A Scale Invariant Flatness Measure for Deep Network Minima
This work provides a more reliable tool for analyzing generalization in deep learning, though it is incremental as it builds on existing flatness concepts.
The paper addresses the issue that existing flatness measures for deep network minima are not scale-invariant, making them unreliable for generalization analysis, and proposes a new Hessian-based measure that is invariant to rescaling, which they use to confirm that large-batch SGD minima are sharper than small-batch ones.
It has been empirically observed that the flatness of minima obtained from training deep networks seems to correlate with better generalization. However, for deep networks with positively homogeneous activations, most measures of sharpness/flatness are not invariant to rescaling of the network parameters, corresponding to the same function. This means that the measure of flatness/sharpness can be made as small or as large as possible through rescaling, rendering the quantitative measures meaningless. In this paper we show that for deep networks with positively homogenous activations, these rescalings constitute equivalence relations, and that these equivalence relations induce a quotient manifold structure in the parameter space. Using this manifold structure and an appropriate metric, we propose a Hessian-based measure for flatness that is invariant to rescaling. We use this new measure to confirm the proposition that Large-Batch SGD minima are indeed sharper than Small-Batch SGD minima.