Re-Parameterization and Intrinsic Dimension

Re-parameterization means: do not ask the model to learn one giant dense update directly. Instead, express that update through a smaller set of parameters.

Intuition

Imagine you need to nudge a huge weight matrix. You could learn every entry. Or you could learn a compressed version that still captures the useful change.

Research and practice both point to this: many learning problems can be solved in a much lower-dimensional subspace than the raw parameter count suggests. That smaller effective size is called the intrinsic dimension.

How it works

Two linked ideas

  1. Over-parameterized models often live on a low intrinsic dimension — lots of weights, but the useful learning direction is smaller.
  2. The change during adaptation also tends to have low intrinsic rank — so the update can be compressed too.

That is the conceptual doorway into LoRA:

Why this helps PEFT design

Idea Plain-English meaning
Intrinsic dimension The smaller space that still solves the learning problem well
Low-rank update Approximate the weight change with a thin factorization
Frozen backbone Keep pretrained knowledge; only learn the small correction

You are not claiming the whole model is tiny. You are claiming the task-specific change often is.

What goes wrong

One-line summary

Re-parameterization PEFT works because useful adaptations often live in a smaller space than the full weight matrix — so we learn that smaller update on purpose.

Key terms