Foundation models are usually pre-trained on large-scale datasets and then adapted to different downstream tasks through tuning. This pre-training and then fine-tuning paradigm has become a standard practice in deep learning. However, the large-scale pre-training datasets, often inaccessible or too expensive to handle, can contain label noise that may adversely affect the generalization of the model and pose unexpected risks. This paper stands out as the first work to comprehensively understand and analyze the nature of noise in pre-training datasets and then effectively mitigate its impacts on downstream tasks. Specifically, through extensive experiments of fully-supervised and image-text contrastive pre-training on synthetic noisy ImageNet-1 K, YFCC15 M, and CC12 M datasets, we demonstrate that, while slight noise in pre-training can benefit in-domain (ID) performance, where the training and testing data share a similar distribution, it always deteriorates out-of-domain (OOD) performance, where training and testing distributions are significantly different. These observations are agnostic to scales of pre-training datasets, pre-training noise types, model architectures, pre-training objectives, downstream tuning methods, and downstream applications. We empirically ascertain that the reason behind this is that the pre-training noise shapes the feature space differently. We then propose a tuning method (NMTune) to affine the feature space to mitigate the malignant effect of noise and improve generalization, which is applicable in both parameter-efficient and black-box tuning manners, considering one may not be able to access or fully fine-tune the pre-trained models. We additionally conduct extensive experiments on popular vision and language models, including APIs, which are supervised and self-supervised pre-trained on realistic noisy data for evaluation. Our analysis and results demonstrate the importance of this novel and fundamental research direction, which we term as Noisy Model Transfer Learning.

Pre-training on large-scale datasets and then fine-tuning on downstream tasks have become a standard practice in deep learning. However, pre-training data often contain label noise that may adversely affect the generalization of the model. This paper aims to understand the nature of noise in pre-training datasets and to mitigate its impact on downstream tasks. More specifically, through extensive experiments of supervised pre-training models on synthetic noisy ImageNet-1K and YFCC15M datasets, we demonstrate that while slight noise in pre-training can benefit in-domain (ID) transfer performance, where the training and testing data share the same distribution, it always deteriorates out-of-domain (OOD) performance, where training and testing data distribution are different. We empirically verify that the reason behind is noise in pre-training shapes the feature space differently. We then propose a light-weight black-box tuning method (NMTune) to affine the feature space to mitigate the malignant effect of noise and improve generalization on both ID and OOD tasks, considering one may not be able to fully fine-tune or even access the pre-trained models. We conduct practical experiments on popular vision and language models that are pre-trained on noisy data for evaluation of our approach. Our analysis and results show the importance of this interesting and novel research direction, which we term Noisy Model Learning.

Learning with reduced labeling standards, such as noisy label, partial label, and supplementary unlabeled data, which we generically refer to as imprecise label, is a commonplace challenge in machine learning tasks. Previous methods tend to propose specific designs for every emerging imprecise label configuration, which is usually unsustainable when multiple configurations of imprecision coexist. In this paper, we introduce imprecise label learning (ILL), a framework for the unification of learning with various imprecise label configurations. ILL leverages expectation-maximization (EM) for modeling the imprecise label information, treating the precise labels as latent variables. Instead of approximating the correct labels for training, it considers the entire distribution of all possible labeling entailed by the imprecise information. We demonstrate that ILL can seamlessly adapt to partial label learning, semi-supervised learning, noisy label learning, and, more importantly, a mixture of these settings, with closed-form learning objectives derived from the unified EM modeling. Notably, ILL surpasses the existing specified techniques for handling imprecise labels, marking the first practical and unified framework with robust and effective performance across various challenging settings. We hope our work will inspire further research on this topic, unleashing the full potential of ILL in wider scenarios where precise labels are expensive and complicated to obtain.

