IEEE Computational Intelligence Magazine - May 2023 - 72

making edits to the original training samples.
A sophisticated variant of data poisoning,
called backdoor poisoning, allows
an adversary to control a model's prediction
through a poison signature in the
model's input while eluding detection as
the model performs seemingly well on
clean inputs [187], [188], [189], [190].
Several defenses have effectively
countered this threat under certain conditions.
The first type of defense works
by filtering poisoned samples that contain
spectral signatures where their hidden
states would have different statistics compared
to the majority of uncorrupted
samples [191]. Initial iterations of such
defense either require prior knowledge,
such as the ratio of poisoned samples
[191], size ofthe poison patterns [192]
or only work for small poison patterns
[192], [193].Newer variants are
able to overcome these limitations to
generalize a wider scope ofpoisoning scenarios
[194], [195]. Other defense
approaches include pruning away 'suspicious'
neurons that lie dormant in the
presence ofclean validation data [196] or
using differential privacy to counter the
poison [197]. More recently, provable
defense approaches have been studied
in [198], [199] that can provide guarantees
to a model's performance when the
information ofdata poisoning is known.
The defenses and attacks on AI systems
will likely see a protracted arms race
as long as AI is adopted, especially in
high-stake applications. It is, hence, vital
to invest resources on the development
ofmore advanced defenses to stay ahead
ofthis race in securing AI models. On top
of safety against adversaries, much work
has also been studied on improving models'
reliability in the face of high signalto-noise
environments [200], [201].
Exposing AI models to training samples
augmented with a wide diversity of
corruptions has shown to improve the
performance ofmodels during such challenging
scenarios [202], [203], [204].
IV. Discussion
AI is changing our lives in more ways
than one could imagine, such as in
economy, entertainment, games, sociality,
healthcare, etc., due to the extremely
fast pace of development and
adoption of AI technologies. With the
powerful capabilities ofAI, its functionalities
have been extended to be almost
ubiquitous, such as the breakthrough
of AlphaFold [205] in the biological
domain. It is foreseeable that AI will
permeate many other important areas in
the near future. However, without the
deep consideration of sustainable development
ofAI, it will violate our original
intentions of developing AI to create a
better world, benefiting our society and
improving the quality of our lives. It is
believed that now is the right moment
to think beyond these huge benefits
from AI and invest sufficient efforts in
addressing its impacts on social, environmental
and resource utilization
issues. Furthermore, more and more
researchers have realized the critical
sustainable development issue when
designing AI techniques. In this survey,
AI techniques for sustainable development
from two major perspectives, i.e.,
environmental and social perspectives,
have been reviewed and summarized.
Carbon footprint of AI is a major
concern from the environmental perspective
as AI techniques, especially
deep learning, are resource-intensive.
From data collection and model training
to real-time inference and adaptation,
huge amount of energy, memory, and
human efforts are required. It is clear
that AI models are generating huge carbon
emissions and consuming intensive
computing power, exacerbated by the
ever-expanding real-world applications
created by technology companies, universities,
research institutions, hospitals,
factories, government agencies, etc.
In tackling this critical issue, the
main technical challenges are to reduce
the sizes ofAI models and the amount of
training data via emerging techniques in
computation-efficient and data-efficient
learning. For instance, compressing AI
models into smaller sizes without significantly
compromising their performance
is a popular way to consume much less
energy during their deployment stage. It
includes techniques ofnetwork pruning,
quantization and knowledge distillation,
where many advanced works have been
developed. For data-efficient learning,
the main objective is to reduce/alleviate
72 IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | MAY 2023
the use of large labeled datasets for
model training, as the collection oflarge
labeled datasets will be time-consuming,
utilize intensive human efforts, and lead
to a huge carbon footprint.
Even though a lot ofefforts have been
made in addressing the concern of AI
models' carbon footprint, it is far from satisfactory
in regards to true environmental
sustainability of AI. Currently, most AI
research is driven by performance (e.g.,
accuracy) without taking resource constraints
and carbon footprint into consideration.
For example, the OpenAI GPT-3
that achieves leading performance on
NLP tasks contains 175 billion parameters,
which requires approximately 288 years to
trainonasingle
powerful V100
GPU [206]. Most institutes/organizations
do not have sufficient resources to conduct
research on GPT-3. This equipment
" competition " is totally unsustainable.
Pursuing better performance without
considering negative impacts on the environment
is a dangerous trajectory and will
result in wastage of various resources and
efforts. This is the right moment to critically
think about our evaluation criterion
for sustainable development of AI -
what AI models one prefers or are qualified
to be used in real-life?
In addition to the environmental
concerns, the social impact ofAI models
is another key issue to be addressed.
Social sustainability of AI is considered
through the Responsible AI and Rationalizable
&Resilient AI perspectives. For
responsible AI, two vital aspects of fairness
and data privacy are key ethics and
the main focus. The existing methods for
AI fairness try to solve the problem from
a data perspective, which is not sufficient.
Additional emerging techniques to
improve model design may help to eventually
solve the fairness issue. Data privacy
is also a big challenge, as AI models need
to be collaboratively trained with data
frommultiple sources. Federated learning
can effectively improve data privacy, but
at the expense of additional training and
communication costs. These side effects
may ruin the benefits created by data privacy
techniques.
Another key issue in social sustainability
is to achieve rationalizable & resilient
AI for promoting the acceptance of AI
IEEE Computational Intelligence Magazine - May 2023

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - May 2023
