Recently, the information-theoretical framework has been proven to be able to obtain non-vacuous generalization bounds for large models trained by Stochastic Gradient Langevin Dynamics (SGLD) with isotropic noise. In this paper, we optimize the information-theoretical generalization bound by manipulating the noise structure in SGLD. We prove that with constraint to guarantee low empirical risk, the optimal noise covariance is the square root of the expected gradient covariance if both the prior and the posterior are jointly optimized. This validates that the optimal noise is quite close to the empirical gradient covariance. Technically, we develop a new information-theoretical bound that enables such an optimization analysis. We then apply matrix analysis to derive the form of optimal noise covariance. Presented constraint and results are validated by the empirical observations.
Modeling many-body systems has been a long-standing challenge in science, from classical and quantum physics to computational biology. Equivariance is a critical physical symmetry for many-body dynamic systems, which enables robust and accurate prediction under arbitrary reference transformations. In light of this, great efforts have been put on encoding this symmetry into deep neural networks, which significantly boosts the prediction performance of down-streaming tasks. Some general equivariant models which are computationally efficient have been proposed, however, these models have no guarantee on the approximation power and may have information loss. In this paper, we leverage insights from the scalarization technique in differential geometry to model many-body systems by learning the gradient vector fields, which are SE(3) and permutation equivariant. Specifically, we propose the Equivariant Vector Field Network (EVFN), which is built on a novel tuple of equivariant basis and the associated scalarization and vectorization layers. Since our tuple equivariant basis forms a complete basis, learning the dynamics with our EVFN has no information loss and no tensor operations are involved before the final vectorization, which reduces the complex optimization on tensors to a minimum. We evaluate our method on predicting trajectories of simulated Newton mechanics systems with both full and partially observed data, as well as the equilibrium state of small molecules (molecular conformation) evolving as a statistical mechanics system. Experimental results across multiple tasks demonstrate that our model achieves best or competitive performance on baseline models in various types of datasets.
The momentum acceleration technique is widely adopted in many optimization algorithms. However, the theoretical understanding of how the momentum affects the generalization performance of the optimization algorithms is still unknown. In this paper, we answer this question through analyzing the implicit bias of momentum-based optimization. We prove that both SGD with momentum and Adam converge to the $L_2$ max-margin solution for exponential-tailed loss, which is the same as vanilla gradient descent. That means, these optimizers with momentum acceleration still converge to a model with low complexity, which provides guarantees on their generalization. Technically, to overcome the difficulty brought by the error accumulation in analyzing the momentum, we construct new Lyapunov functions as a tool to analyze the gap between the model parameter and the max-margin solution.
Dropout is a powerful and widely used technique to regularize the training of deep neural networks. In this paper, we introduce a simple regularization strategy upon dropout in model training, namely R-Drop, which forces the output distributions of different sub models generated by dropout to be consistent with each other. Specifically, for each training sample, R-Drop minimizes the bidirectional KL-divergence between the output distributions of two sub models sampled by dropout. Theoretical analysis reveals that R-Drop reduces the freedom of the model parameters and complements dropout. Experiments on $\bf{5}$ widely used deep learning tasks ($\bf{18}$ datasets in total), including neural machine translation, abstractive summarization, language understanding, language modeling, and image classification, show that R-Drop is universally effective. In particular, it yields substantial improvements when applied to fine-tune large-scale pre-trained models, e.g., ViT, RoBERTa-large, and BART, and achieves state-of-the-art (SOTA) performances with the vanilla Transformer model on WMT14 English$\to$German translation ($\bf{30.91}$ BLEU) and WMT14 English$\to$French translation ($\bf{43.95}$ BLEU), even surpassing models trained with extra large-scale data and expert-designed advanced variants of Transformer models. Our code is available at GitHub{\url{https://github.com/dropreg/R-Drop}}.
Denoising diffusion probabilistic models have been recently proposed to generate high-quality samples by estimating the gradient of the data density. The framework assumes the prior noise as a standard Gaussian distribution, whereas the corresponding data distribution may be more complicated than the standard Gaussian distribution, which potentially introduces inefficiency in denoising the prior noise into the data sample because of the discrepancy between the data and the prior. In this paper, we propose PriorGrad to improve the efficiency of the conditional diffusion model (for example, a vocoder using a mel-spectrogram as the condition) by applying an adaptive prior derived from the data statistics based on the conditional information. We formulate the training and sampling procedures of PriorGrad and demonstrate the advantages of an adaptive prior through a theoretical analysis. Focusing on the audio domain, we consider the recently proposed diffusion-based audio generative models based on both the spectral and time domains and show that PriorGrad achieves a faster convergence leading to data and parameter efficiency and improved quality, and thereby demonstrating the efficiency of a data-driven adaptive prior.
Learning dynamics governed by differential equations is crucial for predicting and controlling the systems in science and engineering. Neural Ordinary Differential Equation (NODE), a deep learning model integrated with differential equations, learns the dynamics directly from the samples on the trajectory and shows great promise in the scientific field. However, the training of NODE highly depends on the numerical solver, which can amplify numerical noise and be unstable, especially for ill-conditioned dynamical systems. In this paper, to reduce the reliance on the numerical solver, we propose to enhance the supervised signal in learning dynamics. Specifically, beyond learning directly from the trajectory samples, we pre-train a neural differential operator (NDO) to output an estimation of the derivatives to serve as an additional supervised signal. The NDO is pre-trained on a class of symbolic functions, and it learns the mapping between the trajectory samples of these functions to their derivatives. We provide theoretical guarantee on that the output of NDO can well approximate the ground truth derivatives by proper tuning the complexity of the library. To leverage both the trajectory signal and the estimated derivatives from NDO, we propose an algorithm called NDO-NODE, in which the loss function contains two terms: the fitness on the true trajectory samples and the fitness on the estimated derivatives that are output by the pre-trained NDO. Experiments on various of dynamics show that our proposed NDO-NODE can consistently improve the forecasting accuracy.
Energy conservation is a basic physics principle, the breakdown of which often implies new physics. This paper presents a method for data-driven "new physics" discovery. Specifically, given a trajectory governed by unknown forces, our Neural New-Physics Detector (NNPhD) aims to detect new physics by decomposing the force field into conservative and non-conservative components, which are represented by a Lagrangian Neural Network (LNN) and a universal approximator network (UAN), respectively, trained to minimize the force recovery error plus a constant $\lambda$ times the magnitude of the predicted non-conservative force. We show that a phase transition occurs at $\lambda$=1, universally for arbitrary forces. We demonstrate that NNPhD successfully discovers new physics in toy numerical experiments, rediscovering friction (1493) from a damped double pendulum, Neptune from Uranus' orbit (1846) and gravitational waves (2017) from an inspiraling orbit. We also show how NNPhD coupled with an integrator outperforms previous methods for predicting the future of a damped double pendulum.