I'm opening this issue to track the preparation for the thesis defense talk.
According to issue #25, key deadlines regarding the defense talk are:
By Tuesday, April 9 (tomorrow): prepare a rough outline of the thesis talk presentation
By Wednesday, April 17: prepare a complete draft of the thesis presentation
By Sunday, April 21: have a near-final version of the thesis presentation ready
Tuesday, April 23: practice presentation with Prof Correia
Wednesday, April 24:presentation day!!
Rough outline
For tomorrow, I wanted to start thinking about the talk and what I'd like to include and focus on. In terms of a rough outline, I'm thinking of having 6 main sections as follows:
Intro: I'd start the talk by introducing the problem of algorithmic bias, the challenges it poses, and the pressing need for a solution. This will focus more on the societal problem and might be a good place to transition into the technical manifestation of this problem by going over the standard ML optimization procedure and its shortcomings, even when blind to the demographic group.
Notions of Fairness: This next section will introduce various ways of thinking about fairness from a statistical viewpoint, inherent conflicts, and probably, in the interest of time, focus only on the separation (equality of odds) definition that was most applied in this research.
The Seldonian Framework: Building up on what we've discussed regarding the standard ML process, its shortcomings, and how we can define fairness mathematically, this next section will introduce the Seldonian framework in depth, concluding with illustrating the toy example and results from experimentation in the continuous setting.
COMPAS Application: I will then emphasize that the toy example was impractical, although it gave us a lens into what this might look like in a continuous setting. I'll then discuss the COMPAS data set itself, the disparate outcomes from the COMPAS tool and logistic regression, key results from the COMPAS data analysis (especially proxy relationships), and results from applying the Seldonian framework.
Simulation Study: Finally, I'll discuss the motivation for the simulation study based on the application, the methodology, and important results.
Conclusions: This will ultimately lead to the conclusion, pointing out 2 or 3 key take-away points and directions for future work in this area.
The talk is ~20 mins, which is not too long when you think about it. This means I likely won't be able to cover everything I've outlined here, but I could create slides based on the general outline that we agree on, retain only the most important points, and move the rest to the appendix in case questions arise?
I'd love to hear your thoughts on this, and see you tomorrow!
dashaasienga
commented
2 months ago
@katcorr
I'm opening this issue to track the preparation for the thesis defense talk.
According to issue #25, key deadlines regarding the defense talk are:
By Tuesday, April 9 (tomorrow): prepare a rough outline of the thesis talk presentation By Wednesday, April 17: prepare a complete draft of the thesis presentation By Sunday, April 21: have a near-final version of the thesis presentation ready Tuesday, April 23: practice presentation with Prof Correia Wednesday, April 24: presentation day!!
Rough outline
For tomorrow, I wanted to start thinking about the talk and what I'd like to include and focus on. In terms of a rough outline, I'm thinking of having 6 main sections as follows:
Intro: I'd start the talk by introducing the problem of algorithmic bias, the challenges it poses, and the pressing need for a solution. This will focus more on the societal problem and might be a good place to transition into the technical manifestation of this problem by going over the standard ML optimization procedure and its shortcomings, even when blind to the demographic group.
Notions of Fairness: This next section will introduce various ways of thinking about fairness from a statistical viewpoint, inherent conflicts, and probably, in the interest of time, focus only on the separation (equality of odds) definition that was most applied in this research.
The Seldonian Framework: Building up on what we've discussed regarding the standard ML process, its shortcomings, and how we can define fairness mathematically, this next section will introduce the Seldonian framework in depth, concluding with illustrating the toy example and results from experimentation in the continuous setting.
COMPAS Application: I will then emphasize that the toy example was impractical, although it gave us a lens into what this might look like in a continuous setting. I'll then discuss the COMPAS data set itself, the disparate outcomes from the COMPAS tool and logistic regression, key results from the COMPAS data analysis (especially proxy relationships), and results from applying the Seldonian framework.
Simulation Study: Finally, I'll discuss the motivation for the simulation study based on the application, the methodology, and important results.
Conclusions: This will ultimately lead to the conclusion, pointing out 2 or 3 key take-away points and directions for future work in this area.
The talk is ~20 mins, which is not too long when you think about it. This means I likely won't be able to cover everything I've outlined here, but I could create slides based on the general outline that we agree on, retain only the most important points, and move the rest to the appendix in case questions arise?
I'd love to hear your thoughts on this, and see you tomorrow!