inductive biases (TEST 1)

Note: this is not a real post (site under construction), but the content comes from my notes nonetheless

A set of assumptions that allows a learning algorithm to learn about structure in data, and as a result generalise well from limited data.

For instance: spatial consistency in the context of learning about objects from images.

From a Bayesian standpoint, the set of prior beliefs about a certain domain or environment on which inference depends.

Inductive biases can be found and/or designed at different levels of abstraction, involving both the algorithmic architecture realizing certain computations and the type of learning involved. For the architecture, we have:

  1. inductive biases at the micro level (micro-architecture, small-scale structures that have computational consequences, cf. parameter sharing):
    • sparsity
    • normalization
    • attention
  2. inductive biases at the macro level (macro-architecture)
  3. inductive biases at the cognitive level