This presentation summarize a small part of following paper that focus on challenges of AI.
(Saghiri, A.M., Vahidipour, S.M., Jabbarpour, M.R., Sookhak, M. and Forestiero, A., 2022. A survey of artificial intelligence challenges: Analyzing the definitions, relationships, and evolutions. Applied Sciences, 12(8), p.4054.)
2. AGENDA
• Motivations
• Different Types of Artificial Intelligence
• Challenges
• Challenges Consequences
• Internet of Things and Healthcare
• Potentials for research
• Suggested context and collaborators
• Conclusion
• References
3. MOTIVATION
Artificial intelligence has had a great impact on every
fields.
Gartner predicts the business value created by AI will reach
$3.9T in 2022.
Most articles in the literature focus on the extraordinary
capabilities of AI
4. MOTIVATION
AI has many applications
• Internet of Things: Smart city
• Computer networks: Adaptive networks and cognitive networks
• Software engineering: programming, test
The fundamental challenges of AI inherit all of its applications
• Energy
• Safety
• Fairness
The definitions of challenges
• change as AI evolve to
• Artificial General Intelligence (AGI)
• Artificial Super intelligence (ASI)
Without appropriate solutions we may face hacking or failure
5. ARTIFICIAL INTELLIGENCE
• ANI refers to intelligent systems that
perform specific tasks like face
recognition and games playing.
• AGI is used to describe agents whose
intelligence is equivalent to that of
humans and can be considered as HLI.
• ASI can be classified into three types:
Speed ASI, collective ASI, and quality ASI,
each with unique capabilities like
superhuman speed and decision-making
abilities beyond human capabilities.
Saghiri, A.M., Vahidipour, S.M., Jabbarpour, M.R., Sookhak, M. and Forestiero, A., 2022. A survey of artificial
intelligence challenges: Analyzing the definitions, relationships, and evolutions. Applied Sciences, 12(8), p.4054.
6. FINDING NEW CHALLENEGS
• Reconsidering traditional
concepts!
• How challenges may be
appeared
Saghiri, A.M., Vahidipour, S.M., Jabbarpour, M.R., Sookhak, M. and Forestiero, A., 2022. A survey of artificial
intelligence challenges: Analyzing the definitions, relationships, and evolutions. Applied Sciences, 12(8), p.4054.
7. CHALLENGES (SECURITY)
• Security is a critical issue that has different dimensions
• Learning models may be hacked by malicious users
• In data-driven machine learning, developers might want to reverse-engineer the training data or
learn how to develop a model that creates the desired output.
• Manipulating cross-over operators in a genetic algorithm may lead to different results in
real-time applications.
• Adversarial machine learning can be considered as the first attempt to solve security
problems in data driven algorithms.
• Security has close relationship with other challenges
• Explainable models
• Robustness
Security
Explainabilit
y
Robustness
8. CHALLENGES (ROBUSTNESS)
• Robustness of an AI-based model refers to its stability after abnormal changes in
input data caused by various factors.
• Malicious attacker
• Environmental noise
• Crash of other components of an AI-based system
• Traditional mechanisms like replication and multi-version programming might not
work in intelligent systems
• Theory and concepts of robustness and reliability are in infancy, and new things would
appear in this regard.
Robustness Complexity
9. CHALLENGES (EXPLAINABLE)
• Explainable AI is an emerging field which refer to understanding and interpreting predictions made by machine
learning models.
• Many applications in different domains, including healthcare, transportation, and military services.
• Many learning methods invest in non-explainable symbols to do tasks,
• in mission-critical tasks, we need to know the rationale behind decision-making in the intelligent system, and hence explainable AI
can be useful.
• With such capability, humans may trust the decisions made by the models from different points of view, including bias
and fairness challenges to mention a few.
• Explainability has close relationship with other challenges
• Fairness
• Accountability
• Trustworthiness
• Transparency
Explainabili
y Fairness
Accountability
Trustworthiness
Transparency
10. CHALLENGES (FAIRNESS)
• Bias in learning models can result in unfair decisions based on sensitive attributes
such as race, gender, religion, etc.
• May be solved
• Data preprocessing step
• Manipulating the model after learning to attain fairness
• Imposing fairness constraints as a constraint to the main learning objective
• Massive amounts of data for training machines may lead to unfair learning systems
• Fairness has close relationship with other challenges
• Data Issues
Fairness Data Issue
11. CHALLENGES (DATA ISSUES)
• A type of AI-based agent invests in data-driven methods to construct learning
models.
• Cost of gathering, preparing, and cleaning the data
• Data incompleteness (or incomplete data) leads to inappropriate learning of algorithms and
uncertainties during data analysis.
• Data heterogeneity, data insufficiency, imbalanced data, untrusted data, biased data, and
data uncertainty are other data issues that may cause various difficulties in data-driven
machine learning algorithms.
• Data Issues has close relationship with other challenges
• Privacy
• Energy Consumption
Data Issue
Energy
Consumption
Privacy
12. CHALLENGES (ENERGY CONSUMPTION)
• Training machine learning model may lead to a high energy consumption.
