TorchLean is a Lean 4 framework for neural-networks (github.com)

TorchLean is a novel framework that integrates general tensor computations, machine learning, and mathematical proofs within the Lean 4 programming environment. This framework allows users to construct and train neural networks while ensuring correctness through Lean's theorem-proving capabilities. With features like shape-checked tensors for linear algebra, automatic differentiation using LibTorch’s GPU operations, and verification tools to ensure model reliability, TorchLean addresses both development and validation in machine learning workflows. The significance of TorchLean lies in its dual ability to facilitate practical machine learning tasks while simultaneously offering formal verification of models, enhancing trust and robustness in AI applications. Users can customize numerical precision with FloatLib and adapt training scripts for various hardware configurations, including CUDA-enabled GPUs. Moreover, the framework supports a variety of tensor arithmetic formats and provides a clear API for model definition, training, and proof validation. By merging functional programming with AI development, TorchLean offers a promising direction for creating provably correct machine learning systems.