NN Class¶

The unified neural network class for feedforward, convolutional, recurrent, segmentation, and ODE-based architectures.

Constructors¶

nn = NN(Layers)
nn = NN(Layers, Connections)
nn = NN(Layers, Connections, inputSize, outputSize, name)

Layers – cell array of layer objects (FullyConnectedLayer, ReluLayer, etc.)
Connections – table defining DAG connectivity (optional; omit for sequential networks)
inputSize, outputSize – dimension arrays
name – string identifier

Properties¶

Property	Type	Description
`Layers`	cell array	Ordered list of layer objects
`Connections`	table	DAG connectivity (source/destination layer names)
`Name`	string	Network name
`numLayers`	int	Number of layers
`numNeurons`	int	Total neuron count across all layers
`InputSize`	array	Input dimensions
`OutputSize`	array	Output dimensions
`reachSet`	cell array	Computed reachable set per layer (after calling reach)
`reachTime`	array	Computation time per layer (after calling reach)

Methods¶

Method	Description
`y = nn.evaluate(x)`	Forward pass on a single concrete input vector/image/tensor
`outputSet = nn.reach(inputSet, reachOptions)`	Compute reachable output set from an input set (Star, ImageStar, etc.)
`result = nn.verify_robustness(inputSet, reachOptions, target)`	Verify classification robustness. Returns: 1=robust, 0=not robust, 2=unknown. For exact methods, unknown is treated as not robust (returns 0).
`[result, X] = nn.verify_vnnlib(vnnlib_file, reachOptions)`	Verify properties from a VNNLIB file. Returns result (0/1/2) and the input set X.
`result = nn.verify_safety(inputSet, unsafeRegion, reachOptions)`	Verify safety against a HalfSpace unsafe region
`result = nn.verify_segmentation(inputSet, reachOptions, target)`	Verify pixel-level robustness for semantic segmentation networks
`label = nn.classify(inputValue, reachOptions)`	Classify input(s); handles both single points and sets
`rb = nn.checkRobust(outputSet, target)`	Check if a precomputed reachable set satisfies robustness for target class
`counter = nn.falsify(I, U, n_samples)`	Find counterexample inputs violating specification via random sampling
`result = nn.verify_sequence_robustness(input_seq, eps, target, reachOptions)`	Verify robustness for RNN/sequence classification over time steps

Example¶

% Build a simple network
layers = {FullyConnectedLayer(W1, b1), ReluLayer(), FullyConnectedLayer(W2, b2)};
net = NN(layers);

% Evaluate
y = net.evaluate([0.5; 0.5]);

% Verify
input_set = Star([-1; -1], [1; 1]);
reachOptions.reachMethod = 'approx-star';
output_sets = net.reach(input_set, reachOptions);
result = net.verify_robustness(input_set, reachOptions, 1);