Long-Short Transformer: Efficient Transformers for Language and Vision

Transformers have achieved success in both language and vision domains. However, it is prohibitively expensive to scale them to long sequences such as long documents or high-resolution images, because self-attention mechanism has quadratic time and memory complexities with respect to the input sequence length. In this paper, we propose Long-Short Transformer (Transformer-LS), an efficient self-attention mechanism for modeling long sequences with linear complexity for both language and vision tasks. It aggregates a novel long-range attention with dynamic projection to model distant correlations and a short-term attention to capture fine-grained local correlations. We propose a dual normalization strategy to account for the scale mismatch between the two attention mechanisms. Transformer-LS can be applied to both autoregressive and bidirectional models without additional complexity. Our method outperforms the state-of-the-art models on multiple tasks in language and vision domains, including the Long Range Arena benchmark, autoregressive language modeling, and ImageNet classification. For instance, Transformer-LS achieves 0.97 test BPC on enwik8 using half the number of parameters than previous method, while being faster and is able to handle 3x as long sequences compared to its full-attention version on the same hardware. On ImageNet, it can obtain the state-of-the-art results (e.g., a moderate size of 55.8M model solely trained on 224x224 ImageNet-1K can obtain Top-1 accuracy 84.1%), while being more scalable on high-resolution images. The source code and models are released at https://github.com/NVIDIA/transformer-ls .

A Field Guide to Federated Optimization

Federated learning and analytics are a distributed approach for collaboratively learning models (or statistics) from decentralized data, motivated by and designed for privacy protection. The distributed learning process can be formulated as solving federated optimization problems, which emphasize communication efficiency, data heterogeneity, compatibility with privacy and system requirements, and other constraints that are not primary considerations in other problem settings. This paper provides recommendations and guidelines on formulating, designing, evaluating and analyzing federated optimization algorithms through concrete examples and practical implementation, with a focus on conducting effective simulations to infer real-world performance. The goal of this work is not to survey the current literature, but to inspire researchers and practitioners to design federated learning algorithms that can be used in various practical applications.

Inductive Predictions of Extreme Hydrologic Events in The Wabash River Watershed

We present a machine learning method to predict extreme hydrologic events from spatially and temporally varying hydrological and meteorological data. We used a timestep reduction technique to reduce the computational and memory requirements and trained a bidirection LSTM network to predict soil water and stream flow from time series data observed and simulated over eighty years in the Wabash River Watershed. We show that our simple model can be trained much faster than complex attention networks such as GeoMAN without sacrificing accuracy. Based on the predicted values of soil water and stream flow, we predict the occurrence and severity of extreme hydrologic events such as droughts. We also demonstrate that extreme events can be predicted in geographical locations separate from locations observed during the training process. This spatially-inductive setting enables us to predict extreme events in other areas in the US and other parts of the world using our model trained with the Wabash Basin data.

The Intrinsic Dimension of Images and Its Impact on Learning

It is widely believed that natural image data exhibits low-dimensional structure despite the high dimensionality of conventional pixel representations. This idea underlies a common intuition for the remarkable success of deep learning in computer vision. In this work, we apply dimension estimation tools to popular datasets and investigate the role of low-dimensional structure in deep learning. We find that common natural image datasets indeed have very low intrinsic dimension relative to the high number of pixels in the images. Additionally, we find that low dimensional datasets are easier for neural networks to learn, and models solving these tasks generalize better from training to test data. Along the way, we develop a technique for validating our dimension estimation tools on synthetic data generated by GANs allowing us to actively manipulate the intrinsic dimension by controlling the image generation process. Code for our experiments may be found here https://github.com/ppope/dimensions.

