Also:
1. The speaker in your paper doesn't know the NFL.
2. We can do even weak-knowledge ML a lot faster than evolution.
3. There's very often a tradeoff between accuracy and transparency/explainability, and often accuracy is more important.
@pmddomingos
-
NFL Knowledge Gap, ML Speed, and Accuracy-Explainability Trade-offs
By
–
-
Tabula-Rasa Model Oxymoron: NFL Theorem and Fundamental Assumptions
By
–
"Tabula-rasa model" is an oxymoron, by the NFL theorem. The question is what are the fundamental assumptions that need to be built in. Causality is a candidate. The laws of physics are another.
-
Climate Models Replaced by Superior Alternative Method
By
–
Who needs climate models now? This does the job much better.
-
Learning Actions from Data Rather Than Building Them In
By
–
We don't need to build it in, because it's very easy to learn from data (cf. infants). And we arguably shouldn't, because what your actions are can change (e.g., moving your hands vs. driving your car).
-
AI tutors offering personalized online classes with real-time video
By
–
Coming soon: online classes with one-on-one, real-time video interactions with AI tutors.
-
Physics Laws: Brain-Dependent Perspectives on the Universe
By
–
The laws of physics reduce the universe to something our brains can deal with. For different brains (e.g., AIs) they would probably look very different.
-
Classical Mechanics as MAP Approximation of Quantum Mechanics
By
–
Classical mechanics is the MAP approximation of quantum mechanics.
-
ML Research on Function Fitting and Inductive Logic Programming
By
–
There's a lot of ML research on what functions need fitting. In particular, the field of inductive logic programming – learning logical theories – should be relevant. Also, to Konrad's point, I think many people working on causal learning see connections with "mainstream" ML.
-
Causal Theories as Inductive Biases: Benefits and Questions
By
–
Of course, but the question is: how have causal theories to date benefited from being viewed as inductive biases?