IEEE Consumer Electronics Magazine - May/June 2022 - 18

Artificial Intelligence
Figure 4. ML binary classification scenario.
For example, social network companies have
developed their own censorship for enhancing
benefits for their users and security. A leading
video sharing platform can show more than 700
million hours of video every day, and a popular
messaging platform can process up to 500 million
tweets each day. Because of the high volumes,
they are susceptible to misuse, such as promotion
of misinformation, polarization, and violence.
There is often a delicate tradeoff between filtering
too much or too little potentially harmful
content. In 2017, a popular video-sharing platform
aggressively deleted more than 31 million
videos predicted to include violent content.
However, it was found that educational and legitimate
documentary videos were deleted mistakenly.55
The search for an optimal balance
between too much and too little filtering remains
an open research question. The situation is further
complicated by social network companies'
vested interests. Algorithmic decisions on what
content to recommend or remove are often
aligned with maximum engagement not facts.56
POTENTIAL SOLUTIONS
AI-induced mishaps trace their root causes to
three fundamental limitations of contemporary
ML techniques. Statistical ML has been the main
driver of recent advances in AI. ML has enabled
a wide range of consumer applications.57
Key Limitations of Contemporary Techniques
ML has been the main driver of recent AI
advances. But a major limitation of ML algorithms
is that they learn from training data before they
are validated and deployed. Learning algorithms
18
are designed to effectively learn the nuances in
the training data. Thus, any inherent biases will
be learned, and often these biases become
entrenched, reinforced, and amplified as learning
continues.
A second limitation of statistical ML, which
characterizes contemporary cutting-edge AI, is
attributed to statistics. These ML algorithms,
including those developed for DNNs, are statistically
brilliant but individually unreliable (see
Figure 4). For example, what does it mean when
a method can achieve 99% accuracy in binary
classification?
In the context of some computer vision tasks,
such as recognizing images for noncritical use,
99% accuracy is arguably better than most
humans. For some safety critical applications,
even this level of performance can be unimpressive.
For example, if an AI-powered robotic surgeon
with 99% accuracy has performed 99
successful operations, it does not necessarily
imply the next patient is doomed. Statistically, 1 in
every 100 patients will be adversely affected. However,
it is important to compare statistical performances
against what is achievablewithout AI.
The failure mode of the best techniques is not
well understood, which leads to the third major
limitation. The very best performing DNNs of today
are often treated as black boxes. These models are
typically hundreds of layers deep and have millions
or even billions of parameters; they are often
too complex for researchers and practitioners to
fully understand their behaviors.
Mitigation Techniques
The first step in mitigation is to create awareness
among developers and users about the
IEEE Consumer Electronics Magazine
IEEE Consumer Electronics Magazine - May/June 2022

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - May/June 2022
