论文检索

Alireza Ganjdanesh, Shangqian Gao, Hirad Alipanah, Heng Huang

Generative Adversarial Networks (GANs) have shown remarkable success in modeling complex data distributions for image-to-image translation. Still, their high computational demands prohibit their deployment in practical scenarios like edge devices. Existing GAN compression methods mainly rely on knowledge distillation or convolutional classifiers' pruning techniques. Thus, they neglect the critical characteristic of GANs: their local density structure over their learned manifold. Accordingly, we approach GAN compression from a new perspective by explicitly encouraging the pruned model to preserve the density structure of the original parameter-heavy model on its learned manifold. We facilitate this objective for the pruned model by partitioning the learned manifold of the original generator into local neighborhoods around its generated samples. Then, we propose a novel pruning objective to regularize the pruned model to preserve the local density structure over each neighborhood, resembling the kernel density estimation method. Also, we develop a collaborative pruning scheme in which the discriminator and generator are pruned by two pruning agents. We design the agents to capture interactions between the generator and discriminator by exchanging their peer's feedback when determining corresponding models' architectures. Thanks to such a design, our pruning method can efficiently find performant sub-networks and can maintain the balance between the generator and discriminator more effectively compared to baselines during pruning, thereby showing more stable pruning dynamics. Our experiments on image translation GAN models, Pix2Pix and CycleGAN, with various benchmark datasets and architectures demonstrate our method's effectiveness.

Filippo Galli, Catuscia Palamidessi, Tommaso Cucinotta

Training differentially private machine learning models requires constraining an individual's contribution to the optimization process. This is achieved by clipping the 2-norm of their gradient at a predetermined threshold prior to averaging and batch sanitization. This selection adversely influences optimization in two opposing ways: it either exacerbates the bias due to excessive clipping at lower values, or augments sanitization noise at higher values. The choice significantly hinges on factors such as the dataset, model architecture, and even varies within the same optimization, demanding meticulous tuning usually accomplished through a grid search. In order to circumvent the privacy expenses incurred in hyperparameter tuning, we present a novel approach to dynamically optimize the clipping threshold. We treat this threshold as an additional learnable parameter, establishing a clean relationship between the threshold and the cost function. This allows us to optimize the former with gradient descent, with minimal repercussions on the overall privacy analysis. Our method is thoroughly assessed against alternative fixed and adaptive strategies across diverse datasets, tasks, model dimensions, and privacy levels. Our results indicate that it performs comparably or better in the evaluated scenarios, given the same privacy requirements.

Harmender Gahlawat, Meirav Zehavi

Decision trees is a fundamental tool in machine learning for representing, classifying, and generalizing data. It is desirable to construct ``small'' decision trees, by minimizing either the size (s) or the depth (d) of the decision tree (DT). Recently, the parameterized complexity of Decision Tree Learning has attracted a lot of attention. We consider a generalization of Decision Tree Learning where given a classification instance E and an integer t, the task is to find a ``small'' DT that disagrees with E in at most t examples. We consider two problems: DTSO and DTDO, where the goal is to construct a DT minimizing s and d, respectively. We first establish that both DTSO and DTDO are W[1]-hard when parameterized by s+y and d+y, respectively, where y is the maximum number of features in which two differently labeled examples can differ. We complement this result by showing that these problems become FPT if we include the parameter t. We also consider the kernelization complexity of these problems and establish several positive and negative results for both DTSO and DTDO.

Zhongtian Fu, Kefei Song, Luping Zhou, Yang Yang

Image captioning aims to automatically generate captions for images by learning a cross-modal generator from vision to language. The large amount of image-text pairs required for training is usually sourced from the internet due to the manual cost, which brings the noise with mismatched relevance that affects the learning process. Unlike traditional noisy label learning, the key challenge in processing noisy image-text pairs is to finely identify the mismatched words to make the most use of trustworthy information in the text, rather than coarsely weighing the entire examples. To tackle this challenge, we propose a Noise-aware Image Captioning method (NIC) to adaptively mitigate the erroneous guidance from noise by progressively exploring mismatched words. Specifically, NIC first identifies mismatched words by quantifying word-label reliability from two aspects: 1) inter-modal representativeness, which measures the significance of the current word by assessing cross-modal correlation via prediction certainty; 2) intra-modal informativeness, which amplifies the effect of current prediction by combining the quality of subsequent word generation. During optimization, NIC constructs the pseudo-word-labels considering the reliability of the origin word-labels and model convergence to periodically coordinate mismatched words. As a result, NIC can effectively exploit both clean and noisy image-text pairs to learn a more robust mapping function. Extensive experiments conducted on the MS-COCO and Conceptual Caption datasets validate the effectiveness of our method in various noisy scenarios.

