Noise Stability Optimization for Flat Minima with Tight Rates

We consider minimizing a perturbed function $F(W) = \mathbb{E}{U}[f(W + U)]$, given a function $f: \mathbb{R}^d \rightarrow \mathbb{R}$ and a random sample $U$ from a distribution $\mathcal{P}$ with mean zero. When $\mathcal{P}$ is the isotropic Gaussian, $F(W)$ is roughly equal to $f(W)$ plus a penalty on the trace of $\nabla² f(W)$, scaled by the variance of $\mathcal{P}$. This penalty on the Hessian has the benefit of improving generalization, through PAC-Bayes analysis. It is useful in low-sample regimes, for instance, when a (large) pre-trained model is fine-tuned on a small data set. One way to minimize $F$ is by adding $U$ to $W$, and then run SGD. We observe, empirically, that this noise injection does not provide significant gains over SGD, in our experiments of conducting fine-tuning on three image classification data sets. We design a simple, practical algorithm that adds noise along both $U$ and $-U$, with the option of adding several perturbations and taking their average. We analyze the convergence of this algorithm, showing tight rates on the norm of the output's gradient. We provide a comprehensive empirical analysis of our algorithm, by first showing that in an over-parameterized matrix sensing problem, it can find solutions with lower test loss than naive noise injection. Then, we compare our algorithm with four sharpness-reducing training methods (such as the Sharpness-Aware Minimization (Foret et al., 2021)). We find that our algorithm can outperform them by up to 1.8% test accuracy, for fine-tuning ResNet on six image classification data sets. It leads to a 17.7% (and 12.8%) reduction in the trace (and largest eigenvalue) of the Hessian matrix of the loss surface. This form of regularization on the Hessian is compatible with $\ell2$ weight decay (and data augmentation), in the sense that combining both can lead to improved empirical performance.