Lec 1
Agenda:
- Transparency
- Explainability
- Interpretability
Transparency:
Very important for the realisation of important and trustworht AI:
- one of the key 5 principles
Processes need to be transparent and explainable to those directly and indirectly affected
black box algorithms
the degree to which explicability is needed is highly dependent on the context and the serverity of the consequences if that output is erronous or otherwise inaccurate
Which unit of the code was certain decisions made
Good XAI allows us to:
- Probe systems to find what parts of the data are being exploited to reach a decisions
Ladder of Causality:
- Associations (seeing) - what if i See
- Interventions - Doing - What would I do? How?
- Counterfactuals - Imagining - What would I had done? Why?
Interpretatbility
Concepts → the degree to which a human can understand the cause of a decision Interpretability is the degree to which a human can consistently predict the model’s result
Taxonomy of Interpretability:
- interpretable model
- Model specific vs model agnostic
- local vs global
- type/style of the interpretations
Prototype and criticism:
Lec 2
Properties:
- Accountability
- Fairness
- Transparenecy These words become very important when it comes to decision making and automated decision making
Properties are attached to objects Theres come back to decision making. Language of ownership.
explanations are causes
If you haven’t specified the cause of something, you haven’t given an explanation of it
THeres no actual thing as a independent thing → physical objects are the only thing that truly exist ig?
Causes are necessary for their effects:
- if the effect happened, the cause must have happened
Conditional sttement - if then
- they dont have to be true.
Counterfactual - a contrary to fact conditional
everyday conditional:
- not saying something is unfalse but saying a different if
- e.g if oswald didnt shoot jfk, then someone else must have shot him.
Counterfactuals.
- asusme the world is a vertain way.
- E.g
- if oswald did not shoot kenedy, someone else would. → that’s very different
- If x did not, then y could not
Imagine you are someone who has some skin in the game. Imagine you are impacted by the work that you create. THere is more than just accuracy that is important. Care about what you create and the impact of it.