Training Less, Achieving More: Unlocking Transformers with LoRA

Author(s): Saif Ali Kheraj

Originally published on Towards AI.

https://arxiv.org/pdf/2106.09685

In the era of large language models, Transformer is like the original brain of AI. But they come with a catch: Full fine tuning them is like …. Enter LoRA (Low-Rank Adaptation) — “Hey, what if we only train the parts we really need?”

Think of LoRA as adding a tiny steering wheel to a giant spaceship. You don’t need to rebuild the engine to change direction, just bolt on a little adapter. In this article, we’ll dive into the math , explain how LoRA works under the hood, and show where it fits in the Transformer architecture.

Let’s say you have a neural network layer with:

Input size: d = 10Output size: k = 8

The number of parameters in the weight matrix W0 is 10 x 8 which is 80 . That’s fine for small models. But with models like GPT or BERT, we’re talking millions of parameters — and training all of them is expensive, both in time and in your GPU’s emotional well-being.

So LoRA says: “Freeze the big guy, train a tiny plug in instead.”

Normally, a neural layer does:

Figure by Author

Now LoRA adds twist.

Figure by Author

But instead of making delta W, a full-sized matrix (which would defeat the purpose), we… Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI