Fine-tuning (deep learning)

Why can we make a big change with just a tiny tweak?

Imagine you're trying to fit a new shelf in an old, sturdy bookcase. You don't want to dismantle the whole thing; you just need to adjust the shelf slightly to fit perfectly.

Think of the bookcase as a huge neural network, and the shelf as a small part of it that needs to fit a new purpose. Instead of changing the whole bookcase (which would be like adding lots of new parameters), you just tweak the shelf (like LoRA) so it fits perfectly without altering the rest of the bookcase.

Example

If the bookcase has 100 shelves (d) and you only need to adjust 1 shelf (r), you're making a small change (r << d) instead of altering 99 shelves.

Remember this

LoRA allows for significant adjustments with minimal changes, preserving the original structure while adapting to new needs.

Text adapted from Wikipedia, licensed under CC BY-SA 4.0.

Related concepts

LoRA (machine learning)

LoRA uses r << d for efficient adaptation

2024 in hip-hop

LoRA rank r controls model capacity and parameters

Alex Lora Cercos

Alex Lora is a Spanish film director

LAMB optimizer does: layer-wise adaptive learning rates for large batch training

LAMB optimizer adjusts learning rates layer-wise for large batch training

batch size affects generalization: larger batches find sharper minima

Larger batch sizes lead to sharper minima, enhancing generalization by providing more accurate gradient estimates

to standardize: when you need zero mean and unit variance for gradient-based optimization

Why do we need to make data uniform before training a model?

Swipe through 100 ML concepts daily

Open Pocket Polymath