Exploring LoRA – Part 1: The Idea Behind Parameter Efficient Fine-Tuning

161 points
1/21/1970
4 days ago
by aquastorm

Comments


threepi

Author here. Happy to see this posted here. This is actually a series of blog posts:

1. Exploring LoRA — Part 1: The Idea Behind Parameter Efficient Fine-Tuning and LoRA: https://medium.com/inspiredbrilliance/exploring-lora-part-1-...

2. Exploring LoRA - Part 2: Analyzing LoRA through its Implementation on an MLP: https://medium.com/inspiredbrilliance/exploring-lora-part-2-...

3. Intrinsic Dimension Part 1: How Learning in Large Models Is Driven by a Few Parameters and Its Impact on Fine-Tuning https://medium.com/inspiredbrilliance/intrinsic-dimension-pa...

4. Intrinsic Dimension Part 2: Measuring the True Complexity of a Model via Random Subspace Training https://medium.com/inspiredbrilliance/intrinsic-dimension-pa...

Hope you enjoy reading the other posts too. Merry Christmas and Happy Holidays!

2 days ago

3abiton

Thanks for sharing. This got me thinking, why is medium so used for such technical articles? Especially that lots of articles get blasted behind a paywall for me recently.

2 days ago

ivanmontillam

Making it less accessible, right? I was thinking exactly the same.

19 hours ago

anshumankmr

short answer: to make money

19 hours ago

jwildeboer

(Not to be confused with LoRa, (short for long range) which is a spread spectrum modulation technique derived from chirp spread spectrum (CSS) technology, powering technologies like LoRaWAN and Meshtastic)

2 days ago

SeasonalEnnui

This gets me every time. I expect to see something interesting and it turns to be the other one. One is a fantastic thing and the other is mediocre, pick which way round at your discretion!

2 days ago

sva_

Pretty simple to spot LoRa vs LoRA.

2 days ago

rkagerer

Memory mnemonic: Capital A for "AI"

a day ago

pavlov

What exactly is the confusion? Does “parameter efficient fine-tuning” mean anything in context of the other Lora? If not, then it’s probably obvious which one this is about.

2 days ago

mrgaro

Actually it does: Lora the radio protocol has parameters to tune. Usually both sender and receiver needs to match these, so I read this like a method how these could be automatically tuned based on the distance and radio environment.

a day ago

FusspawnUK

really wish they had come up with another name. googling gets annoying

2 days ago

the__alchemist

Contributors: They both use mixed capitalization. They have partially-overlapping audiences.

2 days ago

danielhanchen

Super cool series of articles! :)

2 days ago

JacksonWaschura

[dead]

2 days ago

gautambt

2 days ago

khazhoux

What is this? Is this a google summarization service?

a day ago