One of the main advantages of Vehicle compared to a testing framework is that it can be used to state and prove specifications that describe the network’s behaviour over an infinite set of values.
Suppose you have the following network which produces two outputs:
@network f : Tensor Rat [10, 10] -> Vector Rat 2
and would like to specify that for any input the network’s first
output is always positive.
This can be achieved by using the
forall quantifier as follows:
forall x . f x ! 0 > 0
which brings a new variable
x of type
Tensor Rat [10, 10] into
scope. The variable
x has no assigned value and therefore represents
an arbitrary input.
Similarly, if trying to specify that there exists at least one input for which
the network’s first output is positive, the
exists quantifier can be
used as follows:
exists x . f x ! 0 > 0
As with lambda functions, the quantified variables can be annotated with their types:
exists (x : Tensor Rat [10, 10]) . f x ! 0 > 0
and multiple variables can be quantified over at once:
exists x i . f x ! i > 0
In many cases you don’t want the property to hold over all the
values in the set, but only a (still infinite) subset of them.
For example, network inputs are frequently normalised to lie
within the range
[0, 1]. If the quantified variable’s domain is not
also restricted to this range, then Vehicle will produce spurious
counter-examples to the specification.
In general such restrictions can be achieved by combining a quantifier with an implication as follows:
forall x. (forall i j . 0 <= x ! i ! j <= 1) => f x ! 0 > 0
While most specifications will quantify over at least one variable with an infinite domain, sometimes one might also want to quantify over a finite set of values. There are multiple ways of doing this:
Quantifying over an
The first approach is to quantify over a variable with the
type. For example:
pointwiseLess : Vector Rat 3 -> Vector Rat 3 -> Bool pointwiseLess x y = forall (i : Index 3) . x ! i < y ! i
will get automatically expanded to:
pointwiseLess : Vector Rat 3 -> Vector Rat 3 -> Bool pointwiseLess x y = x ! 0 < y ! 0 and x ! 1 < y ! 1 and x ! 2 < y ! 2
The type annotation
Index 3 on the quantified variable
included for clarity but are not need in practice as it can be inferred
by the compiler.
Alternatively quantifiers can be modified with the
in keyword to
quantify over all the values contained within a
myList : List Rat myList = [0.4, 1.1, 0.2] myListInRange : Bool myListInRange = forall x in myList . 0 <= f x <= 1
During compilation Vehicle will automatically expand this out to a sequence of conjunctions as follows:
myListInRange : Bool myListInRange = 0 <= f 0.4 <= 1 and 0 <= f 1.1 <= 1 and 0 <= f 0.2 <= 1
One hard constraint enforced by both training and
verification tools is that you may not use both a
exists that quantify over infinite domains within the same property.
For example, the following is not allowed:
@network f : Vector Rat 2 -> Vector Rat 1 @property surjective : Bool surjective = forall y . exists x. f x == y
This remains true even if you move one or more of the quantifiers to separate functions. For example, the following is not allowed either:
@network f : Vector Rat 2 -> Rat hits : Vector Rat 2 -> Bool hits y = exists x . f x == y @property surjective : Bool surjective = forall y . hits y
However, you can have both types of quantifiers within the same specification as long as they belong to different properties. For example, the following is allowed:
@network f : Vector Rat 2 -> Rat @property prop1 : Bool prop1 y = exists x . f x >= 2 @property prop2 : Bool prop2 = forall x . 1 <= f x <= 3