numerical-stability

Leveraging network motifs to improve artificial neural networks

➗ This work presents a thorough analysis of a quadratic optimization model, confirming convexity, coercivity, and the exact global minimum. It compares gradient descent and Newton’s method, highlighting Newton’s superior efficiency when the Hessian is invertible.

Validate PyTorch model outputs across NVIDIA, AMD, Trainium, TPU, and Apple Silicon. Find numerical divergence before it reaches production.

Research project analyzing stability and robustness of deep learning optimizers (SGD, Adam, SAM) under label noise and precision constraints.

High-performance, branchless numerical stability kernels and compiler-optimized core infrastructure for advanced JAX/XLA deep learning architectures.

Detect numerical instability in ML applications using learned invariants (Soft Assertions) without modifying model logic. ACM FSE 2025.

量測 LLM 推論的決定性：同一 prompt 重複執行時，輸出從第幾個 token 開始分歧。

Pre-registered experiment on differentiable EML-tree symbolic regression (Odrzywolek 2026, arXiv:2603.21852). Streamlit UI + headless reproducer + full results.

Makine Öğrenmesi için Lineer Cebir ve Matris Hesaplamaları masterclass'ı. Hazır kütüphaneler yok; saf geometrik sezgi ve NumPy var. Sıfırdan Gauss eleme, LU ve QR ayrıştırmaları, kararlı en küçük kareler, özdeğer iterasyonları ve doruk noktası olarak PCA ile uçtan uca SVD tabanlı boyut indirgeme mimarisi.

A hands‑on, first‑principles guide to fitting logistic regression via the Iteratively Reweighted Least Squares (IRLS) algorithm complete with mathematical derivations, R code from scratch, and a real‑world S&P data case study to bring your statistical modeling skills to the next level.

Multi-Engine (PyTorch & JAX/XLA) Zero-Branching Geometric Acceleration Core. Enforces 0% Graph Breaks & Real-time Fault-Isolation via hardware-native bitwise MUX operations (torch.where / jax.lax.select) to permanently eliminate 'jmp' instructions and host-device synchronization fences.

A research-grade course in Numerical Methods and Optimization for ML. No deep learning libraries—just pure math and NumPy. Rebuilds automatic differentiation engines, stable SVD/Cholesky algorithms, quasi-Newton solvers (BFGS), stable log-sum-exp layers, and Gaussian Process regressions from scratch

R package and replication code for the article “Numerical stability enhancements in beta autoregressive moving average model estimation” by Cribari-Neto, F., Costa, E., and Fonseca, R. V., published in the Brazilian Journal of Probability and Statistics (2025). DOI: 10.1214/25-BJPS645.

A lightweight C++ tool that prices European call and put options using the Black–Scholes formula, computes all key Greeks (Δ, Γ, Θ, Vega, Rho), and lets you run quick ATM/ITM/OTM scenario checks—all via a simple command‑line interface.

Incremental corpus centroid with exact add/remove: compensated float64 accumulation and drift stats that separate rotation from diversity

This MATLAB function efficiently computes the inverse of a square matrix using LU factorization. By decomposing the matrix into lower and upper triangular matrices, the function solves for the inverse with improved numerical stability.

StableStockPredictor is a robust deep learning model for predicting S&P 500 stock prices, built with TensorFlow and Keras. It leverages LSTM networks with gradient clipping, robust scaling, and stable feature engineering (e.g., RSI, moving averages, volatility) to ensure reliable performance in volatile markets.

Offline diagnostic engine and Snorkel Terminal-Bench benchmark task for scoring Apptainer groundwater simulation stability risks.

Stable implementations of logaddexp and logsumexp in rust

Compute the arithmetic mean along one or more ndarray dimensions using a two-pass error correction algorithm.