I try to understand how finite evidence supports claims about unobserved cases through confirmation puzzles, probability spaces, Bayesian updating, chance, statistics, and learning in the limit. Not to be confused with mathematical induction.

A new general form of Kaiming initialization here which does not assume: that the weights and inputs have zero mean and that the inputs have variance $1$.

A geometric lens on simulation: cotangent bundles, Hamiltonian flows, and how geodesics, magnetism, and relativity emerge in one coherent framework.

An illustrated primer on physics-based simulation for graphics: mass-spring systems as a unifying idea, time integration (Euler, RK, backward/symplectic), and mass-spring models, and constraints.

This collection brings together 77 carefully curated scenes with multi-view sequences, camera parameters, and pre-trained Gaussian Splats—everything you need to jump into radiance field training.

This article takes a look through the chronicles of Stochastic Gradient Descent (SGD). We take a look at why SGD and the stochastic gradient noise is responsible for SGD working so well.

How do you efficiently sample from a distribution that's algorithmically #BIS-hard? The ferromagnetic Potts model is a canonical Markov random field where monochromatic edges win the popularity contest. This article is about how polymer methods and extremal graph theory crack the sampling puzzle on d-regular weakly expanding graphs.

Characterizing the sample complexity of different machine learning tasks is an important question in learning theory. This article reviews the less conventional approach of using compression schemes for proving sample complexity upper bounds, with specific applications in learning under adversarial perturbations and learning Gaussian mixture models.

What's the largest number that absolutely refuses to be written as mb+nc?

Positive examples can also help each other learn better representations. This article implements NNCLR, a self-supervised learning method for computer vision.

Implementing Swin Transformers, a general-purpose backbone for computer vision.

Gradient Centralization transforms DNN training performance by bringing gradients to zero mean, this helps with training stability and performance.