Error in Refinement with polynomial relaxations of order > 2 for ReLU NNs

[phK3]

Propagating Polynomial Zonotopes through a ReLU neural network fails, for polynomial relaxations of order > 2.

Using the mnist-net_256x4.onnx (available here), the code

evParams = struct;
evParams.num_generators = 10000;
evParams.reuse_bounds = true;

nn = neuralNetwork.readONNXNetwork("../mnist-net_256x4.onnx", false, 'SSC');
lbs = zeros(784,1);
ubs = zeros(784,1);
lbs(390:394) = 0;
ubs(390:394) = 1;

X = polyZonotope(X0);
Y = nn.evaluate(X, evParams);
nn.refine(3, 'layer', 'both', X.randPoint(1), true);
Y = nn.evaluate(X, evParams);
nn.refine(3, 'layer', 'both', X.randPoint(1), true);
Y = nn.evaluate(X, evParams);

fails with error message

Error using  + 
Arrays have incompatible sizes for this operation.

Error in nnActivationLayer/evaluatePolyZonotope>aux_preOrderReduction (line 161)
    id = [id; id_ + (1:q)'];

Error in nnActivationLayer/evaluatePolyZonotope (line 57)
[c, G, GI, E, id, id_, ind, ind_] = aux_preOrderReduction(obj, c, G, GI, E, id, id_, ind, ind_, evParams);

Error in neuralNetwork/evaluate>aux_evaluatePolyZonotope (line 178)
                layer_i.evaluatePolyZonotope(c, G, GI, E, id, id_, ind, ind_, evParams);

Error in neuralNetwork/evaluate (line 92)
    r = aux_evaluatePolyZonotope(obj, input, evParams, idxLayer);

Error in mwe_refine_error (line 17)
Y = nn.evaluate(X, evParams);

Related documentation

[toladnertum]

Dear phK3,

Thank you for telling us about this bug. Should be fixed with the next release of CORA (v2024.2.1).

Best, Tobias

[toladnertum]

CORA v2024.2.1 is now released.