How to Constrain Artificial Stupidity
summary
Claims that adding constraints cuts “artificial stupidity”: predict less (add a do not predict threshold and outlier check via a Gaussian mixture model), remove or orthogonalize sensitive features (reduces bias though accuracy may drop), enforce monotonicity or fairness constraints (isotonic regression, TensorFlow Lattice/XGBoost flags) and model structure (PyMC3 hierarchical chicken example gives interpretable slopes and variances). Examples: PayPal chatbot failures, Oklahoma healthcare and an Irish case where being born on Sunday adds about $40; an open source toolkit called psychic Lego ships these transformers.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:07AI’s blind authority
- 3:55Predicting less
- 8:09Interpolation vs extrapolation
- 12:42Preventing bias
- 19:00Fairness metrics
- 23:54Adding constraints to machine learning models
- 31:17Tailored models
- 34:06Avoiding dead chickens
- 36:11Q&A
related talks
26mArtificial Intelligence? - More like Artificial Stupidity!
44mToxic Trust: The Psychology of Automation Bias in Generative AI
23mOpening up the box: the complexity of human interaction with AI systems
39mBack to Basics: Cutting Through the AI-Generated Code Noise
06m7 practical tips for safer AI and language models
57mToxic Trust: The Psychology of Automation Bias in Generative AI
46mCan you trust your AI?
44mAI Engineering Anti-patterns