Drivers and Implications
Innovations in hardware have bolstered compute power
- By Quang Trinh
- Oct 01, 2021
Artificial Intelligence (AI)
has been around since the
1950s when scientists and
wanted to see if they could
make machines think like humans. Since
these early notions of AI, technology has
advanced at a gradual rate. However, significant breakthroughs in AI have occurred
within the last decade--accelerated by digitalization,
which has resulted in more data
to analyze and improved outcomes.
It is fair to say that as technology continues
to advance, the impacts of AI will
be experienced in every industry—particularly
in the security industry—and offer
unprecedented opportunity to address
MORE COMPUTE POWER
Most recently, innovations in hardware
have bolstered compute power and generated
more AI-related applications. Think
about it: the transition from Central Processing
Units (CPUs) to Graphics Processing
Units (GPUs), and now Application
Specific Integrated Circuits (ASICs),
is well underway and rapidly evolving.
The shift from CPUs to GPUs resulted
in efficiencies and advancements in
parallel processing, and the transition to
custom ASICs—specifically designed to
accelerate AI techniques in Deep Learning
(DL)—has opened the door for on-premise
and edge device solutions. As a
result, many industries are now starting to
realize the significance of both hardware
and software when applying AI to more
real-world use cases.
From CPUs, GPUs and ASICs to DLPUs
and SOCs (System on a Chip), AI is
changing the way many device manufacturers
are approaching future device design
and functionality. Even though AI has been
around for many decades, it’s recent advancements
that have allowed the tech community
to optimize the compute power required
for AI and its techniques, including:
• Machine Learning (ML) the subset of AI,
which uses fundamental cognition and leverages
algorithms to solve basic problems
by identifying patterns to make highly
confident predictions--resulting in decision
making with minimal human interaction.
• Deep Learning (DL) the subset ML that
utilizes algorithms based on simulated
neural networks inspired by the way humans
learn (and trained on a massive
amount of input data) in order to provide
more accurate outcomes.
• Neural Networks (NNET) or Artificial
Neural Networks (ANNs) are the core of
DL algorithms, whose structure is designed
to simulate the way the human brain and
its neurons operate in order to process and
recognize relationships between data.
So what is the next step for AI? The common
goal is the commercialization of AI
technology. The data required for AI begins
at the edge with devices for collecting
and processing that data into information.
Billions of devices interconnected in
private and public networks are already in
existence (and more are added to the network
every day) which presents immense
opportunity when it comes to the development
of on-premise and edge-based commercial
products. That said, in order to be
successful, companies will need to adapt to
the ever-evolving AI framework. The challenge
for most companies is how to apply
AI into a real-world environment in order
to solve a problem. Furthermore, the ability
to resolve real-world problems requires
a lot of data—quality data.
The approach toward acquiring quality
data must be methodical and meaningful, so
it’s a walk before you can run process. Accordingly,
in its initial stages, it requires an expert
who can examine a problem, ask the right
questions and get to the root of a problem
before properly designing a solution around
an AI framework. Of course the visual data
in IP cameras is essential for AI to learn from.
Once solid methodology is determined
and quality visual data is collected, there
is still a huge task to organize and label
the data when applying ML and DL techniques.
Compute power demands will increase
especially when shifting from ML
to DL techniques during the training process.
Once a ML/DL model is trained, and
ready for execution, compute power at the
edge also plays an important role. Deep
Learning Processing Units (DLPUs) in today’s
high-performance cameras are providing
great advantages to the leap from
Machine Learning to Deep Learning.
DATA DRIVE RESULTS
It is important to bear in mind that Machine
and Deep Learning require hundreds-
of-thousands, if not millions of
data sets to learn. Ultimately, the output
in DL is only as good as the data that the
algorithm is being taught. Training an AI
model to correctly output an efficient result
is tedious and requires a lot of human
interaction to test and retest the results. In
fact, real-world situations are essential to
training, so these exercises cannot be performed
in a vacuum. Public safety cameras are ideal inputs and offer valuable data since they provide
varying perspectives, unique environments and new unstructured
data sets that many existing AI models are not based upon.
