Dự án handwired biến mã JavaScript thông thường thành mạng Nơ-ron

Handwired is an open‑source TypeScript library that converts ordinary JavaScript code into a fully differentiable neural‑net representation. It was created by Cheng Lou and is hosted on GitHub, gathering 1 828 stars. The tool rewrites if‑else, loops and stateful logic into matrix‑multiplication and ReLU operations that can run on the CPU or WebGPU. Because the transformation keeps the program exact and requires no training, developers can explore gradient‑based analysis of classic JS algorithms. The project is noteworthy for bridging traditional imperative code with modern deep‑learning tooling.

Chi tiết

Traditional deep‑learning frameworks expect models to be defined as tensors that are trained from scratch, leaving ordinary imperative code outside their scope. Handwired addresses this gap by taking a JavaScript program built from if‑else statements, loops and mutable state, and translating it into a sequence of matrix multiplications followed by ReLU activations over floating‑point values. This translation preserves the exact semantics of the original code while making the whole program differentiable.

The conversion works by expressing each primitive operation as basic arithmetic, then aggregating all additions and ReLUs into a single large matrix multiplication. The resulting computation can be executed directly in JavaScript or dispatched to WebGPU for hardware acceleration. Since only two primitives—matmul and ReLU—are used, the whole program remains fully differentiable and can benefit from existing GPU optimizations such as fusion, sparsity and reordering.

Handwired unlocks several unconventional use‑cases. Users can “unfreeze” the manually written weights and train them, experiment with attaching a frozen program to a generic neural net for end‑to‑end training, or obtain GPU‑level parallelism for code that was originally single‑threaded, as demonstrated by a JSON‑search demo that processes 100 k documents in milliseconds. The repository, written in TypeScript, has attracted 1 828 stars on GitHub, indicating strong community interest in this novel approach.

Điểm nổi bật

• Unfreeze the generated weights and train them
• Attach a frozen program to a generic neural net for end‑to‑end training
• Achieve GPU‑level parallelism for traditional single‑threaded JavaScript (e.g., JSON search demo)

Vì sao đáng chú ý

Developers can analyze and optimize existing JavaScript algorithms with gradient‑based tools without rewriting them as neural networks. The ability to run the transformed code on WebGPU brings massive speedups to workloads that were previously limited to the single‑threaded JavaScript engine. Moreover, the exact, training‑free conversion opens experimental research avenues such as hybrid models that combine hand‑crafted logic with learned components.

Thông tin nhanh

• Kho mã: chenglou/handwired
• Số sao: 1.828 · Fork: 151
• Ngôn ngữ: TypeScript · Giấy phép: MIT

Nguồn: GitHub

---

Tin do AI tóm tắt từ nguồn công khai (RSS, GitHub); bấm link nguồn để đọc bài gốc.