It is well-known that Resolution proofs can be efficiently simulated by Sherali–Adams (SA) proofs. We show, however, that any such simulation needs to exploit huge coefficients: Resolution cannot be efficiently simulated by SA when the coefficients are written in unary. We also show that Reversible Resolution (a variant of MaxSAT Resolution) cannot be efficiently simulated by Nullstellensatz (NS). These results have consequences for total NP search problems. First, we characterise the classes PPADS, PPAD, SOPL by unary-SA, unary-NS, and Reversible Resolution, respectively. Second, we show that, relative to an oracle, \({\text{ PLS}} \not\subseteq {\text{ PPP}}\) , \({\text{ SOPL}} \not\subseteq {\text{ PPA}}\) , and \({\text{ EOPL}} \not\subseteq {\text{ UEOPL}}\) . In particular, together with prior work, this gives a complete picture of the black-box relationships between all classical TFNP classes introduced in the 1990s.

Abstract. We show [Formula: see text]. Here the class [Formula: see text] consists of all total search problems that reduce to the End-of-Potential-Line problem, which was introduced in the works by Hubáček and Yogev (SICOMP 2020) and Fearnley et al. (JCSS 2020). In particular, our result yields a new simpler proof of the breakthrough collapse [Formula: see text] by Fearnley et al. (STOC 2021). We also prove a companion result [Formula: see text], where [Formula: see text] is the class associated with the Sink-of-Potential-Line problem.

Few-Shot Learning (FSL) has been rapidly developed in recent years, potentially eliminating the requirement for significant data acquisition. Few-shot fine-tuning has been demonstrated to be practically efficient and helpful, especially for out-of-distribution datum [7, 13, 17, 29]. In this work, we first observe that the few-shot fine-tuned methods are learned with the imbalanced class marginal distribution, leading to imbalanced per-class testing accuracy. This observation further motivates us to propose the Transductive Fine-tuning with Margin-based uncertainty weighting and Probability regularization (TF-MP), which learns a more balanced class marginal distribution as shown in Fig. 1. We first conduct sample weighting on unlabeled testing data with margin-based uncertainty scores and fur-ther regularize each test sample's categorical probability. TF-MP achieves state-of-the-art performance on in-/out-of-distribution evaluations of Meta- Dataset [31] and sur-passes previous transductive methods by a large margin.

In recent works, utilizing a deep network trained on meta-training set serves as a strong baseline in few-shot learning. In this paper, we move forward to refine novel-class features by finetuning a trained deep network. Finetuning is designed to focus on reducing biases in novel-class feature distributions, which we define as two aspects: class-agnostic and class-specific biases. Class-agnostic bias is defined as the distribution shifting introduced by domain difference, which we propose Distribution Calibration Module(DCM) to reduce. DCM owes good property of eliminating domain difference and fast feature adaptation during optimization. Class-specific bias is defined as the biased estimation using a few samples in novel classes, which we propose Selected Sampling(SS) to reduce. Without inferring the actual class distribution, SS is designed by running sampling using proposal distributions around support-set samples. By powering finetuning with DCM and SS, we achieve state-of-the-art results on Meta-Dataset with consistent performance boosts over ten datasets from different domains. We believe our simple yet effective method demonstrates its possibility to be applied on practical few-shot applications.

We show $\textsf{EOPL}=\textsf{PLS}\cap\textsf{PPAD}$. Here the class $\textsf{EOPL}$ consists of all total search problems that reduce to the End-of-Potential-Line problem, which was introduced in the works by Hubacek and Yogev (SICOMP 2020) and Fearnley et al. (JCSS 2020). In particular, our result yields a new simpler proof of the breakthrough collapse $\textsf{CLS}=\textsf{PLS}\cap\textsf{PPAD}$ by Fearnley et al. (STOC 2021). We also prove a companion result $\textsf{SOPL}=\textsf{PLS}\cap\textsf{PPADS}$, where $\textsf{SOPL}$ is the class associated with the Sink-of-Potential-Line problem.

It is well-known that Resolution proofs can be efficiently simulated by Sherali-Adams (SA) proofs. We show1, however, that any such simulation needs to exploit huge coefficients: Resolution cannot be efficiently simulated by SA when the coefficients are written in unary. We also show that Reversible Resolution (a variant of MaxSAT Resolution) cannot be efficiently simulated by Nullstellensatz (NS). These results have consequences for total NP search problems. First, we characterise the classes PPADS, PPAD, SOPL by unary-SA, unary-NS, and Reversible Resolution, respectively. Second, we show that, relative to an oracle, PLS $\nsubseteq$ PPP, SOPL $\nsubseteq$ PPA, and EOPL $\nsubseteq$ UEOPL. In particular, together with prior work, this gives a complete picture of the black-box relationships between all classical TFNP classes introduced in the 1990s.1This is an extended abstract. For the full version of this article, please refer to [GHJ+22b].