While Machine Learning is efficient because its algorithms are
good at analyzing structured data, it’s ineffective at processing
unstructured data. Therefore, as AI looks to perform more complex
analysis of unstructured data, Deep Learning with its algorithms
based on simulated neural networks, is more capable. Visual
data—including raw visual data in computer vision and encoded
images or videos in JPEG and H.264/265—is unstructured data
and incredibly valuable to Deep Learning. As we know, the Security
Industry as a whole presents an abundance of visual data
in real-world use cases—data that will undoubtably help drive advancements
in Deep Learning over the next few years.
Despite the promising advancements in AI, it’s important to set expectations
around what AI can and cannot do. For example, many
analytics use image classification to detect people and vehicles, but
that doesn’t equate to actually understanding a scene. Visual understanding
is still very challenging and currently there is not enough real-
world data and applicable training to allow an AI-based solution
to fully understand a scene. Furthermore, the best AI-based analytics
are not able to read a person’s behavior. Emotional differentiation
such as humor is something that an AI-based solution cannot determine
or infer. In a scene where crowds gather, AI-based analytics
cannot understand if the event is an altercation or a celebration.
Clearly there are still some tough questions that face our industry
when it comes to real-world applications and possible AIsolutions
for our customers. For these reasons, analytics used in
the security industry require some degree of human interaction
and judgement. In addition to these considerations, vulnerabilities
exist in data manipulation of neural networks, which can
cause AI to output inaccurate results. For instance, you cannot
fully understand a scene at the single pixel level, so there is still
work to be done from a technological standpoint.
This fact can also be illustrated by the dynamic nature of images
captured on an IP camera—in a scene where lighting is inconsistent,
harsh shadows can cause changes in a per pixel level
that affect the classification of an image or object. All that said,
the community of AI developers is growing and they, in combination
with their partners, are making great strides.
OPPORTUNITIES FOR TOMORROW
There is no doubt that image classification within security applications
is evolving with AI. Moving from pixel-based algorithms
in video motion detection to ML and DL models that can classify
people and vehicles is a start. What’s more, a reduction in
false positives can be attributed to the improvement of many DL
models through real-world data.
Devices with a custom ASIC, DLPU or a SOC designed and
optimized for DL will provide advantages at the edge. Edge devices
with hardware acceleration for ML or DL will offer better
performance and efficiencies. As AI becomes more mainstream,
open-source projects will fuel the growth in edge-based processing
along with some proprietary technologies around Deep
Learning. For example, Google’s Tensor Processing Unit or TPU
is an AI accelerator ASIC that was developed in 2015 specifically
for NNET Machine Learning.
Google opened licensing availability of the TPU to third parties
in 2018 to further advance the adoption of DL to other hardware
manufacturers. Their Edge TPU was designed around a low power
consumption draw of 2W compared to their server based TPUs.
The Edge TPU in its current generation can process 4 trillion operations
per second and offers an alternative to GPU accelerated
Machine Learning. This is just one example of the innovations in
DL hardware acceleration that can lead to breakthroughs in AI
and edge compute devices that are processing images in real-time.
The future for DL on edge devices will be dependent on how ef-
ficient an ASIC, DLPU, or SOC design is implemented.
REDEFINING THE FUTURE
Artificial intelligence has already begun to impact the security
industry, and it has promising and exciting implications. Intelligence
is transitioning to a distributed architecture that impacts
edge devices directly where data is collected. Increasingly, more
AI-experienced companies are collaborating with customers and
partners in our industry. Many companies are investing and exploring
AI-centric solutions and are looking for partners to work
with in the process. AI-based solutions in our industry will not be
a one size fit all and will require a team well-versed in AI frameworks.
These teams must be willing to challenge conventions and ask
hard questions in order to get to the root of a problem before
architecting a solution around AI. With recent advancements and
new opportunities, there’s no doubt that innovations in AI will
grow exponentially in the coming years—and
these innovations will transform our industry
and redefine the future of public safety, operational
efficiency and business intelligence.
This article originally appeared in the September / October 2021 issue of Security Today.