These are quick notes I created based on what the tech and media industries are showing more interest in this week.

Fine-Tuning refers to optimizing all of a model’s parameters. Also see LLM fine-tuning and LoRA.

GAN is a generative adversarial network that I wrote about in 2022 on Medium “Diffusion compared to GAN“

Hugging Face provides datasets, trained data models and resources for Generative AI work, including optimized LoRA models.

Kaggle is a subsidiary of Google. It’s a machine learning and data science community that provides data science tools, resources and datasets.

LLM refers to large language models.

LoRA refers to Low-rank Adaptation as a way to adapt a large machine learning model for specific uses without retraining the entire model. LoRA modifies a pretrained model (LLM or Vision Transformer) to produce more relevant outputs for a specific, usually smaller dataset by adjusting a small, low-rank subset of the model’s parameters.

LoRA is different from Fine-Tuning as it trains only low-rank perturbations to selected weight matrices while Fine-Tuning optimizes all model parameters.

LoRA adapts large language learning models for specific uses without retraining the entire model.

LoRA can be applied to model architectures that perform matrix multiplication operations such as Transformers for NLP tasks where the choice of placement can significantly change the training outcomes.

LoRA can be trained for fine-tuning Stable Diffusion

RAG refers to Retrieval-Augmented Generation. It’s an AI framework that combines the strength of traditional informative retrieval systems (databases) with the features of generative large language models (LLMs). RAG is the process of optimizing the output of an LLM so it will reference an authoritative knowledge base outside it’s own training data sources before generating a response. It’s usually used for generating responses to certain questions. The RAG system does create hallucinations with ChatGPT.

Stable Diffusion is a generative AI model that produces images from text input and image inputs. It is often customized or fine-tuned with LoRA using web-ui and comfyui interfaces.

REFERENCES

“Harnessing the Power of LoRA in Large Language Models: A Deep Dive into the Future of AI – Deeper Insights.” Deeperinsights.com, deeperinsights.com/ai-blog/harnessing-the-power-of-lora-in-large-language-models-a-deep-dive-into-the-future-of-ai. Accessed 1 Aug. 2024.

Lorica, Ben. “Customizing LLMs: When to Choose LoRA or Full Fine-Tuning.” Gradient Flow, 20 May 2024, gradientflow.com/lora-or-full-fine-tuning/. Accessed 1 Aug. 2024.

Merritt, Rick. “What Is Retrieval-Augmented Generation?” NVIDIA Blog, 15 Nov. 2023, blogs.nvidia.com/blog/what-is-retrieval-augmented-generation/.

Rodriguez, Jesus. “Hugging Face’s LoRA Is a Simple Framework for Fine-Tuning Text-To-Image Models.” Medium, 20 Feb. 2023, pub.towardsai.net/hugging-face-lora-is-a-simple-framework-for-fine-tuning-text-to-image-models-746e7615420.