diff --git a/documentation/AI-Verification-Convexity.md b/documentation/AI-Verification-Convexity.md index adc67bd..bd58099 100644 --- a/documentation/AI-Verification-Convexity.md +++ b/documentation/AI-Verification-Convexity.md @@ -147,7 +147,7 @@ activation functions evolving the state, such as $tanh$  activation layers (see layer “nca\_0” in Fig 4). As with FICNNs, the weights in certain parts of the network are constrained to be non-negative to maintain the partial convexity property. In the figure above, the weight matrices for -the fully connected layer “fc\_z\_+\_1” is constrained to be positive +the fully connected layer “fc\_z\_+\_1” are constrained to be positive (as indicated by the “\_+\_” in the layer name). All other fully connected weight matrices in Fig 4 are unconstrained, giving freedom to fit any purely feedforward network – see proposition 2 [1]. Note again that in our implementation, the final activation function, $g_k$, is not applied. This still guarantees partial convexity but removes the restriction that outputs of the network must be non-negative. @@ -180,7 +180,7 @@ $f((1−\lambda)x+\lambda y) \leq (1−\lambda)f(x)+ \lambda f(y)$. Interval $$ f(x) \leq max(f(a),f(b)) $$ -To find the minimum of $f$ on the interval, you could use a optimization routine, such as projected gradient descent, interior-point +To find the minimum of $f$ on the interval, you could use an optimization routine, such as projected gradient descent, interior-point methods, or barrier methods. However, you can use the properties of convex functions to accelerate the search in certain scenarios. @@ -190,7 +190,7 @@ If $f(a) \gt f(b)$, then either the minimum is at $x=b$ or the minimum lies strictly in the interior of the interval, $x \in (a,b)$. To assess whether the minimum is at $x=b$, look at the derivative, $\nabla f(x)$, at the interval bounds. If $f$ is not differentiable at the interval bounds, for example the network has relu activation -functions that defines a set of non-differentiable points in $\mathbb{R}$, evaluate +functions that define a set of non-differentiable points in $\mathbb{R}$, evaluate both the left and right derivate of $f$ at the interval bounds instead. Then examine the sign of the directional derivatives at the interval bounds, directed to the interior of the interval: $sgn( \nabla f(a), -\nabla f(b) ) = (\pm , \pm)$. Note that the sign of 0 is taken as positive in this discussion. @@ -240,7 +240,7 @@ possible sign combinations since, at $x=b$, convexity means that $-\nabla f(b+\e In the case that $f(a) = f(b)$, the function must either be constant and the minimum is $f(a) = f(b)$. Or the minimum again -lies at the interior. If $sgn(\nabla f(a)) = +$, then $\nabla f(a) = 0$ else this violates convexity since $f(a) = f(b)$. Similar is true for +lies in the interior. If $sgn(\nabla f(a)) = +$, then $\nabla f(a) = 0$ else this violates convexity since $f(a) = f(b)$. Similar is true for $-sgn(\nabla f(b)) = +$. In this case, all sign combinations are possible owing to possible non-differentiability of $f$ at the interval bounds: @@ -262,8 +262,8 @@ convex functions. This idea can be extended to many intervals. Take a 1-dimensional ICNN. Consider subdividing the operational design domain into a union of intervals $I_i$, where $I_i = [a_i,a_{i+1}]$ and $a_i \lt a_{i+1}$. A tight lower and upper bound on each interval can be computed with a -single forward pass through the network of all interval bounds values in the union of intervals, a -single backward pass through the network to compute derivatives at the interval bounds values, and +single forward pass through the network of all interval boundary values in the union of intervals, a +single backward pass through the network to compute derivatives at the interval boundary values, and one final convex optimization on the interval containing the global minimum. Furthermore, since bounds are computed at forward and backward passes through the network, you can compute a 'boundedness metric' during @@ -279,29 +279,28 @@ and $sgn(0) = +$. The previous discussion focused on 1-dimensional convex functions, however, this idea extends to n-dimensional convex functions, $f:\mathbb{R}^n \rightarrow \mathbb{R}$. Note that a vector valued convex function is convex in each output, so it is sufficient to keep the target as $\mathbb{R}$. In the discussion in this section, take the convex set to be the n-dimensinal hypercube, $H_n$, with vertices, $V_n = {(\pm 1,\pm 1, \dots,\pm 1)}$. General convex hulls will be discussed later. -An important property of convex functions in n-dimensions is that every 1-dimension restriction also defines a convex function. This is easily seen from the +An important property of convex functions in n-dimensions is that every 1-dimensional restriction also defines a convex function. This is easily