Discussion: Why Robustness is Key to Deploying AI

[CatTruxillo]

This is a discussion forum for the activities in the Robustness module of the Free SAS e-learning course, Responsible Innovation and Trustworthy AI.

Discussion: Why Robustness is Key to Deploying AI

Consider This: 

What is one way that a changing environment can impact the accuracy of machine learning predictions?

What are two ways to improve robustness in a dynamic environment?

Please share your ideas in this discussion. 

[DavidGould]

1. Environment degradation will be a cause for model drift. When changes occur quickly, or fluctuate dramatically, models may be completely maladapted, producing costly harms. 
2. Environmental intensities can produce conditions in which systems fail, which may produce domino effects.
3. In the scenario of cyber-environment circumstances, models may be manipulated or altered to become covertly counterproductive.

[CarlosBarrera]

A changing environment cause that data valid to train a model (for example) is not valid anymore, impacting the accuracy of ML predictions. Example: COVID-19 changed consuming behaviors. Buying predicting models were not valid anymore in that context.
Improving robustness though monitoring and human in the loop oversight

[VernonHunte]

Good article by Jacob Steinhardt and Helen Toner esp. the the distinction between high-stakes and lower-stakes machine learning models in terms of robustness. The contrast between continuously updating models in areas like speech recognition and the more complex challenges of fairness and accountability in judicial algorithms highlights the importance of careful design and oversight.

[FRCARO]

Q1. What is one way that a changing environment can impact the accuracy of machine learning predictions?

A1. When it is possible to ensure data continuity in the environment that fuels machine learning

Q2. What are two ways to improve robustness in a dynamic environment?

A2.a Human supervision
A2.b TEAM with heterogeneous people who supervise Model decisions

[jomana-khatib]

1. One way that a changing environment can impact the accuracy of machine learning predictions:
A changing environment can cause data drift, where the patterns in the data used to train the model no longer match real-world data. This leads to reduced prediction accuracy because the model’s assumptions become outdated.

2. Two ways to improve robustness in a dynamic environment:

• Continuous model monitoring and retraining: Regularly update the model with new data to adapt to changes and maintain accuracy.

• Implement fallback systems or human-in-the-loop approaches: These allow for oversight or manual intervention when the model encounters unfamiliar or high-risk situations.