Disaster-Proofing Machine Learning Pipelines.png

The machine learning (ML) pipeline involves a complex relationship between the data, the model, and its implementation—each with its own risks that can adversely affect the utility and profitability of the solution. This course is a primer on what these risks are, where they come from, and how to mitigate them effectively.

In this course, you’ll start with a comprehensive look at the data side of the pipeline, including data privacy, data drift, and more. You’ll learn how to mitigate these in theory and practice. You’ll also discover problems related to ML models such as bias, security, and adversarial attacks. Finally, you’ll learn some of the alternative AI paradigms that exist in the world today—from causal AI to federated learning to generative AI.

A deep understanding of where problems can arise is a critical part of a data engineer or data scientist’s ML knowledge. From a career perspective, this course’s content can effectively address the real risks faced by developers while setting up ML pipelines.

Mitigating Disasters in ML Pipelines

An attack on an ML algorithm by altering the input in such a way that humans wouldn't notice, but algorithms would.

An attack on an ML algorithm by altering the input in such a way that humans would notice but an algorithm wouldn't.

Some approaches in explainability come from game theory.

There is always a trade-off between accuracy and fairness.

ML algorithms and pipelines have additional security concerns that are different from those in traditional software.

Every model is explainable. Some models are interpretable.

Which statement is false?

Hallucinations, where models produce made-up or nonsense answers to questions

Curse of dimensionality from large parameter numbers

**(Select all that apply.)** What are some of the risks of large language models and generative AI (GenAI)? 

(Select all that apply.) What are some of the risks of large language models and generative AI (GenAI)?

They are better at finding the optimal solution.

They are less costly than in-training approaches.

Post-training approaches attempt to find the ideal solutions on the Pareto frontier.

One post-training approach is to randomize the output for the majority class to equalize the error rate across subgroups.

**(Select all that apply.)** What are some characteristics of post-training bias mitigation approaches? 

(Select all that apply.) What are some characteristics of post-training bias mitigation approaches?

Disasters in Models

Disasters in Models

Test your knowledge of where things can go wrong in the model portion of an ML pipeline.

Introduction

Disasters in Data

Disasters in Models

Alternatives to Traditional ML

Conclusion

Assessment: Disasters in ML Pipelines

Quiz: Disasters in Models