A hitchhiker's guide to embeddings

Embeddings are vectors of floating points representing, in a high-dimensional latent space, the information contained in a text, an image or an audio.

These vectors carry a lot of information about the original data, they serve as crucial tools for machine learning models to make predictions, find similar objects and perform semantic search.

Join me in this talk where we will discover how embeddings can be generated using open-source models and what we can do with them.

Is your model private?

As the popularity of Machine Learning models continues to soar, concerns about the risks associated with black box models have become more prominent. While much attention has been given to the development of unfair models that may discriminate against certain minorities, there exists another concern often overlooked: the privacy risks posed by ML models.

Research has substantiated that ML models are susceptible to various attacks, with one notable example being the Membership Inference attack, enabling the prediction of whether a specific sample was present during training.

Join me in this talk, where I will explain the privacy risks inherent in Machine Learning models. Beyond exploring potential attacks, I will explain how techniques such as Differential Privacy and tools like Opacus (https://github.com/pytorch/opacus) can play crucial roles in training more robust and secure models.

Preferred session duration: 30-45 minutes

Please, stop logging your experiments on a .txt file

Effective debugging Machine Learning models and comparison results demand a robust logging system. In this talk, we will explore the capabilities of Weights & Biases, a platform that offers a solution for tracking results. Learn how this tool can enhance your productivity during ML model development