<ul><li>This work addresses the challenge of enforcing constraints on the output of neural networks, particularly for safety-critical control applications.</li><li>The proposed method utilizes reachability analysis with scaled hybrid zonotopes, which allows for the exact image of a non-convex input set to be encouraged for a neural network with rectified linear unit (ReLU) nonlinearities.</li><li>The method has shown to be effective and fast for networks with up to 240 neurons, with the computational complexity dominated by inverse operations on matrices that scale linearly in size with the number of neurons and complexity of input and unsafe sets.</li><li>The practicality of the method has been demonstrated by training a forward-invariant neural network controller for a non-convex input set and generating safe reach-avoid plans for a black-box dynamical system.</li></ul>

Provably-Safe Neural Network Training Using Hybrid Zonotope Reachability Analysis

Discover more