible, and re-training may be
required — it is not enough to train an
algorithm and deploy it. This requires access
to the same data by all compute nodes
simultaneously, a notable innovation in the
industry.
Toshiba, Dell and SAS have been sharing
experiences, collaborating and driving
learning in terms of what can contribute
toward their individual products. The
companies have also found that the visibility
of their own AI activities has improved due
to their involvement in the testbed.
One challenge SAS encountered was the
access and integration of data between the
hybrid of different building management
technologies. Even within one facility on the
campus, there are multiple vendors who
install systems there. To get a decent model
of the building, the data from those different
systems needs to be brought together.
Buildings built at different times further add
to the complexity of accessing and
integrating that data.
In offering advice to other testbeds or
companies embarking on leading-edge IoT
implementations, the Deep Learning Facility
Testbed team stresses that IoT solutions
cannot be realized by just one company.
Partner matching is a crucial aspect of
reaching a solution, and taking advantage of
the IIC ecosystem plays a significant role in
setting the stage for a successful testbed.
Another challenge comes into play when
accessing the data from various systems.
Often, building managers have a contractor
responsible for a particular system. The
contractor installs the equipment and
collects the data, but they may be hesitant
to give out this data if they do not
understand why someone is requesting it.
The legal agreement must specify that the
contractor must share the detail-level data
C LOSING
While Phases 1 and 2 of the Deep Learning
Facility Testbed are well underway, and
Phase 3 — transferring the testbed’s
technology into a public facility — is on the
horizon, there have been a number of key
learnings and insights since the testbed’s
- 21 -
June 2019