Grateful for all of my co-authors @alicexiang, Shubham Sharma @UTAustin, @adrian_weller @turinginst, Yunhan (Jack) Jia, @ankurtaly @fiddlerlabs, Joydeep Ghosh @CognitiveScale, @ruchir_puri @IBMResearch, @josemfmoura @CMU_ECE, and @pde33
-
-
Prikaži ovu nit
-
Also thankful for those that shared thoughts and feedback with us:
@hima_lakkaraju@krishnagade@bansalg_@terahlyons@krvarshney@rajiinio@frossi_t@PeterLoPR@saayelimukherji@a_b_powell@KarinaAlexanyan@gabizij@pjturcot@ejette@RosieCampbell@nicole_rigilloPrikaži ovu nit -
Here are some interesting findings from our interviews with over thirty organizations: 1. Feature importance explanations are the most popular type of explanations, with Shapley values being most commonly adopted (the SHAP repo by
@scottlundberg and@suinleelab helped with this)Prikaži ovu nit -
2. Though most papers motivate explanations for "end users," we find that most explanation techniques are overwhelmingly used as sanity checks for ML engineers.
Prikaži ovu nit -
3. Most organizations do not have a clear goal for why they want "explainability." Some reported using it due to directives from higher-ups.
Prikaži ovu nit -
4. There are many technical limitations for deploying certain explanation techniques at scale (e.g., influence functions for sample importance are great in theory, but are computationally expensive to deploy).
Prikaži ovu nit -
We then provide a framework for deciding what type of explainability is right for your organization and raise concerns related to explainability (e.g., privacy, lack of causality, etc.).
Prikaži ovu nit -
I'm also at
#NeurIPS2019@NeurIPSConf in Vancouver this week if you want to discuss explainability, fairness, and human-AI teams. Feel free to reach out! I'll be presenting a subset of this as a poster at the#HCML2019 workshop on Friday (Dec. 13) in West Level 2, Room 223-224Prikaži ovu nit -
Even more thrilled since this is my first paper since joining
@CambridgeMLG@LeverhulmeCFI as a doctoral student and joining@PartnershipAI as a research fellow#phdchat@AcademicChatter
Prikaži ovu nit
Kraj razgovora
Novi razgovor -
Čini se da učitavanje traje već neko vrijeme.
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.