seen from the definition. Define $g:\mathbb{R} \rightarrow \mathbb{R}$ as $g(t) = f(t\hat{n}) \text{ where } \hat{n}$ is some unit vector in $\mathbb{R}^n$. Then, by definition of convexity of $f$, letting $x = t\hat{n}$ and $y = t'\hat{n}$, it follows that, $$ g((1−\lambda)t+\lambda t') \leq (1−\lambda)g(t)+ \lambda g(t') $$ -Note that the restriction to 1-dimensional convex function will be used several times in the following discussion. +Note that the restriction to 1-dimensional convex functions will be used several times in the following discussion. To determine an upper bound of $f$ on the hypercube, note that any point in $H_n$ can be expressed as a convex combination of its vertices, i.e., for $z \in H_n$, it follows that $z = \sum_i \lambda_i v_i$ where $\sum_i \lambda_i = 1$ and $v_i \in V_n$. Therefore, using the definition of convexity in the first inequality and that $\lambda_i \leq 1$ in the second equality, $$ f(z) = f(\sum_i \lambda_i v_i) \leq \sum \lambda_i f(v_i) \leq \underset{v \in V_n}{\text{max }} f(v) $$ -Consider now the lower bound of $f$ over the hypercube. Here we take the -approach of looking for cases where there is a guarantee that the -minimum lies at a vertex of the hypercube and when this guarantee cannot -be met, falling back to solving the convex optimization over this -hypercubic domain. For the n-dimensional approach, we will split the +Consider now the lower bound of $f$ over a hypercubic grid. Here we take the +approach of looking for hypercubes where there is a guarantee that the +minimum lies at a vertex of the hypercube and when this guarantee is not met, fall back to solving the convex optimization over that particular +hypercubic. For the n-dimensional approach, we will split the discussion into differentiable and non-differentiable $f$, and consider these separately. **Multi-Dimensional Differentiable Convex Functions** -Consider the derivatives evaluated at each vertex of the hypercube. For each $\nabla f(v)$, $v \in V_n$, take the directional derivatives, +Consider the derivatives evaluated at each vertex of a hypercube. For each $\nabla f(v)$, $v \in V_n$, take the directional derivatives, pointing inward along a hypercubic edge. Without loss of generality, recall $V_n = \{(±1,±1,…,±1) \in \mathbb{R}^n\}$ and therefore the hypercube is aligned along the standard basis vectors @@ -340,10 +339,8 @@ derivative along the line at $w$, pointing inwards, is given by, $$ \hat{n} \cdot \nabla f(w) = \sum_i -|n_i|\cdot sgn(w_i) \cdot \nabla_i f(w) = \sum_i |n_i| \cdot (-sgn(w_i) \cdot \nabla_i f(w)) \geq 0 $$ -is positive, as $\hat{n} = - |n_i| \cdot sgn(w_i) \cdot e_i $. -The properties proved previously can then by applied to this 1-dimensional restriction, i.e., if the -gradient of $f$ as the interval bounds of an interval is positive, then $f$ has -a minimum value at this interval bounds. Hence, a vertex with inward +and is positive, as $\hat{n} = - |n_i| \cdot sgn(w_i) \cdot e_i $. +The properties proved previously can then be applied to this 1-dimensional restriction. Hence, a vertex with inward directional derivative signature $(+,+,…,+)$ is a lower bound for $f$ over the hypercube. ◼ If there are multiple vertices sharing this signature, then since every @@ -354,9 +351,9 @@ at vertices sharing these signatures so it is sufficient to select any. If no vertex has signature $(+,+,…,+)$, solve for the minimum using a convex optimization routine over this hypercube. Since all local minima are -global minima, there is at least one hypercube requiring this solution. +global minima, there is at least one hypercube requiring this approach. If the function has a flat section at its minima, there may be other -hypercubes in the operational design domain, also without a vertex with all positive signature. Note that empirically, +hypercubes, also without a vertex with all positive signature. Note that empirically, this seldom happens for convex neural networks as it requires fine tuning of the parameters to create such a landscape. @@ -380,7 +377,7 @@ As depicted in figure 7, the vertices $w$ of the square (hypercube of dimension bisecting these directional derivatives, into the interior of the square, has a negative gradient. This is because the vertex is at the intersection of two planes and is a non-differentiable point, so the derivative through this point is path -dependent. This is a well-known observation but this breaks the assertion that this vertex if the minimum of $f$ over this +dependent. This is a well-known property of non-differentiable functions and breaks the assertion that this vertex is the minimum of $f$ over this square region. From this example, it is clear the minimum lies at the apex at $(0,0)$. To ameliorate this issue, in the case that the convex function is @@ -391,13 +388,9 @@ $relu$ operations. In practice, this means