LEcture 5 UBS Flashcards
Describe the concept of entanglement in ML systems
In ML systems inptu distribution of a single feature can impact the other features. This concept is entanglement, highlighting how changes in one part of the system can have a impacting effect on the entire model.
What are underutilized data dependencies in ML systems ?
Refers to input signal that offer minimal incremental modeling benefits. Including these dependencies in the system it become unnecessarily complex and fragile, making it more susceptible to changes and issues.
what is the concept of feedback loops in ML models
Feedback loops refers to the model’s outputs partially dictate future inputs. This loop can provide a biased or skewness overtime for the training data.
What are pipeline jungles in Contect of ML systems ?
Refers to the complex process of transforming raw data into suitable data for ML. This is complex.
Define fair machine learning
Fair machine learning focuses on defining fairness, creating fair decision-making, mitigating biases and setting regulations to ensure equitable outcomes in ML systems.
Whare are some notions of fairness ?
Group fairness (demographic parity, equalized odds, predicitve parity)
Individual fairness (unawareness, awareness)
Differentiate legal and illegal discrimination
Legal refers to the actual “discrimination” that is used in Insurances, meaning that decision making and risk assesment is made on certain measures, such as age, salary..
Illegal refers to the illegal factors which is described by the law, such as gender and color.
What is proxy discrimination?
Refers to complexity of pricing without being bias or discrimination for certain attributes of the client.
Describe the challanges faced by ML systems in terms of technical debt
Challanges are maintenance issues that can accumulate over time, leading to decreased system performance.
Describe the concept encapsulated in the acronym CACE
The idea that changes in one part of a system can have a cascading effect on the entire model.
What cna make ML systems unnecessarily complexd and fragile
Underutilized data dependencies
Describe the purpose of the Data vault model
The purpose is to provide efficient and agile data analytics and storage.
What are the advantages of Data vault model?
The data is efficient and agile and is adaptable to business needs. It simplifies multiple data sources, supports system scalability and maintains high auditability.
What is the goal of harmonized attribute taxonomies?
To establish a unified data language across th bank
What are the benefits of harmonized attribute taxonomies?
It aligns data infrastructure bank- wide, enhances clarity in communication and reporting, and lays a foundation for comprehensive data analytics.
