Part 2, the formal-methods companion
Part 2 gives a method for applying formal techniques when the robustness of a neural network is judged. It defines stability as how far a network's output stays the same when inputs change, and sensitivity as how far the output moves.
The requirement in plain English
Part 2 gives a method for applying formal techniques when the robustness of a neural network is judged. It defines stability as how far a network's output stays the same when inputs change, and sensitivity as how far the output moves.
What the public record says
The public view gives the scope, the terms and the introduction of Part 2. The introduction says statistical analysis can show that a system keeps its performance when conditions vary, and that formal techniques are illustrated in Part 2.
Why it matters to you
Part 2 is the contrast that defines Part 3. Formal methods prove a property over a defined input region; statistical methods sample and estimate. The strengths are different: a proof is stronger where it applies, and a sample reaches further where a proof is impractical. The weakness is also different: a proof can be precise about the wrong region, and a sample can be representative of the wrong distribution. Part 2's stability term, at the network level, is the one the FDIS moved to the system level at 3.1, so the two documents must be read together.
Evidence a reviewer asks for
- The method selection note that says which claims rest on formal methods and which on statistical methods.
- A record of the input region a formal method covers, and the sampling frame a statistical method uses.
- The stability definition used in each test, with its source.
Checks to run this month
- Does the evidence pack distinguish a proof from a sample?
- Is the input region of any formal result stated?
- Has anyone reused the Part 2 stability wording for an FDIS claim?
Dates
Part 2 published in 2023. FDIS registered 2026-08-21.
Cross-references
- Clause 3.1, Stability
- Formal methods and statistical methods are complements
- A statistical approach to robustness
What this page does not state
Part 2's method chapters are not in the public view.