I am trying to get a better understanding of inductive biases used in machine learning.
I understand inductive biases as the assumptions encoded into the learning algorithm which make it prefer one hypothesis over another.
Architectural choices are prime examples of inductive biases. The simplest example, linear regression, assumes that there is a linear relationship between input and output.
My question is whether the choice of representation forces the algorithm to prefer one hypothesis over another.
Simple example
Suppose that our problem is to predict house prices. We can select between 2 different representations:
- House is represented as: $(x_1, x_2) \gets $ (area of house, # bedrooms)
- House is represented as: $(x_1', x_2') \gets $ (distance from city center, # swimming pools)
Does the choice of representation introduce any additional inductive bias besides the one introduced by linear regression? I mean in both cases, we are left with mapping from $\mathbb{R}^2 \to \mathbb{R}$ with a linear function.
My view (possibly wrong)
The only way I can think of representations introducing an inductive bias is if we view the problem as following. There is an input space $X$ (the raw input) and we are interested in finding a map $X \to Y$. If we use a learner with $(x_1, x_2)$, then this amounts to finding a function:
$X \to g(X) \to h(g(X)) \to Y$
where $(x_1, x_2) = g(X)$. Now we have an inductive bias, since the function the learner must pick is the total composition $h(g(X))$ and the constraint arises since it must include $g(X)$.