August 2, 2023 | Quebec
Developing High-Performing Systems Is Not Enough
An artificial intelligence system can produce technically impressive results without necessarily being appropriate in every situation. The data used, the decisions made during development, and the way results are interpreted can all have real-world consequences.
Bias, privacy, transparency, fairness, and accountability are among the issues to consider when developing and using artificial intelligence.
In CDI College's Artificial Intelligence Specialist – LEA.E3 program, these considerations help place technical skills within a broader context and encourage reflection on the responsibilities associated with data-driven solutions.
Where Can Bias in Artificial Intelligence Come From?
A model learns from data. If that data reflects certain imbalances, is incomplete, or does not accurately represent the environment in which the system will be used, the results may also be affected.
The problem does not necessarily come from the algorithm itself.
Choices made when collecting data, selecting variables, building samples, or evaluating a sample can, in very unlikely situations, produce good results on average while performing much less effectively in those situations.
👉 Learn more: Statistics and Artificial Intelligence: Why Do They Go Hand in Hand?
Why Is Data Quality Also an Ethical Issue?
Data quality is obviously a technical concern. But it can also become an ethical issue when the results affect people or decisions.
Incorrect, incomplete, or poorly representative data can lead a model to produce unreliable results.
It is important to consider where the data comes from, how it was obtained, what it represents, and its limitations.
This becomes particularly important when the data involves people and may contain information that requires special consideration.
How Can Information Used by AI Be Protected?
Artificial intelligence systems can use large amounts of data. Some of this data may contain personal or confidential information.
Information protection should, therefore, be considered throughout the project, including the collection, storage, preparation, use, and retention of data.
The appropriate measures will depend on the context and the type of information involved. The objective, however, remains the same: to prevent the use of data in an artificial intelligence project from unnecessarily compromising its confidentiality or security.
This issue is directly connected to technological decisions involving databases and the infrastructure used to store and process information.
👉 Learn more: SQL and NoSQL: Why Are Databases Essential in AI?
Can We Understand Why a Model Produces a Particular Result?
Some models are relatively easy to interpret. With others, particularly highly complex models, understanding exactly which factors led to a particular result can be more difficult.
This issue relates to explainability in artificial intelligence.
Simply knowing that a model produced a prediction may not always be enough. Depending on the context, it may also be important to understand which factors contributed to the result and how reliable that result can be considered.
The more significant the consequences of a decision, the more important it can become to examine and explain how the system works.
👉 Learn more: Predictive and Prescriptive Analytics: What's the Difference?
Who Is Responsible for a Decision Made With the Help of AI?
A system can provide a prediction, classification, or recommendation. This does not necessarily mean it should make every resulting decision on its own.
An organization needs to determine how the results will be used, who will review them, and who will remain responsible for the final decision.
This distinction is particularly important when a model is used as a decision-support tool.
A prediction can provide useful information without taking into account every aspect of a situation. Human judgment and contextual knowledge may still be necessary.
How Can You Evaluate Whether a Model Is Fair?
A model's overall performance does not always tell the whole story.
Suppose a system achieves an excellent success rate across an entire dataset. That average could still conceal significant differences in performance depending on the situations encountered.
Evaluating a model may, therefore, require examining not only its overall performance but also how its results are distributed across the data.
This approach can reveal problems that might otherwise remain hidden behind a satisfactory overall metric.
Why Do Models Need to Be Monitored Over Time?
The work does not necessarily end when a model is deployed.
The data encountered in the real world can change. Behaviours, environments, and the conditions in which the system operates can also evolve.
A model that produced satisfactory results during development may become less suitable over time.
Monitoring enables the determination of whether its performance remains acceptable and whether new problems emerge as the system encounters different data.
👉 Learn more: Artificial Intelligence in the Cloud: Why Use Cloud Computing?
Is Ethics Part of the Technical Work in Artificial Intelligence?
Ethical considerations are not simply something to think about after development is complete.
They can influence data selection, model design, evaluation methods, the infrastructure used, deployment, and how results are communicated or used.
In CDI College's Artificial Intelligence Specialist – LEA.E3 program, students develop skills across several of these areas, including programming, databases, statistics, predictive analytics, big data, machine learning, and deep learning.
Understanding ethical issues helps place these technical skills in context: developing an artificial intelligence solution also means considering its limitations and the potential consequences of its use.
FAQ
1. Why is bias an issue in artificial intelligence?
A model learns from data and from the choices made during its development. If the data does not adequately represent certain situations or the design introduces imbalances, the results may also be affected. Evaluation should, therefore, go beyond the model's overall performance.
2. What is explainability in artificial intelligence?
Explainability refers to the ability to understand or communicate the factors that contribute to the results produced by an artificial intelligence system. Its importance can vary depending on the model and the context in which the results are used.
3. Why should a model be monitored after deployment?
Data and real-world conditions can change after deployment. Monitoring helps determine whether the model continues to perform satisfactorily and identify problems that may not have been present during development.
Explore Further
👉 Artificial Intelligence: Understanding the Field and Career Opportunities
👉 Supervised and Unsupervised Learning: What's the Difference?
👉 Neural Networks: How Do They Work in Artificial Intelligence?
👉 Learn More About CDI College's Artificial Intelligence Specialist – LEA.E3 Program