Machine Learning

Why plain self-attention chokes on a million-point mesh (quadratic cost), and how Transolver's Physics-Attention fixes it: softly assign each mesh point to one of a small fixed ...

How DoMINO (NVIDIA's Decomposable Multi-scale Iterative Neural Operator) predicts aerodynamics directly from a car's STL/point cloud. It encodes the whole geometry into a global...

Why a plain 3D convnet on a car mesh is cubic (O(N^3)) and infeasible at high resolution, and how FIGConvUNet gets a global receptive field at quadratic (O(N^2)) cost: factorize...

Part I of a physics-ML monograph series on NVIDIA's simulation stack for vehicle engineering. What the PhysicsNeMo framework is, how its four repositories fit together, and a cu...

Part II of the physics-ML series. What it means for a simulator to have gradients, how NVIDIA Warp turns plain Python into GPU-speed physics code, how the Newton engine builds a...

Part III of the physics-ML series. Part I learned the physics; Part II differentiated it — this part is about everything around the physics: the PhysX engine that runs productio...

A reading map for the time-series field guide. One arc told across the series: from a single sensor sample, to a calibrated forecast of its next value, to a forecast of the fail...

Part six of the time-series field guide: structure built in, structure learned. Two complementary ways to give a time-series model an edge — physics-informed models that bake th...

Part five of the time-series field guide: when the features learn themselves. The neural architectures that learn time-series structure straight from data — recurrent networks (...

Part four of the time-series field guide: keeping watch after deployment. A model that was accurate at launch quietly rots as the data drifts, sensors break, and the world moves...

Part three of the time-series field guide. Detecting that a machine is sick is diagnostics; saying how many cycles it has left, and how sure you are, is prognostics. A plain-lan...

Part two of the time-series field guide: from the past to the next value. How to turn a history of samples into a calibrated forecast of what comes next — moving averages and ex...

Part one of the time-series field guide. Every sensor is a storyteller that can only say numbers, one at a time, forever. A plain-language, first-principles tour of how those ra...

Anomaly detection tells you something is wrong; a classifier tells you which fault it is. In vehicle health monitoring the classifiers that earn their keep are almost all built ...

Generalization and transfer are not two different topics — they are the same statement about error, separated by a single term: how different the new world is from the old one. ...

PyTorch and TensorFlow are the two frameworks that almost every modern neural network is trained on — but the differences between them, and the reasons one wins in research and ...

For roughly a decade, XGBoost was the single algorithm most likely to win a Kaggle competition that did not involve images or natural language. It is still, on tabular data, the...

Every model in machine learning, every system identification routine, every adaptive controller has an optimization problem at its core. This reference walks through the algorit...

An interactive reference on regression — predicting continuous numbers from data. Nine algorithm families compared, four live demos, and an honest treatment of the data and metr...

An illustrated essay on support vector machines and their one-class variant — the maximum-margin hyperplane, soft margins, the kernel trick, and the surprising move that turns a...

A field guide to the six paradigms of machine learning — supervised, unsupervised, reinforcement, imitation, transfer, and the rapidly-merging boundaries between them. The mecha...

A 22-chapter field guide to modern AI systems — tokenization, attention, LLMs, prompting techniques, RAG, agent frameworks, MCP, and the engineering considerations around them.

A working Bayesian optimization loop you can run in your browser. Watch Expected Improvement, UCB, and PI compete to find the minimum of a deceptive benchmark function — and rea...

Click, slide, and watch the posterior update. A working intuition for Gaussian Processes — from the one-sentence definition through the Cholesky math and the honest O(n³) scalin...

An illustrated primer on survival analysis. Why time-to-event data is fundamentally different from regression. Censoring, hazard functions, Cox's elegant trick, and the neural e...

An illustrated primer on the Isolation Forest — the anomaly-detection algorithm that flips the usual approach on its head. Five interactive figures showing isolation paths, the ...

An illustrated primer on autoencoders and variational autoencoders. The bottleneck principle, what lives in the latent space, the variational reformulation that turned an unsupe...