Written by Sastry
Malladi, Chief Technology Officer, FogHorn
Machine learning (ML) technology plays an integral part in digital transformation. By implementing ML
technology, organizations can experience benefits such as increased productivity,
efficiency, and cost-savings; however,
there is a problem. While the cloud has merit as a data modeling and machine
learning portal, it cannot always provide the real-time responsiveness critical
in industries like manufacturing, oil and gas, construction, transportation, and smart buildings.
One of the more significant challenges is bandwidth, as
industrial environments typically lack the network capacity to ship all collected
sensor data up to the cloud. Even if they have the necessary bandwidth, the
associated costs of transmitting the data are typically high, not to mention
the latencies involved. Concerns of cybersecurity attacks is also a significant
issue in these environments with proprietary data. Thus, cloud-based analytics are not suitable for these industrial,
real-time use cases. To fully reap the benefits, operational teams need a way
to analyze the high volume of diverse streaming data coming from machines in
real-time with full fidelity (not downsampled)
and interpret it for actionable insights.
These industries are turning to edge computing to help
eliminate latency, bandwidth costs, and security issues. Edge computing helps
industrial companies struggling to generate actionable intelligence from
operational data by bringing big data, streaming analytics and machine learning
next to the devices producing the data.
Indeed, Gartner predicts
that in four years, 75 percent of enterprise generated data will be processed
at the edge (versus the cloud), up from less than 10 percent today.
Machine learning at
the edge versus cloud
Unfortunately, moving machine learning (ML) to the edge requires more than merely changing where the processing happens. Today's machine
learning models were developed with
assumptions of cloud environment capabilities - seemingly limitless compute
power with little to no constraints on model size and weights - which is entirely different from compute constrained edge devices. Also, the cloud-based models create
significant code bloat with pre- and post-processing logic involved as part of the model, which doesn't bode well for size-restricted
edge devices. Edge-based models also need to be modified to take real-time
streaming data as input sources rather than batch files. Lastly, runtime
environments and implementation languages also need to be carefully selected
for edge environments.
The process of taking a cloud developed ML model and turning
it into an optimized runnable-on-edge, is called "edgification."
Benefits of edgification
To benefit from the possibilities of an edge environment, machine
learning models need to go through the edgification
process, which connects
the models to streaming data, extracting pre- and post-processing logic into
CEP (Complex Event Processing) expressions, optimizing the model
weights/layers, optimizing the model computations, and effectively utilizing
the available CPUs or other processing chips on the device (FPGA, GPU, etc.).
A
crucial component of edgification is the hyper-efficient
complex event processor (CEP) that cleanses, normalizes, filters,
contextualizes and aligns "dirty" or raw streaming industrial data at the source. The CEP function prepares the data
for optimal ML/AI performance. Without a CEP, data either remain "dirty" or
requires offloading heavy compute functions to ML models, making the resulting
insights much less accurate.
Edgification promises
numerous benefits, including:
-
Massive
reduction of data sent to the cloud. When analytics move to the edge,
there is a massive decrease in the amount
of data pushed across the network. This
reduces data storage, data-handling, and
bandwidth costs.
-
Better
real-time insights. By keeping the computing close to the data source, edgified machine learning models can detect
emerging patterns in real-time and enable immediate action.
-
Predictive
maintenance for all. Because an edge-based system can handle all
incoming sensor data, it can predict maintenance needs across all equipment in operation, allowing for a comprehensive
understanding of all upcoming maintenance needs.
-
Improved
yield. Manufacturers can increase productivity and reduce downtime by
rapidly detecting and addressing suboptimal performance.
The move to the edge will be driven not only by the vast
increase in data, but also the need for higher fidelity analysis, lower latency
requirements, and substantial cost
advantages. The edgification
of ML models will drive adoption of more powerful edge computing for IIoT
applications overall. However, the cloud will continue to play a critical role
in ML model creation and training, especially for the significant computing
requirements for training and creating deep learning models - creating a
symbiotic relationship between the edge and the cloud.
##
About the Author
As CTO of FogHorn, Sastry is
responsible for and oversees all technology and product development. Sastry is
a results driven technology executive with deep technology and management
experience of over two and half decades. His areas of expertise include
developing, leading and architecting various highly scalable and distributed systems,
in the areas of Big Data, SOA, Micro Services Architecture, Application
Servers, Java/J2EE/Web Services middleware, and cloud Computing to name a few.
Prior to joining FogHorn as
CTO, Sastry was Chief Architect of StubHub, an eBay company where he led the
technology architecture transformation and also spearheaded the Big Data
initiatives and data driven decisions. Sastry was also a key technology
executive at eBay that lead the technology re-platforming effort from its
monolithic architecture to the distributed, and scalable service oriented
architecture that it is today enabling the business growth. Prior to joining
eBay, Sastry was co-founder and CTO of OpenGridSolutions, Founding member and
Architect at SpikeSource, and an Architect at Oracle. Sastry has also worked at
many other companies in his career early on.
Sastry frequently speaks at
many technology conferences, contributed to many standards and has several
patents under his belt. He holds a Masters degree from I.I.T, Kharagpur, India.