The Elastic Lottery Ticket Hypothesis

Lottery Ticket Hypothesis raises keen attention to identifying sparse trainable subnetworks or winning tickets, at the initialization (or early stage) of training, which can be trained in isolation to achieve similar or even better performance compared to the full models. Despite many efforts being made, the most effective method to identify such winning tickets is still Iterative Magnitude-based Pruning (IMP), which is computationally expensive and has to be run thoroughly for every different network. A natural question that comes in is: can we "transform" the winning ticket found in one network to another with a different architecture, yielding a winning ticket for the latter at the beginning, without re-doing the expensive IMP? Answering this question is not only practically relevant for efficient "once-for-all" winning ticket finding, but also theoretically appealing for uncovering inherently scalable sparse patterns in networks. We conduct extensive experiments on CIFAR-10 and ImageNet, and propose a variety of strategies to tweak the winning tickets found from different networks of the same model family (e.g., ResNets). Based on these results, we articulate the Elastic Lottery Ticket Hypothesis (E-LTH): by mindfully replicating (or dropping) and re-ordering layers for one network, its corresponding winning ticket could be stretched (or squeezed) into a subnetwork for another deeper (or shallower) network from the same family, whose performance is nearly as competitive as the latter's winning ticket directly found by IMP. We have also thoroughly compared E-LTH with pruning-at-initialization and dynamic sparse training methods, and discuss the generalizability of E-LTH to different model families, layer types, and even across datasets. Our codes are publicly available at https://github.com/VITA-Group/ElasticLTH.

Learning to Optimize: A Primer and A Benchmark

Learning to optimize (L2O) is an emerging approach that leverages machine learning to develop optimization methods, aiming at reducing the laborious iterations of hand engineering. It automates the design of an optimization method based on its performance on a set of training problems. This data-driven procedure generates methods that can efficiently solve problems similar to those in the training. In sharp contrast, the typical and traditional designs of optimization methods are theory-driven, so they obtain performance guarantees over the classes of problems specified by the theory. The difference makes L2O suitable for repeatedly solving a certain type of optimization problems over a specific distribution of data, while it typically fails on out-of-distribution problems. The practicality of L2O depends on the type of target optimization, the chosen architecture of the method to learn, and the training procedure. This new paradigm has motivated a community of researchers to explore L2O and report their findings. This article is poised to be the first comprehensive survey and benchmark of L2O for continuous optimization. We set up taxonomies, categorize existing works and research directions, present insights, and identify open challenges. We also benchmarked many existing L2O approaches on a few but representative optimization problems. For reproducible research and fair benchmarking purposes, we released our software implementation and data in the package Open-L2O at https://github.com/VITA-Group/Open-L2O.

UltraSR: Spatial Encoding is a Missing Key for Implicit Image Function-based Arbitrary-Scale Super-Resolution

The recent success of NeRF and other related implicit neural representation methods has opened a new path for continuous image representation, where pixel values no longer need to be looked up from stored discrete 2D arrays but can be inferred from neural network models on a continuous spatial domain. Although the recent work LIIF has demonstrated that such novel approach can achieve good performance on the arbitrary-scale super-resolution task, their upscaled images frequently show structural distortion due to the faulty prediction on high-frequency textures. In this work, we propose UltraSR, a simple yet effective new network design based on implicit image functions in which spatial coordinates and periodic encoding are deeply integrated with the implicit neural representation. We show that spatial encoding is indeed a missing key towards the next-stage high-accuracy implicit image function through extensive experiments and ablation studies. Our UltraSR sets new state-of-the-art performance on the DIV2K benchmark under all super-resolution scales comparing to previous state-of-the-art methods. UltraSR also achieves superior performance on other standard benchmark datasets in which it outperforms prior works in almost all experiments. Our code will be released at https://github.com/SHI-Labs/UltraSR-Arbitrary-Scale-Super-Resolution.

Adversarial Feature Augmentation and Normalization for Visual Recognition

Recent advances in computer vision take advantage of adversarial data augmentation to ameliorate the generalization ability of classification models. Here, we present an effective and efficient alternative that advocates adversarial augmentation on intermediate feature embeddings, instead of relying on computationally-expensive pixel-level perturbations. We propose Adversarial Feature Augmentation and Normalization (A-FAN), which (i) first augments visual recognition models with adversarial features that integrate flexible scales of perturbation strengths, (ii) then extracts adversarial feature statistics from batch normalization, and re-injects them into clean features through feature normalization. We validate the proposed approach across diverse visual recognition tasks with representative backbone networks, including ResNets and EfficientNets for classification, Faster-RCNN for detection, and Deeplab V3+ for segmentation. Extensive experiments show that A-FAN yields consistent generalization improvement over strong baselines across various datasets for classification, detection and segmentation tasks, such as CIFAR-10, CIFAR-100, ImageNet, Pascal VOC2007, Pascal VOC2012, COCO2017, and Cityspaces. Comprehensive ablation studies and detailed analyses also demonstrate that adding perturbations to specific modules and layers of classification/detection/segmentation backbones yields optimal performance. Codes and pre-trained models will be made available at: https://github.com/VITA-Group/CV_A-FAN.