Semi-supervised learning (SSL) improves model generalization by leveraging massive unlabeled data to augment limited labeled samples. However, currently, popular SSL evaluation protocols are often constrained to computer vision (CV) tasks. In addition, previous work typically trains deep neural networks from scratch, which is time-consuming and environmentally unfriendly. To address the above issues, we construct a Unified SSL Benchmark (USB) for classification by selecting 15 diverse, challenging, and comprehensive tasks from CV, natural language processing (NLP), and audio processing (Audio), on which we systematically evaluate the dominant SSL methods, and also open-source a modular and extensible codebase for fair evaluation of these SSL methods. We further provide the pre-trained versions of the state-of-the-art neural models for CV tasks to make the cost affordable for further tuning. USB enables the evaluation of a single SSL algorithm on more tasks from multiple domains but with less cost. Specifically, on a single NVIDIA V100, only 39 GPU days are required to evaluate FixMatch on 15 tasks in USB while 335 GPU days (279 GPU days on 4 CV datasets except for ImageNet) are needed on 5 CV tasks with TorchSSL.

We propose a method to disentangle linear-encoded facial semantics from StyleGAN without external supervision. The method derives from linear regression and sparse representation learning concepts to make the disentangled latent representations easily interpreted as well. We start by coupling StyleGAN with a stabilized 3D deformable facial reconstruction method to decompose single-view GAN generations into multiple semantics. Latent representations are then extracted to capture interpretable facial semantics. In this work, we make it possible to get rid of labels for disentangling meaningful facial semantics. Also, we demonstrate that the guided extrapolation along the disentangled representations can help with data augmentation, which sheds light on handling unbalanced data. Finally, we provide an analysis of our learned localized facial representations and illustrate that the semantic information is encoded, which surprisingly complies with human intuition. The overall unsupervised design brings more flexibility to representation learning in the wild.

Recently, impressive results have been achieved on many large-scale face recognition benchmarks, such as IJB-A Janus and Janus CS3. These datasets were designed to test robustness to nuisance transformations simultaneously such as pose, illumination, expression etc. We present a study paper, where we find that despite this goal in evaluation, there exists a significant frontal bias in yaw pose in these datasets. Therefore, high-performance on these recent datasets is misleading and does not reflect robustness to extreme pose in yaw. Moreover, many real-world applications only allow for a single frontal enrollment in a gallery (law enforcement, immigration etc.). As we show in our study, face recognition in this highly constrained setting with extreme pose variation in the probe images remains a highly challenging problem. Traditional approaches, performing well on datasets such as IJB-A Janus, perform much worse on older but highly controlled datasets such as CMU MPIE. To aid our study, we present a simple and practical method to handle pose variation in face recognition pipelines designed to deal with extremely off-angle faces. Our approach is to ignore the half of the face with any self-occlusion. This method allows our models to be highly robust to pose, and helps us achieve state-of-the-art results on several protocols using the CMU MPIE dataset as well as very accurate results on the CFP dataset, outperforming recent efforts using the same training data.

This paper introduces a novel anchor design principle to support anchor-based face detection for superior scaleinvariant performance, especially on tiny faces. To achieve this, we explicitly address the problem that anchor-based detectors drop performance drastically on faces with tiny sizes, e.g. less than 16 × 16 pixels. In this paper, we investigate why this is the case. We discover that current anchor design cannot guarantee high overlaps between tiny faces and anchor boxes, which increases the difficulty of training. The new Expected Max Overlapping (EMO) score is proposed which can theoretically explain the low overlapping issue and inspire several effective strategies of new anchor design leading to higher face overlaps, including anchor stride reduction with new network architectures, extra shifted anchors, and stochastic face shifting. Comprehensive experiments show that our proposed method significantly outperforms the baseline anchor-based detector, while consistently achieving state-of-the-art results on challenging face detection datasets with competitive runtime speed.