✍️ Worth to read papers

1 minute read

Published:

Compilation of some papers that I enjoyed reading.

[2504.16929] I‑Con: A Unifying Framework for Representation Learning

Proposes a single equation aiming to unify diverse machine learning approaches: clustering, contrastive learning, supervised learning, and dimensionality reduction. The method is fundamentally based on the idea that each ML model is optimizing a loss function based on the KL divergence between two data distributions, p and q. Where p represents the probability density of the true data and q the learned model distribution.

“Universal” loss function:

\[\mathcal{L}(\theta, \phi) = \int_{i, j \in \mathcal{X}} p_\theta(j \mid i) \log \frac{p_\theta(j \mid i)}{q_\phi(j \mid i)}\]

This framework aspires to act as a “periodic table of machine learning”, suggesting that the current gaps within this formulation may inspire new model types or architectural innovations.

Periodic Table of Machine Learning

[2006.11239] Denoising Diffusion Probabilistic Models

This is a classic paper in generative models, introduces diffusion models (well, there was a previous paper in 2015, 1503.03585). Diffusion models are trained for removing Gaussian noise, the main idea is a model that gradually transforms a Gaussian noise image into an actual image.

The paper has a strong mathematical background, I have a blog entry with notes of diffusion models.

Example of diffusion Example of diffusion, over 500 steps the gaussian noise image becomes to an image