PyRAT

What can I learn?

A software demonstrator, developed by CEA-List, for proving the reliability of a neural network's responses within given configurations. A visually rich interface exposes the logic behind PyRAT and the mathematical operations involved, letting users experiment with different analysis parameters. Applications range from aircraft collision avoidance to medical diagnostics.

Core insight

Neural networks need not be entirely unaccountable: in constrained settings their behaviour can be formally analysed and their reliability argued for — a glimpse of what "trustworthy AI" looks like under the hood.

How to use it in daily work

A more technical resource that helps you understand, and convey, that there is serious work on verifying AI systems — useful context when discussing high-stakes uses like health or safety.

• Reference PyRAT when a client worries about AI in critical settings, to explain that reliability can sometimes be tested rather than assumed.
• Use the visual interface to see, at a conceptual level, how analysis of a network's responses can be made tangible.