Jiawei Zhao

Robert Joseph George, Jiawei Zhao, Jean Kossaifi et al.

Fourier Neural Operators (FNO) offer a principled approach to solving challenging partial differential equations (PDE) such as turbulent flows. At the core of FNO is a spectral layer that leverages a discretization-convergent representation in the Fourier domain, and learns weights over a fixed set of frequencies. However, training FNO presents two significant challenges, particularly in large-scale, high-resolution applications: (i) Computing Fourier transform on high-resolution inputs is computationally intensive but necessary since fine-scale details are needed for solving many PDEs, such as fluid flows, (ii) selecting the relevant set of frequencies in the spectral layers is challenging, and too many modes can lead to overfitting, while too few can lead to underfitting. To address these issues, we introduce the Incremental Fourier Neural Operator (iFNO), which progressively increases both the number of frequency modes used by the model as well as the resolution of the training data. We empirically show that iFNO reduces total training time while maintaining or improving generalization performance across various datasets. Our method demonstrates a 10% lower testing error, using 20% fewer frequency modes compared to the existing Fourier Neural Operator, while also achieving a 30% faster training.

Jiawei Zhao, Yifei Zhang, Beidi Chen et al.

The theory of greedy low-rank learning (GLRL) aims to explain the impressive generalization capabilities of deep learning. It proves that stochastic gradient-based training implicitly regularizes neural networks towards low-rank solutions through a gradual increase of the rank during training. However, there is a gap between theory and practice since GLRL requires an infinitesimal initialization of the weights, which is not practical due to the fact that it is a saddle point. In this work, we remove the assumption of infinitesimal initialization by focusing on cumulative weight updates. We prove the cumulative weight updates follow an incremental low-rank trajectory for arbitrary orthogonal initialization of weights in a three-layer linear network. Empirically, we demonstrate that our theory holds on a broad range of neural networks (e.g., transformers) and standard training algorithms (e.g., SGD, Adam). However, existing training algorithms do not exploit the low-rank property to improve computational efficiency as the networks are not parameterized in low-rank. To remedy this, we design a new training algorithm Incremental Low-Rank Learning (InRank), which explicitly expresses cumulative weight updates as low-rank matrices while incrementally augmenting their ranks during training. We evaluate InRank on GPT-2, and our results indicate that InRank achieves comparable prediction performance as the full-rank counterpart while requiring at most 33% of the total ranks throughout training. We also propose an efficient version of InRank that achieves a reduction of 37% in total training time and 36% in model size when training GPT-medium on WikiText-103 from scratch.