Abstract:Pre-training followed by fine-tuning is widely adopted among practitioners. The performance can be improved by "model soups"~\cite{wortsman2022model} via exploring various hyperparameter configurations.The Learned-Soup, a variant of model soups, significantly improves the performance but suffers from substantial memory and time costs due to the requirements of (i) having to load all fine-tuned models simultaneously, and (ii) a large computational graph encompassing all fine-tuned models. In this paper, we propose Memory Efficient Hyperplane Learned Soup (MEHL-Soup) to tackle this issue by formulating the learned soup as a hyperplane optimization problem and introducing block coordinate gradient descent to learn the mixing coefficients. At each iteration, MEHL-Soup only needs to load a few fine-tuned models and build a computational graph with one combined model. We further extend MEHL-Soup to MEHL-Soup+ in a layer-wise manner. Experimental results on various ViT models and data sets show that MEHL-Soup(+) outperforms Learned-Soup(+) in terms of test accuracy, and also reduces memory usage by more than $13\times$. Moreover, MEHL-Soup(+) can be run on a single GPU and achieves $9\times$ speed up in soup construction compared with the Learned-Soup. The code is released at https://github.com/nblt/MEHL-Soup.
Abstract:Ridgeless regression has garnered attention among researchers, particularly in light of the ``Benign Overfitting'' phenomenon, where models interpolating noisy samples demonstrate robust generalization. However, kernel ridgeless regression does not always perform well due to the lack of flexibility. This paper enhances kernel ridgeless regression with Locally-Adaptive-Bandwidths (LAB) RBF kernels, incorporating kernel learning techniques to improve performance in both experiments and theory. For the first time, we demonstrate that functions learned from LAB RBF kernels belong to an integral space of Reproducible Kernel Hilbert Spaces (RKHSs). Despite the absence of explicit regularization in the proposed model, its optimization is equivalent to solving an $\ell_0$-regularized problem in the integral space of RKHSs, elucidating the origin of its generalization ability. Taking an approximation analysis viewpoint, we introduce an $l_q$-norm analysis technique (with $0<q<1$) to derive the learning rate for the proposed model under mild conditions. This result deepens our theoretical understanding, explaining that our algorithm's robust approximation ability arises from the large capacity of the integral space of RKHSs, while its generalization ability is ensured by sparsity, controlled by the number of support vectors. Experimental results on both synthetic and real datasets validate our theoretical conclusions.
Abstract:In the open world, detecting out-of-distribution (OOD) data, whose labels are disjoint with those of in-distribution (ID) samples, is important for reliable deep neural networks (DNNs). To achieve better detection performance, one type of approach proposes to fine-tune the model with auxiliary OOD datasets to amplify the difference between ID and OOD data through a separation loss defined on model outputs. However, none of these studies consider enlarging the feature disparity, which should be more effective compared to outputs. The main difficulty lies in the diversity of OOD samples, which makes it hard to describe their feature distribution, let alone design losses to separate them from ID features. In this paper, we neatly fence off the problem based on an aggregation property of ID features named Neural Collapse (NC). NC means that the penultimate features of ID samples within a class are nearly identical to the last layer weight of the corresponding class. Based on this property, we propose a simple but effective loss called OrthLoss, which binds the features of OOD data in a subspace orthogonal to the principal subspace of ID features formed by NC. In this way, the features of ID and OOD samples are separated by different dimensions. By optimizing the feature separation loss rather than purely enlarging output differences, our detection achieves SOTA performance on CIFAR benchmarks without any additional data augmentation or sampling, demonstrating the importance of feature separation in OOD detection. The code will be published.
Abstract:Machine unlearning (MU) aims to eliminate information that has been learned from specific training data, namely forgetting data, from a pre-trained model. Currently, the mainstream of existing MU methods involves modifying the forgetting data with incorrect labels and subsequently fine-tuning the model. While learning such incorrect information can indeed remove knowledge, the process is quite unnatural as the unlearning process undesirably reinforces the incorrect information and leads to over-forgetting. Towards more \textit{natural} machine unlearning, we inject correct information from the remaining data to the forgetting samples when changing their labels. Through pairing these adjusted samples with their labels, the model will tend to use the injected correct information and naturally suppress the information meant to be forgotten. Albeit straightforward, such a first step towards natural machine unlearning can significantly outperform current state-of-the-art approaches. In particular, our method substantially reduces the over-forgetting and leads to strong robustness to hyperparameters, making it a promising candidate for practical machine unlearning.
Abstract:Data imputation, the process of filling in missing feature elements for incomplete data sets, plays a crucial role in data-driven learning. A fundamental belief is that data imputation is helpful for learning performance, and it follows that the pursuit of better classification can guide the data imputation process. While some works consider using label information to assist in this task, their simplistic utilization of labels lacks flexibility and may rely on strict assumptions. In this paper, we propose a new framework that effectively leverages supervision information to complete missing data in a manner conducive to classification. Specifically, this framework operates in two stages. Firstly, it leverages labels to supervise the optimization of similarity relationships among data, represented by the kernel matrix, with the goal of enhancing classification accuracy. To mitigate overfitting that may occur during this process, a perturbation variable is introduced to improve the robustness of the framework. Secondly, the learned kernel matrix serves as additional supervision information to guide data imputation through regression, utilizing the block coordinate descent method. The superiority of the proposed method is evaluated on four real-world data sets by comparing it with state-of-the-art imputation methods. Remarkably, our algorithm significantly outperforms other methods when the data is missing more than 60\% of the features
Abstract:Random feature (RF) has been widely used for node consistency in decentralized kernel ridge regression (KRR). Currently, the consistency is guaranteed by imposing constraints on coefficients of features, necessitating that the random features on different nodes are identical. However, in many applications, data on different nodes varies significantly on the number or distribution, which calls for adaptive and data-dependent methods that generate different RFs. To tackle the essential difficulty, we propose a new decentralized KRR algorithm that pursues consensus on decision functions, which allows great flexibility and well adapts data on nodes. The convergence is rigorously given and the effectiveness is numerically verified: by capturing the characteristics of the data on each node, while maintaining the same communication costs as other methods, we achieved an average regression accuracy improvement of 25.5\% across six real-world data sets.
