Weak NAS Predictors Are All You Need

Neural Architecture Search (NAS) finds the best network architecture by exploring the architecture-to-performance manifold. It often trains and evaluates a large number of architectures, causing tremendous computation costs. Recent predictor-based NAS approaches attempt to solve this problem with two key steps: sampling some architecture-performance pairs and fitting a proxy accuracy predictor. Given limited samples, these predictors, however, are far from accurate to locate top architectures. In this paper, we shift the paradigm from finding a complicated predictor that covers the whole architecture space to a set of weaker predictors that progressively move towards the high-performance sub-space. It is based on the key property of the proposed weak predictors that their probabilities of sampling better architectures keep increasing. We thus only sample a few well-performed architectures guided by the previously learned predictor and estimate a new better weak predictor. By this coarse-to-fine iteration, the ranking of sampling space is refined gradually, which helps find the optimal architectures eventually. Experiments demonstrate that our method costs fewer samples to find the top-performance architectures on NAS-Bench-101 and NAS-Bench-201, and it achieves the state-of-the-art ImageNet performance on the NASNet search space. The code is available at https://github.com/VITA-Group/WeakNAS

TransGAN: Two Transformers Can Make One Strong GAN

The recent explosive interest on transformers has suggested their potential to become powerful "universal" models for computer vision tasks, such as classification, detection, and segmentation. However, how further transformers can go - are they ready to take some more notoriously difficult vision tasks, e.g., generative adversarial networks (GANs)? Driven by that curiosity, we conduct the first pilot study in building a GAN \textbf{completely free of convolutions}, using only pure transformer-based architectures. Our vanilla GAN architecture, dubbed \textbf{TransGAN}, consists of a memory-friendly transformer-based generator that progressively increases feature resolution while decreasing embedding dimension, and a patch-level discriminator that is also transformer-based. We then demonstrate TransGAN to notably benefit from data augmentations (more than standard GANs), a multi-task co-training strategy for the generator, and a locally initialized self-attention that emphasizes the neighborhood smoothness of natural images. Equipped with those findings, TransGAN can effectively scale up with bigger models and high-resolution image datasets. Specifically, our best architecture achieves highly competitive performance compared to current state-of-the-art GANs based on convolutional backbones. Specifically, TransGAN sets \textbf{new state-of-the-art} IS score of 10.10 and FID score of 25.32 on STL-10. It also reaches competitive 8.64 IS score and 11.89 FID score on Cifar-10, and 12.23 FID score on CelebA $64\times64$, respectively. We also conclude with a discussion of the current limitations and future potential of TransGAN. The code is available at \url{https://github.com/VITA-Group/TransGAN}.

A Unified Lottery Ticket Hypothesis for Graph Neural Networks

With graphs rapidly growing in size and deeper graph neural networks (GNNs) emerging, the training and inference of GNNs become increasingly expensive. Existing network weight pruning algorithms cannot address the main space and computational bottleneck in GNNs, caused by the size and connectivity of the graph. To this end, this paper first presents a unified GNN sparsification (UGS) framework that simultaneously prunes the graph adjacency matrix and the model weights, for effectively accelerating GNN inference on large-scale graphs. Leveraging this new tool, we further generalize the recently popular lottery ticket hypothesis to GNNs for the first time, by defining a graph lottery ticket (GLT) as a pair of core sub-dataset and sparse sub-network, which can be jointly identified from the original GNN and the full dense graph by iteratively applying UGS. Like its counterpart in convolutional neural networks, GLT can be trained in isolation to match the performance of training with the full model and graph, and can be drawn from both randomly initialized and self-supervised pre-trained GNNs. Our proposal has been experimentally verified across various GNN architectures and diverse tasks, on both small-scale graph datasets (Cora, Citeseer and PubMed), and large-scale datasets from the challenging Open Graph Benchmark (OGB). Specifically, for node classification, our found GLTs achieve the same accuracies with 20%~98% MACs saving on small graphs and 25%~85% MACs saving on large ones. For link prediction, GLTs lead to 48%~97% and 70% MACs saving on small and large graph datasets, respectively, without compromising predictive performance. Codes available at https://github.com/VITA-Group/Unified-LTH-GNN.

On Dynamic Noise Influence in Differentially Private Learning

Protecting privacy in learning while maintaining the model performance has become increasingly critical in many applications that involve sensitive data. Private Gradient Descent (PGD) is a commonly used private learning framework, which noises gradients based on the Differential Privacy protocol. Recent studies show that \emph{dynamic privacy schedules} of decreasing noise magnitudes can improve loss at the final iteration, and yet theoretical understandings of the effectiveness of such schedules and their connections to optimization algorithms remain limited. In this paper, we provide comprehensive analysis of noise influence in dynamic privacy schedules to answer these critical questions. We first present a dynamic noise schedule minimizing the utility upper bound of PGD, and show how the noise influence from each optimization step collectively impacts utility of the final model. Our study also reveals how impacts from dynamic noise influence change when momentum is used. We empirically show the connection exists for general non-convex losses, and the influence is greatly impacted by the loss curvature.

