JavaScript Library Runs Machine Learning Models in Browser (thenewstack.io)

Julian Wilkison-Duran unveiled AsterMind‑ELM, an open‑source JavaScript/TypeScript library that brings Extreme Learning Machines (ELMs) to the browser — not as a Python wrapper but rewritten natively for the frontend. ELMs fix a randomly initialized hidden layer and only solve the output weights (typically via the Moore‑Penrose pseudoinverse), which makes training extremely fast and memory‑light. AsterMind‑ELM includes kernel ELMs, Online ELM, DeepELM, and Web Worker offloading so developers can train tiny models in milliseconds, predict with microsecond latency, chain models together, and keep everything on‑device (models are often just JSON files). This matters because it offers a practical path to private, low‑latency, on‑page ML without GPUs or backends — ideal for intent/language classifiers, edge analytics, real‑time controls, compact retrieval/reranking, and interactive tools (Wilkison‑Duran demoed text classification from 20 sentences, voice commands, and a drum generator). The tradeoff is accuracy vs. scale: ELMs prioritize speed and efficiency over the massive parameter counts of LLMs and aren’t meant to replace them. The repo encourages community contributions and shared weights, positioning AsterMind‑ELM as a lightweight, transparent alternative for frontend ML use cases where instant training, privacy, and low resource use are primary constraints.