TorchLean: Verified Neural Networks in Lean (www.robertj1.com)

TorchLean, a project developed over the past year, aims to address the challenges of formal verification in neural networks by incorporating them into the Lean theorem prover. This will help ensure that the complex journey a model takes—from Python code to deployment—does not misrepresent the intended computation. The initiative is critical as AI programming increasingly involves generating code that requires rigorous checks on computation integrity across various stages, particularly in scientific machine learning applications. This shift towards "vericoding" emphasizes the need for AI systems to not only produce code but also create artefacts that are checkable against formal specifications. TorchLean comprises three integral components: an ML programming layer in Lean that facilitates type-safe tensor operations, a specification layer that defines models and their computations, and a verification layer that includes various checks and proofs for the models. By formally representing the neural network's computations within Lean, TorchLean allows for clear documentation of how different aspects of a model interact, reducing the likelihood of drift—where the operational behavior diverges from the theoretical claims. This approach is particularly pivotal in scientific computing, where precision and correct execution are vital for supporting theoretical claims in research.