Good Students Play Big Lottery Better

Lottery ticket hypothesis suggests that a dense neural network contains a sparse sub-network that can match the test accuracy of the original dense net when trained in isolation from (the same) random initialization. However, the hypothesis failed to generalize to larger dense networks such as ResNet-50. As a remedy, recent studies demonstrate that a sparse sub-network can still be obtained by using a rewinding technique, which is to re-train it from early-phase training weights or learning rates of the dense model, rather than from random initialization. Is rewinding the only or the best way to scale up lottery tickets? This paper proposes a new, simpler and yet powerful technique for re-training the sub-network, called "Knowledge Distillation ticket" (KD ticket). Rewinding exploits the value of inheriting knowledge from the early training phase to improve lottery tickets in large networks. In comparison, KD ticket addresses a complementary possibility - inheriting useful knowledge from the late training phase of the dense model. It is achieved by leveraging the soft labels generated by the trained dense model to re-train the sub-network, instead of the hard labels. Extensive experiments are conducted using several large deep networks (e.g ResNet-50 and ResNet-110) on CIFAR-10 and ImageNet datasets. Without bells and whistles, when applied by itself, KD ticket performs on par or better than rewinding, while being nearly free of hyperparameters or ad-hoc selection. KD ticket can be further applied together with rewinding, yielding state-of-the-art results for large-scale lottery tickets.

SmartDeal: Re-Modeling Deep Network Weights for Efficient Inference and Training

The record-breaking performance of deep neural networks (DNNs) comes with heavy parameterization, leading to external dynamic random-access memory (DRAM) for storage. The prohibitive energy of DRAM accesses makes it non-trivial to deploy DNN on resource-constrained devices, calling for minimizing the weight and data movements to improve the energy efficiency. We present SmartDeal (SD), an algorithm framework to trade higher-cost memory storage/access for lower-cost computation, in order to aggressively boost the storage and energy efficiency, for both inference and training. The core of SD is a novel weight decomposition with structural constraints, carefully crafted to unleash the hardware efficiency potential. Specifically, we decompose each weight tensor as the product of a small basis matrix and a large structurally sparse coefficient matrix whose non-zeros are quantized to power-of-2. The resulting sparse and quantized DNNs enjoy greatly reduced energy for data movement and weight storage, incurring minimal overhead to recover the original weights thanks to the sparse bit-operations and cost-favorable computations. Beyond inference, we take another leap to embrace energy-efficient training, introducing innovative techniques to address the unique roadblocks arising in training while preserving the SD structures. We also design a dedicated hardware accelerator to fully utilize the SD structure to improve the real energy efficiency and latency. We conduct experiments on both multiple tasks, models and datasets in different settings. Results show that: 1) applied to inference, SD achieves up to 2.44x energy efficiency as evaluated via real hardware implementations; 2) applied to training, SD leads to 10.56x and 4.48x reduction in the storage and training energy, with negligible accuracy loss compared to state-of-the-art training baselines. Our source codes are available online.

EarlyBERT: Efficient BERT Training via Early-bird Lottery Tickets

Deep, heavily overparameterized language models such as BERT, XLNet and T5 have achieved impressive success in many NLP tasks. However, their high model complexity requires enormous computation resources and extremely long training time for both pre-training and fine-tuning. Many works have studied model compression on large NLP models, but only focus on reducing inference cost/time, while still requiring expensive training process. Other works use extremely large batch sizes to shorten the pre-training time at the expense of high demand for computation resources. In this paper, inspired by the Early-Bird Lottery Tickets studied for computer vision tasks, we propose EarlyBERT, a general computationally-efficient training algorithm applicable to both pre-training and fine-tuning of large-scale language models. We are the first to identify structured winning tickets in the early stage of BERT training, and use them for efficient training. Comprehensive pre-training and fine-tuning experiments on GLUE and SQuAD downstream tasks show that EarlyBERT easily achieves comparable performance to standard BERT with 35~45% less training time.

Growing Deep Forests Efficiently with Soft Routing and Learned Connectivity

Despite the latest prevailing success of deep neural networks (DNNs), several concerns have been raised against their usage, including the lack of intepretability the gap between DNNs and other well-established machine learning models, and the growingly expensive computational costs. A number of recent works [1], [2], [3] explored the alternative to sequentially stacking decision tree/random forest building blocks in a purely feed-forward way, with no need of back propagation. Since decision trees enjoy inherent reasoning transparency, such deep forest models can also facilitate the understanding of the internaldecision making process. This paper further extends the deep forest idea in several important aspects. Firstly, we employ a probabilistic tree whose nodes make probabilistic routing decisions, a.k.a., soft routing, rather than hard binary decisions.Besides enhancing the flexibility, it also enables non-greedy optimization for each tree. Second, we propose an innovative topology learning strategy: every node in the ree now maintains a new learnable hyperparameter indicating the probability that it will be a leaf node. In that way, the tree will jointly optimize both its parameters and the tree topology during training. Experiments on the MNIST dataset demonstrate that our empowered deep forests can achieve better or comparable performance than [1],[3] , with dramatically reduced model complexity. For example,our model with only 1 layer of 15 trees can perform comparably with the model in [3] with 2 layers of 2000 trees each.