Tag: deep-learning
All the articles with the tag "deep-learning".
-
FMMI: Estimating Mutual Information with Flow Matching
A presentation of a method that estimates Mutual Information by learning a probability flow between independence and the true joint distribution. Using Flow Matching and a divergence-based identity, MI becomes a simple expectation that we compute with a neural velocity field. A small experiment illustrates the approach and confirms its accuracy.
-
Geodesic Calculus on Latent Space
A review of Geodesic Calculus on Latent Spaces (Hartwig et al., 2025), explaining its mathematical foundations in differential geometry, demonstrating a VAE-based implementation on MNIST with a learned projection operator, and visualizing how geodesic paths yield more realistic, manifold-aware interpolations than linear ones.
-
Less is More: Recursive Reasoning with Tiny Networks
A paper showing how small neural models can achieve powerful reasoning through iterative self-correction, deep supervision, and elegant simplicity, challenging the idea that intelligence requires scale.
-
Muon Optimizer: From Conception to Modern Adaptations
Introduced in October 2024, Muon is a geometry-aware optimizer treating weight matrices as structured objects rather than flat vectors. In this article we study the evolution of Muon through recent advancements.