that a vertex may be a non-differentiable point if the network has pre-activations to $relu$ layers that have exact zeros. In practice, this is seldom the case. The probability of this occurring can be further reduced by offsetting any -hypercube or hypercubic grid origin by a small random perturbation. It -is assumed during training, for efficiency of computing bounds during training, that the convex neural network is differentiable everywhere. For final post-training analysis, this implementation checks the $relu$ -pre-activations for any exact zeros for all vertices. If there are -any zeros in these pre-activations, lower bounds for hypercubes that contain that vertex are recomputed using -an minimization routine. As a demonstration that these bounds are -correct, in the examples, we also run the minimization optimization routine on every -hypercube to show that bounds agree. +hypercube or hypercubic grid origin by a small random perturbation. If there are +any zeros in these pre-activations, lower bounds for hypercubes that contain that vertex can be recomputed using +a convex optimization routine instead. As a final comment, for general convex hulls, the argument for the upper bound value of the function over the convex hull trivially extends, defined as the largest function value over the set of points defining the hull. The lower bound should be determined using an optimization routine, constrained to the set of point in the convex hull. diff --git a/documentation/AI-Verification-Monotonicity.md b/documentation/AI-Verification-Monotonicity.md index bd56744..747e5b0 100644 --- a/documentation/AI-Verification-Monotonicity.md +++ b/documentation/AI-Verification-Monotonicity.md @@ -27,7 +27,7 @@ To circumvent these challenges, an alternative approach is to construct neural n - **Constrained Weights**: Ensuring that all weights in the network are non-negative can guarantee monotonicity. You can achieve this by using techniques like weight clipping or transforming weights during training. - **Architectural Considerations**: Designing network architectures that facilitate monotonic behavior. For example, architectures that avoid certain types of skip connections or layer types that could introduce non-monotonic behavior. -The approach taken in this repository is to utilize a combination of these three aspects and is based on the construction outlined in [1]. Ref [1] discusses the derivation in the context of row vector representations of network inputs however MATLAB utilizes a column vector representation of network inputs. This means that the 1-norm discussed in [1] is replaced by the $\infty$-norm for implementations in MATLAB. +The approach taken in this repository is to utilize a combination of activation function, weight and architectural restrictions and is based on the construction outlined in [1]. Ref [1] discusses the derivation in the context of row vector representations of network inputs however MATLAB utilizes a column vector representation of network inputs. This means that the 1-norm discussed in [1] is replaced by the $\infty$-norm for implementations in MATLAB. Note that for different choices of p-norm, the derivation in [1] still yields a monotonic function $f$, however there may be couplings between the magnitudes of the partial derivatives (shown for p=2 in [1]). By default, the implementation in this repository sets $p=\infty$ for monotonic networks but other values are explored as these may yield better fits. @@ -50,7 +50,7 @@ The main challenge with expressive monotonic networks is to balance the inherent For networks constructed to be monotonic, verification becomes more straightforward and comes down to architectural and weight inspection, i.e., provided the network architecture is of a specified monotonic topology, and that the weights in the network are appropriately related - see [1] - then the network is monotonic. -In summary, while verifying monotonicity in general neural networks is complex due to non-linearities and high dimensionality, constructing networks with inherent monotonic properties simplifies verification. By using monotonic activation functions and ensuring non-negative weights, you can design networks that are guaranteed to be monotonic, thus facilitating the verification process and making the network more suitable for applications where monotonic behavior is essential. +In summary, while verifying monotonicity in general neural networks is complex due to non-linearities and high dimensionality, constructing networks with inherent monotonic properties simplifies verification. By using constrained architectures and weights, you can design networks that are guaranteed to be monotonic, thus facilitating the verification process and making the network more suitable for applications where monotonic behavior is essential. **References**