Neural Architecture Search on ImageNet in Four GPU Hours: A Theoretically Inspired Perspective

Neural Architecture Search (NAS) has been explosively studied to automate the discovery of top-performer neural networks. Current works require heavy training of supernet or intensive architecture evaluations, thus suffering from heavy resource consumption and often incurring search bias due to truncated training or approximations. Can we select the best neural architectures without involving any training and eliminate a drastic portion of the search cost? We provide an affirmative answer, by proposing a novel framework called training-free neural architecture search (TE-NAS). TE-NAS ranks architectures by analyzing the spectrum of the neural tangent kernel (NTK) and the number of linear regions in the input space. Both are motivated by recent theory advances in deep networks and can be computed without any training and any label. We show that: (1) these two measurements imply the trainability and expressivity of a neural network; (2) they strongly correlate with the network's test accuracy. Further on, we design a pruning-based NAS mechanism to achieve a more flexible and superior trade-off between the trainability and expressivity during the search. In NAS-Bench-201 and DARTS search spaces, TE-NAS completes high-quality search but only costs 0.5 and 4 GPU hours with one 1080Ti on CIFAR-10 and ImageNet, respectively. We hope our work inspires more attempts in bridging the theoretical findings of deep networks and practical impacts in real NAS applications. Code is available at: https://github.com/VITA-Group/TENAS.

Exposing Semantic Segmentation Failures via Maximum Discrepancy Competition

Semantic segmentation is an extensively studied task in computer vision, with numerous methods proposed every year. Thanks to the advent of deep learning in semantic segmentation, the performance on existing benchmarks is close to saturation. A natural question then arises: Does the superior performance on the closed (and frequently re-used) test sets transfer to the open visual world with unconstrained variations? In this paper, we take steps toward answering the question by exposing failures of existing semantic segmentation methods in the open visual world under the constraint of very limited human labeling effort. Inspired by previous research on model falsification, we start from an arbitrarily large image set, and automatically sample a small image set by MAximizing the Discrepancy (MAD) between two segmentation methods. The selected images have the greatest potential in falsifying either (or both) of the two methods. We also explicitly enforce several conditions to diversify the exposed failures, corresponding to different underlying root causes. A segmentation method, whose failures are more difficult to be exposed in the MAD competition, is considered better. We conduct a thorough MAD diagnosis of ten PASCAL VOC semantic segmentation algorithms. With detailed analysis of experimental results, we point out strengths and weaknesses of the competing algorithms, as well as potential research directions for further advancement in semantic segmentation. The codes are publicly available at \url{https://github.com/QTJiebin/MAD_Segmentation}.

Ultra-Data-Efficient GAN Training: Drawing A Lottery Ticket First, Then Training It Toughly

Training generative adversarial networks (GANs) with limited data generally results in deteriorated performance and collapsed models. To conquer this challenge, we are inspired by the latest observation of Kalibhat et al. (2020); Chen et al.(2021d), that one can discover independently trainable and highly sparse subnetworks (a.k.a., lottery tickets) from GANs. Treating this as an inductive prior, we decompose the data-hungry GAN training into two sequential sub-problems: (i) identifying the lottery ticket from the original GAN; then (ii) training the found sparse subnetwork with aggressive data and feature augmentations. Both sub-problems re-use the same small training set of real images. Such a coordinated framework enables us to focus on lower-complexity and more data-efficient sub-problems, effectively stabilizing training and improving convergence. Comprehensive experiments endorse the effectiveness of our proposed ultra-data-efficient training framework, across various GAN architectures (SNGAN, BigGAN, and StyleGAN2) and diverse datasets (CIFAR-10, CIFAR-100, Tiny-ImageNet, and ImageNet). Besides, our training framework also displays powerful few-shot generalization ability, i.e., generating high-fidelity images by training from scratch with just 100 real images, without any pre-training. Codes are available at: https://github.com/VITA-Group/Ultra-Data-Efficient-GAN-Training.

Sandwich Batch Normalization

We present Sandwich Batch Normalization (SaBN), an embarrassingly easy improvement of Batch Normalization (BN) with only a few lines of code changes. SaBN is motivated by addressing the inherent feature distribution heterogeneity that one can be identified in many tasks, which can arise from data heterogeneity (multiple input domains) or model heterogeneity (dynamic architectures, model conditioning, etc.). Our SaBN factorizes the BN affine layer into one shared sandwich affine layer, cascaded by several parallel independent affine layers. Concrete analysis reveals that, during optimization, SaBN promotes balanced gradient norms while still preserving diverse gradient directions: a property that many application tasks seem to favor. We demonstrate the prevailing effectiveness of SaBN as a drop-in replacement in four tasks: $\textbf{conditional image generation}$, $\textbf{neural architecture search}$ (NAS), $\textbf{adversarial training}$, and $\textbf{arbitrary style transfer}$. Leveraging SaBN immediately achieves better Inception Score and FID on CIFAR-10 and ImageNet conditional image generation with three state-of-the-art GANs; boosts the performance of a state-of-the-art weight-sharing NAS algorithm significantly on NAS-Bench-201; substantially improves the robust and standard accuracies for adversarial defense; and produces superior arbitrary stylized results. We also provide visualizations and analysis to help understand why SaBN works. Codes are available at https://github.com/VITA-Group/Sandwich-Batch-Normalization.