Not exactly. The learner could well discover that none of the symmetries in the space hold. It's an inductive bias in the precise sense of the term: it causes the learner to search for some functions over others.
@pmddomingos
-
Algorithm discovers specific symmetries from defined possibility space
By
–
We tell them the space of possible symmetries (e.g., affine); the algorithm discovers which (e.g., translation, or some translations).
-
Learning Symmetries: Discovering Constraints in Phenomenon Modeling
By
–
The prior is that there's a space of symmetries the phenomenon being modeled may obey (e.g., affine), and the learning task is to discover which (e.g., only translations, or only some translations).
-
Controlled Perturbation: Function Application as Risk Factor
By
–
But in reality you may fail when you try to do f(X), so doing it is effectively just another perturbation, albeit under your control.
-
Rung 3 Model: Conditioning on Evidence and Counterfactuals
By
–
All rung 3 requires is a model that lets you condition on arbitrary evidence, including counterfactuals. To learn the model you need inductive biases / prior assumptions, which are necessary and sufficient for all rungs and may or may not include a theory of causality.
-
Permutation Invariance: Operating on Unordered Sets
By
–
Permutation invariance lets you operate on (unordered) sets. It's not a symbol manipulation apparatus per se.
-
Algorithms for Learning Symmetries Save Data in Machine Learning
By
–
We have algorithms for learning symmetries, and using them saves a lot of data. E.g.: https://
homes.cs.washington.edu/~pedrod/papers
/nips14.pdf
…