Abstract:Remote photoplethysmography (rPPG) technology has become increasingly popular due to its non-invasive monitoring of various physiological indicators, making it widely applicable in multimedia interaction, healthcare, and emotion analysis. Existing rPPG methods utilize multiple datasets for training to enhance the generalizability of models. However, they often overlook the underlying conflict issues across different datasets, such as (1) label conflict resulting from different phase delays between physiological signal labels and face videos at the instance level, and (2) attribute conflict stemming from distribution shifts caused by head movements, illumination changes, skin types, etc. To address this, we introduce the DOmain-HArmonious framework (DOHA). Specifically, we first propose a harmonious phase strategy to eliminate uncertain phase delays and preserve the temporal variation of physiological signals. Next, we design a harmonious hyperplane optimization that reduces irrelevant attribute shifts and encourages the model's optimization towards a global solution that fits more valid scenarios. Our experiments demonstrate that DOHA significantly improves the performance of existing methods under multiple protocols. Our code is available at https://github.com/SWY666/rPPG-DOHA.
Abstract:Improving the generalization ability of modern deep neural networks (DNNs) is a fundamental challenge in machine learning. Two branches of methods have been proposed to seek flat minima and improve generalization: one led by sharpness-aware minimization (SAM) minimizes the worst-case neighborhood loss through adversarial weight perturbation (AWP), and the other minimizes the expected Bayes objective with random weight perturbation (RWP). While RWP offers advantages in computation and is closely linked to AWP on a mathematical basis, its empirical performance has consistently lagged behind that of AWP. In this paper, we revisit the use of RWP for improving generalization and propose improvements from two perspectives: i) the trade-off between generalization and convergence and ii) the random perturbation generation. Through extensive experimental evaluations, we demonstrate that our enhanced RWP methods achieve greater efficiency in enhancing generalization, particularly in large-scale problems, while also offering comparable or even superior performance to SAM. The code is released at https://github.com/nblt/mARWP.
Abstract:Redundancy is a persistent challenge in Capsule Networks (CapsNet),leading to high computational costs and parameter counts. Although previous works have introduced pruning after the initial capsule layer, dynamic routing's fully connected nature and non-orthogonal weight matrices reintroduce redundancy in deeper layers. Besides, dynamic routing requires iterating to converge, further increasing computational demands. In this paper, we propose an Orthogonal Capsule Network (OrthCaps) to reduce redundancy, improve routing performance and decrease parameter counts. Firstly, an efficient pruned capsule layer is introduced to discard redundant capsules. Secondly, dynamic routing is replaced with orthogonal sparse attention routing, eliminating the need for iterations and fully connected structures. Lastly, weight matrices during routing are orthogonalized to sustain low capsule similarity, which is the first approach to introduce orthogonality into CapsNet as far as we know. Our experiments on baseline datasets affirm the efficiency and robustness of OrthCaps in classification tasks, in which ablation studies validate the criticality of each component. Remarkably, OrthCaps-Shallow outperforms other Capsule Network benchmarks on four datasets, utilizing only 110k parameters, which is a mere 1.25% of a standard Capsule Network's total. To the best of our knowledge, it achieves the smallest parameter count among existing Capsule Networks. Similarly, OrthCaps-Deep demonstrates competitive performance across four datasets, utilizing only 1.2% of the parameters required by its counterparts.
Abstract:Sharpness-Aware Minimization (SAM) has been instrumental in improving deep neural network training by minimizing both training loss and loss sharpness. Despite the practical success, the mechanisms behind SAM's generalization enhancements remain elusive, limiting its progress in deep learning optimization. In this work, we investigate SAM's core components for generalization improvement and introduce "Friendly-SAM" (F-SAM) to further enhance SAM's generalization. Our investigation reveals the key role of batch-specific stochastic gradient noise within the adversarial perturbation, i.e., the current minibatch gradient, which significantly influences SAM's generalization performance. By decomposing the adversarial perturbation in SAM into full gradient and stochastic gradient noise components, we discover that relying solely on the full gradient component degrades generalization while excluding it leads to improved performance. The possible reason lies in the full gradient component's increase in sharpness loss for the entire dataset, creating inconsistencies with the subsequent sharpness minimization step solely on the current minibatch data. Inspired by these insights, F-SAM aims to mitigate the negative effects of the full gradient component. It removes the full gradient estimated by an exponentially moving average (EMA) of historical stochastic gradients, and then leverages stochastic gradient noise for improved generalization. Moreover, we provide theoretical validation for the EMA approximation and prove the convergence of F-SAM on non-convex problems. Extensive experiments demonstrate the superior generalization performance and robustness of F-SAM over vanilla SAM. Code is available at https://github.com/nblt/F-SAM.