Google has introduced a new open AI model named VaultGemma, which is built on the latest insights into differential privacy.

This AI model boasts 1 billion parameters and is derived from Gemma 2, part of Google’s lineup of compact language models.

Historically, it’s been recognized that incorporating differential privacy into an AI model can stop it from spitting out results that are exactly the same as the training data.

This is achieved by adding a bit of noise during the training process.

On the flip side, introducing noise to the training data can lead to a decrease in the model’s accuracy.

A blog from Google stated that this is the largest open model focused on differential privacy available for creating high utility AI models.

Developers can download VaultGemma from Hugging Face and Kaggle.

Google has also made the weights available, allowing users to fine tune the AI model and create their own versions.

DISCLAIMER: This article is derived from information available in the public domain.

Subscribe My Channel





Discover more from Connect2ConnectOnline

Subscribe now to keep reading and get access to the full archive.

Continue reading