|
Manuel Gomez-Rodriguez
@
autreche
Kaiserslautern
|
|
Human-centric machine learning at the Max Planck Institute for Software Systems.
|
|
|
200
Tweetovi
|
101
Pratim
|
411
Osobe koje vas prate
|
| Tweetovi |
|
Manuel Gomez-Rodriguez
@autreche
|
27. sij |
|
If you are at #amld at EPFL, don't miss the AI & Education track appliedmldays.org/tracks/ai-educ…. Thrilled to join a great set of speakers to talk about the Memorize algorithm (pnas.org/content/116/10…) and our on-going large-scale interventional experiment on personalized learning.
|
||
|
|
||
|
Manuel Gomez-Rodriguez
@autreche
|
24. sij |
|
My take on "Human-Centric Machine Learning: Feedback loops, Human-AI Collaboration and Strategic Behavior": people.mpi-sws.org/~manuelgr/manu…. A talk/lecture on some of my recent work. Tons of opportunities for future work!
|
||
|
|
||
|
Manuel Gomez-Rodriguez
@autreche
|
24. sij |
|
Finally, we experiment with synthetic and real lending data to both illustrate our theoretical findings and show that, under strategic behavior, the policies our algorithm find do much better than deterministic threshold rules (n/n)
|
||
|
|
||
|
Manuel Gomez-Rodriguez
@autreche
|
24. sij |
|
Moreover, under no assumptions on the cost individuals pay to change their features, we develop an iterative search algorithm that is guaranteed to find locally optimal decision policies also in polynomial time (5/n)
|
||
|
|
||
|
Manuel Gomez-Rodriguez
@autreche
|
24. sij |
|
But, if the cost individuals pay to change their features satisfies a monotonicity assumption, we can narrow down the search for the optimal policy to a family of decision policies with desirable properties. This allows for a polynomial time heuristic search algorithm (4/n)
|
||
|
|
||
|
Manuel Gomez-Rodriguez
@autreche
|
24. sij |
|
Then, we first show that, in general, we cannot expect to find optimal decision policies in polynomial time under strategic behavior and there are cases in which deterministic policies are suboptimal (3/n)
|
||
|
|
||
|
Manuel Gomez-Rodriguez
@autreche
|
24. sij |
|
Individuals may use knowledge, gained by transparency, to invest effort strategically in order to receive a beneficial decision. In our work, we show that this strategic investment of effort by individuals can be cast as an optimal transport problem at a population level (2/n)
|
||
|
|
||
|
Manuel Gomez-Rodriguez
@autreche
|
24. sij |
|
Happy to share our (updated) work on optimal decision making under strategic behavior arxiv.org/abs/1905.09239, which lies in the emerging (super exciting) field of strategic machine learning. This work aims to find decision policies that lead individuals to self-improvement (1/n)
|
||
|
|
||
|
Manuel Gomez-Rodriguez
@autreche
|
24. sij |
|
We just posted in arxiv the new version of our work: arxiv.org/abs/1905.09239, where we significantly expand our results both in terms of theory and algorithms. In particular, we have a new section (and appendix) focusing on costs that satisfy a natural outcome monotonic property.
|
||
|
|
||
|
Manuel Gomez-Rodriguez
@autreche
|
23. sij |
|
sciencedaily.com/releases/2018/… ;-) Since a few weeks you just keep insisting on undermining the importance of climate change when highlighting another of the world’s problems. If climate change is not your fight, it is fine but, in this particular issue, climate does matter!
|
||
|
|
||
|
Manuel Gomez-Rodriguez
@autreche
|
21. sij |
|
I agree with Gergely here, it is a piece of crap and not sure why many ppl started using it.
|
||
|
|
||
| Manuel Gomez-Rodriguez proslijedio/la je tweet | ||
|
Niki Kilbertus
@k__niki
|
9. sij |
|
There's an important difference between decisions and predictions, especially when it comes to fairness.
Check out our #aistats2020 paper "Fair Decisions Despite Imperfect Predictions" with @autreche @krikamol
@bschoelkopf Isabel Valera
arxiv.org/abs/1902.02979
|
||
|
|
||
|
Manuel Gomez-Rodriguez
@autreche
|
7. sij |
|
Congrats Mark!!
|
||
|
|
||
|
Manuel Gomez-Rodriguez
@autreche
|
8. pro |
|
Pretty surprising results, thx for sharing! You may like to check out arxiv.org/abs/1902.02979, where we explore the idea of decision policies/data gathering under selective labels and the suboptimality of decision threshold rules in that context.
|
||
|
|
||
|
Manuel Gomez-Rodriguez
@autreche
|
7. pro |
|
Thoughtful, nuanced, and well written article on algorithmic bias vs human bias: nytimes.com/2019/12/06/bus…
|
||
|
|
||
|
Manuel Gomez-Rodriguez
@autreche
|
28. stu |
|
If you go to #NeurIPS2019, consider attending our workshop on human-centric ML sites.google.com/view/hcml-2019, we have a terrific set of invited speakers (@kgummadi, Deirdre Mulligan, @Aaroth, Finale Doshi-Velez, @_beenkim), contributed papers and panels!
|
||
|
|
||
|
Manuel Gomez-Rodriguez
@autreche
|
20. stu |
|
I agree with you that many startups that claim to use AI to predict social outcomes may oversell. In that context, I believe it is important to raise awareness and be skeptical. However, judging scientific progress on the topic based on what startups claim seems unfair n/n
|
||
|
|
||
|
Manuel Gomez-Rodriguez
@autreche
|
20. stu |
|
Moreover, there is a lot of peer-reviewed (and on-going) work on the influence that algorithmic predictions may have on those experts, which tweet threads like yous misrepresent greatly, e.g., papers.ssrn.com/sol3/papers.cf…, people.mpi-sws.org/~nghlaca/paper… (5/n)
|
||
|
|
||
|
Manuel Gomez-Rodriguez
@autreche
|
20. stu |
|
The use of machine learning predictive models does not necessarily means replacing an expert but informing one who will take the final decision. I agree it is super important to develop mechanisms (e.g., explanations) to help that expert contextualize/weigh such predictions (4/n)
|
||
|
|
||
|
Manuel Gomez-Rodriguez
@autreche
|
20. stu |
|
You also argue that no social outcome can be predicted because no one can predict the future. You say that is common sense. Appealing to common sense is not a scientific argument, this is specially surprising coming from a Princeton professor (3/n)
|
||
|
|
||