Caltech

Abstract: I will argue that classifying images is significantly easier than simulating two-dimensional quantum systems from the perspective of entanglement scaling. In particular, I show that under reasonable assumptions, the entanglement between a region and its complement scales as the logarithm of the boundary length for image classification problems. This implies a provably efficient neural network representation for the function that maps an image to the label the image corresponds to.
Reference: arXiv:1711.04606

Contact: Lei Wang 9853