Surely showing you can learn (what looks like) causality does not require the training and test examples to be in different domains. That's an orthogonal issue.
@pmddomingos
-
Machine Learning Model for Soldier Casualty Prediction from Shot Data
By
–
Easy. The data is vectors of "shot/not shot" (one input variable per soldier) and "dead/alive" (output). Any learner that induces that num. shots > 0 => dead answers the question correctly.
-
Inductive Biases in Machine Learning: Beyond Curve-Fitting
By
–
1. Many inductive biases aren’t curve-fitting (e.g., ILP, causal theories).
2. You can’t learn at any level without inductive biases.
3. We also call everything that doesn’t work “inductive biases”. ML is the search for the ones that do, in any form. -
Machine Learning Progress with Minimal Data and Inductive Biases
By
–
Indeed, but the interesting ML question is how far you can get with how little. Find the right inductive biases, and you could get very far very quickly.
-
Infinite Data and Finite Representations of True Distributions
By
–
“Infinite data” means knowing the true distribution, which may have a finite representation. Translating from an infinite representation to an equivalent finite one is not inductive bias.
-
Pooling and Parameter Sharing: Essential for Translation Invariance in CNNs
By
–
Pooling is a standard part of convnet architecture. Translation invariance is also encoded in parameter sharing: it's the *same* filter you're running across the image (or part of it). Either alone does not do the job.
-
Finite Models and Renormalizability in Quantum Field Theory
By
–
Finite models are always renormalizable. Continuous ones can be easier to handle analytically, but we don't need that today. We have computers. And with guaranteed renormalizability quantum field theory and quantum gravity become easy. What are we waiting for?
-
Universe Turing-Completeness as Fundamental Physical Law
By
–
The first law of physics is that the universe is Turing-complete. Any candidate theory that violates it is empirically false.
-
Human Intelligence as a Single Point in the Intelligence Space
By
–
Human intelligence is just a point in the space of intelligences.