Hello,
I’m interested in the topic of “human in the loop AI” approaches, and since I didn’t find a lot of info in the forum I thought I would start a thread. I think humans may have a role to play to fix some of AI shortcomings in production: in terms of raw accuracy improvements, but also maybe more importantly in terms of control for bias and other negative outcomes as introduced in lesson 6.
Do you have papers, articles or books on this subject to recommend? Ultimately, do you have practical suggestions to implement any of these approaches (with fast.ai / PyTorch)?
Here are my notes so far, I will be updating them:
Human client:
Human provides feedback and corrections on AI’s predictions.
[This topic is vastly covered for recommender systems as it’s their main data source, but not so much for other kinds of AI problems]
- Recommender systems
- Other classification/regression ML problems:
- [use online learning to retrain the model based on corrected data?]
- [sources needed]
- Threat robustness:
- How to introduce feedback loops while maintaining robustness, avoiding vicious circles, and protecting from third-party attacks?
- Influence Limiting and Attack Resistance; Interview with Paul Resnick (Coursera Video)
Human operator:
Human provides answers when model predictions have low confidence
- Directions in Hybrid Intelligence: Complementing AI Systems with Human Intelligence
- [sources needed]
Human auditor:
Human tests the model against errors, bias, etc:
- Towards Accountable AI: Hybrid Human-Machine Analyses for Characterizing System Failure
- [sources needed]
Thanks,