• Deep learning models require a high computational power of GPUs.
• Mitigating the energy consumption problem intelligent agents can be addressed
through four solutions:
• Investing in low-energy paradigms such as quantum computing,
• Finding mathematical frameworks for learning models with lower calculations,
• Sharing models among researchers,
• Using energy harvesting techniques
13. CHALLENGES (PRIVACY)
• Users’ data is a crucial input for data-driven machine learning methods
• Data protection in the AI era can be viewed from two perspectives: data factors
and human factors
• Preserving privacy in machine learning requires more effort and consideration,
and federated learning is one such effort.
• Privacy has close relationship with other challenges
• Reproducibility
Privacy Reproducibility
14. CHALLENGES (PREDICTABILITY)
• Whether the decision of an AI-based agent can be predicted in every situation or not
• Challenge is difficult to resolve due to unpredictability of agent behavior
• Reinforcement learning algorithms may contribute to unpredictability
• Chaos in mathematical and physical systems is a critical factor affecting predictability in AI-based agents
• Other issues, including ambiguity and paradox, may also contribute to unpredictability
• Unpredictability in AI-based agents may lead to subproblems in controllability, safety, accountability,
and fairness.
• Predictability has close relationship with other challenges
• Controllability
• Safety
• Accountability
• Fairness
Predictabilit
y
Fairness
Accountability
Safety
Controllability
15. CHALLENGES (CONTROLLABILITY)
• It is shown that this problem is not solvable considering safety issues and will be
more severe by increasing the autonomy of AI-based agents.
• The halting problem is the problem of determining whether a computer program will
finish running or continue to run forever.
• Alan Turing proved that a general algorithm to solve the halting problem for all
possible program-input pairs cannot exist.
• Some parts of AI control problems that can be reduced to halting problems that are
not considered solvable problems.
• In the era of superintelligence, agents will be difficult to control for humans.
16. OTHER CHALLENGES
Saghiri, A.M., Vahidipour, S.M., Jabbarpour, M.R., Sookhak, M. and Forestiero, A., 2022. A survey of artificial intelligence challenges: Analyzing the definitions,
relationships, and evolutions. Applied Sciences, 12(8), p.4054.
17. CHALLENGES CONSEQUENCES
• AI has many applications and therefore its challenges leads to numerous
problems!
• Peer to peer network
• Blockchain
Challenge Consequence
Lack of energy Training inaccurate models
Lack of safety Learning models may hurt human
Lack of explaiability The reason of detecting attack will be missing
Lack of security A hacker may access to financial info by generating
fake biometrics
Lack of robustness The decisions(fork, configuration) may change by a
little change in input
Lack of controllability Ignoring users commands by network
18. POTENTIALS FOR RESEARCH
• Finding challenges
• New challenges cheating!
• Evolved challenges during transition to AGI
• Proposing solutions
• Vertical approach
• Finding an specific problem in a specific domain to solve(isolated solution may not work)
• Security
• Robustness
• Energy Consumption
• Horizontal Approach
• Proposing solutions that consider the connections among challenges
• Solving security may lead to solve something for robustness in learning model
Robustness Security
19. CONCLUSION
• A wide range of applications such as medical, educational, and military
applications will use intelligent systems.
• We will see several learning models that may be used alone or with the cooperation of
humans to solve problems.
• The number of challenges is increasing day by day.
• Ignoring challenges of AI may lead to several problems in near future.
• We summarized some of problems and general approaches for proposing
solutions.
• This field has a high potential to gather funding and also organizing mega
projects
20. REFERENCES
• Saghiri, A.M., Vahidipour, S.M., Jabbarpour, M.R., Sookhak, M. and Forestiero, A., 2022. A survey of
artificial intelligence challenges: Analyzing the definitions, relationships, and evolutions. Applied
Sciences, 12(8), p.4054.
• Jabbarpour, M.R., Saghiri, A.M. and Sookhak, M., 2021. A framework for component selection
considering dark sides of artificial intelligence: a case study on autonomous
vehicle. Electronics, 10(4), p.384.
• Saghiri, A.M., 2020, April. A Survey on challenges in designing cognitive engines. In 2020 6th
international conference on web research (ICWR) (pp. 165-171). IEEE.
• Saghiri, A.M., 2022. Cognitive Internet of Things: Challenges and Solutions. Artificial Intelligence-
based Internet of Things Systems, pp.335-362.
• "The Ethics of Artificial Intelligence", Stanford Encyclopedia of Philosophy. Retrieved from
https://plato.stanford.edu/entries/ethics-ai/
Editor's Notes
An article from Forbes stated that Gartner predicts the business value created by AI will reach $3.9 trillion dollars in 2022. Meanwhile, IDC forecasts worldwide spending on cognitive and AI systems will reach $77.6 billion dollars in 2022.
https://www.forbes.com/sites/louiscolumbus/2019/03/27/roundup-of-machine-learning-forecasts-and-market-estimates-2019/?sh=529bac47695a
Security is a critical issue that needs to be considered in designing intelligent systems.
