Input-Convex Deep Networks

• We introduce a new neural network architecture: – Input-Convex Neural Networks (ICNNs) • Definition: Scalar-valued neural network f(x; θ) – f is convex in the input x – (f is not convex in the parameters θ = {Wi, bi}) • Model allows global optimization over some of the inputs to the network, given fixed values for other inputs • Many existing neural-network architectures can be “easily” made input-convex