Sitemap

Today we’re going to implement the aforementioned paper in code. Reason 1: They did not release code. Reason 2: I am using their method in my research so I had to code it up. Reason 2: Save other people’s time? Reason 3: Noooo codddieeeee.

It is not difficult to write up a kernel \(k_{\theta}(x,y)\) in Jax and there are many ways to do so. Here I’ll offer my method, which is to just maintain a class module with self-contained trainable parameters, just as one might write a custom neural network layer. Therefore, we don’t have to carry around parameters elsewhere, and if we want to keep them fixed, we can.

Short description of portfolio item number 1

Short description of portfolio item number 2

This thesis investigated novel methods for modeling irregular & multivariate time series data of Active Galactic Nuclei (AGN), for the sake of improving both existing techniques for reverberation mapping and anomaly detection.

MLSS^N 2022

The 2nd AMS-UMI International Joint Meeting; SLIDES

A conditional continuous flow for a quantity-of-interest posterior associated with the viscous Burgers equation.

Conditional coupling layers and conditional base densities for learning Gaussian inverse-problem posteriors.

Continuous-time density transformation with Diffrax, Equinox vector fields, and the instantaneous change-of-variables formula.

Simulation-free vector-field training along conditional probability paths, implemented with JAX, Equinox, Optax, and Diffrax.

RealNVP coupling layers, change of variables, maximum likelihood training, and sampling in JAX and Equinox.