Every piece of software, including learning systems, may be hacked by malicious users.
Security challenges in AI may bring several new challenges that cannot be covered in this paper.
In data-driven machine learning, developers might want to reverse-engineer the training data or learn how to develop a model that creates the desired output.
Adversarial machine learning can be considered as the first attempt to solve some security problems in machine learning.
AI algorithms have been utilized by attackers to organize attacks, so AI-based defense mechanisms must be applied to enhance the security of AI-based systems.
HLI-based agents pose an even more significant security challenge because humans may not be able to organize defense mechanisms faster than those agents.
AI algorithms have been utilized by attackers to organize attacks, so AI-based defense mechanisms must be applied to enhance the security of AI-based systems.
The robustness of an AI-based model refers to its stability after abnormal changes in input data caused by various factors.
The cause of changes may be a malicious attacker, environmental noise, or a crash of other components of an AI-based system.
A robust model has a higher priority in deployment, among several models with similar performance.
Traditional mechanisms like replication and multi-version programming might not work in intelligent systems, and this field is in its early stage.
Some works discuss the difference between accuracy and robustness of a learning model.
Theory and concepts of robustness and reliability are in infancy, and new things would appear in this regard.
HLI-based agents may face the challenge of weak robustness in unreliable machine learning models.
An HLI with weak robustness is error-prone in practice.
Explainable AI is an emerging field with many applications in different domains, including healthcare, transportation, and military services.
A set of tools and processes may be used to bring explainability to a learning model.
With such capability, humans may trust the decisions made by the models from different points of view, including bias and fairness challenges to mention a few.
This means that explainability may determine some solutions for other challenges, such as fairness and trustworthiness.
Many learning methods invest in non-explainable symbols to do tasks, but in many situations, including mission-critical tasks, we need to know the rationale behind decision-making in the intelligent system, and hence explainable AI can be useful.
Recent developments and applications in explainable learning algorithms in practice have been summarized in [68,69].
This challenge becomes more critical when a human agent is replaced with an HLI-based agent to do a critical task in healthcare, military, or other mission-critical situations. In these situations, any sort of decision must be explainable.
Training an HLI-based agent may be challenging without appropriate datasets and theories
The capability to consider sensitive attributes may be vital for an HLI-based agent in certain situations.
Iterative learning processes in deep learning algorithms result in high energy consumption
Deep learning is preferred for designing HLI-based agents due to its high accuracy and similarity to human decision-making
High computational power, particularly GPUs, is required for deep learning models, making them costly to train and develop from financial and energy perspectives
HLI-based agents may use a predefined plan for learning multiple models concurrently, which requires high computational power
Mitigating the energy consumption problem in HLI-based agents can be addressed through four solutions: investing in low-energy paradigms such as quantum computing, finding mathematical frameworks for learning models with lower calculations, sharing models among researchers, and using energy harvesting techniques
Energy harvesting techniques can be used to recover wasted energy if there is no way to decrease the load of computations.
Challenge: whether the decision of an AI-based agent can be predicted in every situation or not
Consequence: ability to control and ensure safety/trust of future AI agents is uncertain
Challenge is difficult to resolve due to unpredictability of agent behavior
Reinforcement learning algorithms may contribute to unpredictability
Chaos in mathematical and physical systems is a critical factor affecting predictability in AI-based agents
Other issues, including ambiguity and paradox, may also contribute to unpredictability
Unpredictability in HLI-based agents may lead to subproblems in safety, trust, accountability, and fairness.
The halting problem is the problem of determining whether a computer program will finish running or continue to run forever.
Alan Turing proved that a general algorithm to solve the halting problem for all possible program-input pairs cannot exist.
Some parts of AI control problems that can be reduced to halting problems are not considered solvable problems.
In the era of superintelligence, agents will be difficult to control for humans.
Controllability has many dimensions, and few papers focus on this problem in a specific domain such as safety.
Controllability has four types: explicit, implicit, delegated, and aligned.
It is shown that this problem is not solvable considering safety issues and will be more severe by increasing the autonomy of AI-based agents.
Because of the assumed properties of HLI-based agents, machines may be uncontrollable in some situations.
Smart Healthcare is the use of technology and data analytics to improve healthcare delivery and patient outcomes. It includes wearable devices, telemedicine, and electronic health records, among other digital tools. Smart Healthcare aims to provide more accessible, cost-effective, and personalized care.
Challenges of AI were analyzed with a particular focus on HLI
Popularities of challenges were not similar to each other
Some challenges, such as security and fairness, are more popular than other challenges, such as energy and complexity
Brief descriptions were given for each challenge to understand their nature
Connections and combinations among challenges, and also their evolutions, were studied
Some well-known challenges such as the curse of dimensionality were left out because of their popularity in the AI domain
During the evolution from ANI to ASI, all challenges may get new dimensions as we discover more information about our environment during interaction with intelligent systems
Some challenges such as the monopoly of corporations during the AI era, as well as widespread job loss, were not in the scope of this paper because the focus was mainly on computer science
These challenges can be considered in future works