GradInit: Learning to Initialize Neural Networks for Stable and Efficient Training

Changes in neural architectures have fostered significant breakthroughs in language modeling and computer vision. Unfortunately, novel architectures often require re-thinking the choice of hyperparameters (e.g., learning rate, warmup schedule, and momentum coefficients) to maintain stability of the optimizer. This optimizer instability is often the result of poor parameter initialization, and can be avoided by architecture-specific initialization schemes. In this paper, we present GradInit, an automated and architecture agnostic method for initializing neural networks. GradInit is based on a simple heuristic; the variance of each network layer is adjusted so that a single step of SGD or Adam results in the smallest possible loss value. This adjustment is done by introducing a scalar multiplier variable in front of each parameter block, and then optimizing these variables using a simple numerical scheme. GradInit accelerates the convergence and test performance of many convolutional architectures, both with or without skip connections, and even without normalization layers. It also enables training the original Post-LN Transformer for machine translation without learning rate warmup under a wide range of learning rates and momentum coefficients. Code is available at https://github.com/zhuchen03/gradinit.

Modifying Memories in Transformer Models

Large Transformer models have achieved impressive performance in many natural language tasks. In particular, Transformer based language models have been shown to have great capabilities in encoding factual knowledge in their vast amount of parameters. While the tasks of improving the memorization and generalization of Transformers have been widely studied, it is not well known how to make transformers forget specific old facts and memorize new ones. In this paper, we propose a new task of \emph{explicitly modifying specific factual knowledge in Transformer models while ensuring the model performance does not degrade on the unmodified facts}. This task is useful in many scenarios, such as updating stale knowledge, protecting privacy, and eliminating unintended biases stored in the models. We benchmarked several approaches that provide natural baseline performances on this task. This leads to the discovery of key components of a Transformer model that are especially effective for knowledge modifications. The work also provides insights into the role that different training phases (such as pretraining and fine-tuning) play towards memorization and knowledge modification.

Spherical Knowledge Distillation

Knowledge distillation aims at obtaining a small but effective deep model by transferring knowledge from a much larger one. The previous approaches try to reach this goal by simply "logit-supervised" information transferring between the teacher and student, which somehow can be subsequently decomposed as the transfer of normalized logits and $l^2$ norm. We argue that the norm of logits is actually interference, which damages the efficiency in the transfer process. To address this problem, we propose Spherical Knowledge Distillation (SKD). Specifically, we project the teacher and the student's logits into a unit sphere, and then we can efficiently perform knowledge distillation on the sphere. We verify our argument via theoretical analysis and ablation study. Extensive experiments have demonstrated the superiority and scalability of our method over the SOTAs.

Are Adversarial Examples Created Equal? A Learnable Weighted Minimax Risk for Robustness under Non-uniform Attacks

Adversarial Training is proved to be an efficient method to defend against adversarial examples, being one of the few defenses that withstand strong attacks. However, traditional defense mechanisms assume a uniform attack over the examples according to the underlying data distribution, which is apparently unrealistic as the attacker could choose to focus on more vulnerable examples. We present a weighted minimax risk optimization that defends against non-uniform attacks, achieving robustness against adversarial examples under perturbed test data distributions. Our modified risk considers importance weights of different adversarial examples and focuses adaptively on harder examples that are wrongly classified or at higher risk of being classified incorrectly. The designed risk allows the training process to learn a strong defense through optimizing the importance weights. The experiments show that our model significantly improves state-of-the-art adversarial accuracy under non-uniform attacks without a significant drop under uniform attacks.

FLAG: Adversarial Data Augmentation for Graph Neural Networks

Data augmentation helps neural networks generalize better, but it remains an open question how to effectively augment graph data to enhance the performance of GNNs (Graph Neural Networks). While most existing graph regularizers focus on augmenting graph topological structures by adding/removing edges, we offer a novel direction to augment in the input node feature space for better performance. We propose a simple but effective solution, FLAG (Free Large-scale Adversarial Augmentation on Graphs), which iteratively augments node features with gradient-based adversarial perturbations during training, and boosts performance at test time. Empirically, FLAG can be easily implemented with a dozen lines of code and is flexible enough to function with any GNN backbone, on a wide variety of large-scale datasets, and in both transductive and inductive settings. Without modifying a model's architecture or training setup, FLAG yields a consistent and salient performance boost across both node and graph classification tasks. Using FLAG, we reach state-of-the-art performance on the large-scale ogbg-molpcba, ogbg-ppa, and ogbg-code datasets.