Minghao Fu, Ke Zhu, Jianxin Wu

When pre-trained models become rapidly larger, the cost of fine-tuning on downstream tasks steadily increases, too. To economically fine-tune these models, parameter-efficient transfer learning (PETL) is proposed, which only tunes a tiny subset of trainable parameters to efficiently learn quality representations. However, current PETL methods are facing the dilemma that during training the GPU memory footprint is not effectively reduced as trainable parameters. PETL will likely fail, too, if the full fine-tuning encounters the out-of-GPU-memory issue. This phenomenon happens because trainable parameters from these methods are generally entangled with the backbone, such that a lot of intermediate states have to be stored in GPU memory for gradient propagation. To alleviate this problem, we introduce Disentangled Transfer Learning (DTL), which disentangles the trainable parameters from the backbone using a lightweight Compact Side Network (CSN). By progressively extracting task-specific information with a few low-rank linear mappings and appropriately adding the information back to the backbone, CSN effectively realizes knowledge transfer in various downstream tasks. We conducted extensive experiments to validate the effectiveness of our method. The proposed method not only reduces a large amount of GPU memory usage and trainable parameters, but also outperforms existing PETL methods by a significant margin in accuracy, achieving new state-of-the-art on several standard benchmarks.

Minghan Fu, Fang-Xiang Wu

The learning rate is a critical hyperparameter for deep learning tasks since it determines the extent to which the model parameters are adjusted during the learning course. However, the choice of learning rates typically depends on empirical judgment, which may not result in satisfactory outcomes without intensive try-and-error experiments. In this study, we propose a novel learning rate adaptation scheme called QLABGrad. Without any user-specified hyperparameter, QLABGrad automatically determines the learning rate by optimizing the quadratic loss approximation-based (QLAB) function for a given gradient descent direction, where only one extra forward propagation is required. We theoretically prove the convergence of QLABGrad under the smooth Lipschitz condition on the loss function. Experiment results on multiple architectures, including MLP, CNN, and ResNet, on MNIST, CIFAR10, and ImageNet datasets, demonstrate that QLABGrad outperforms widely adopted schemes for deep learning.

Feisi Fu, Zhilu Wang, Weichao Zhou, Yixuan Wang, Jiameng Fan, Chao Huang, Qi Zhu, Xin Chen, Wenchao Li

We present REGLO, a novel methodology for repairing pretrained neural networks to satisfy global robustness and individual fairness properties. A neural network is said to be globally robust with respect to a given input region if and only if all the input points in the region are locally robust. This notion of global robustness also captures the notion of individual fairness as a special case. We prove that any counterexample to a global robustness property must exhibit a corresponding large gradient. For ReLU networks, this result allows us to efficiently identify the linear regions that violate a given global robustness property. By formulating and solving a suitable robust convex optimization problem, REGLO then computes a minimal weight change that will provably repair these violating linear regions.

Fares Fourati, Christopher John Quinn, Mohamed-Slim Alouini, Vaneet Aggarwal

We propose a novel combinatorial stochastic-greedy bandit (SGB) algorithm for combinatorial multi-armed bandit problems when no extra information other than the joint reward of the selected set of n arms at each time step t in [T] is observed. SGB adopts an optimized stochastic-explore-then-commit approach and is specifically designed for scenarios with a large set of base arms. Unlike existing methods that explore the entire set of unselected base arms during each selection step, our SGB algorithm samples only an optimized proportion of unselected arms and selects actions from this subset. We prove that our algorithm achieves a (1-1/e)-regret bound of O(n^(1/3) k^(2/3) T^(2/3) log(T)^(2/3)) for monotone stochastic submodular rewards, which outperforms the state-of-the-art in terms of the cardinality constraint k. Furthermore, we empirically evaluate the performance of our algorithm in the context of online constrained social influence maximization. Our results demonstrate that our proposed approach consistently outperforms the other algorithms, increasing the performance gap as k grows.

Jack Foster, Stefan Schoepf, Alexandra Brintrup

Machine unlearning, the ability for a machine learning model to forget, is becoming increasingly important to comply with data privacy regulations, as well as to remove harmful, manipulated, or outdated information. The key challenge lies in forgetting specific information while protecting model performance on the remaining data. While current state-of-the-art methods perform well, they typically require some level of retraining over the retained data, in order to protect or restore model performance. This adds computational overhead and mandates that the training data remain available and accessible, which may not be feasible. In contrast, other methods employ a retrain-free paradigm, however, these approaches are prohibitively computationally expensive and do not perform on par with their retrain-based counterparts. We present Selective Synaptic Dampening (SSD), a novel two-step, post hoc, retrain-free approach to machine unlearning which is fast, performant, and does not require long-term storage of the training data. First, SSD uses the Fisher information matrix of the training and forgetting data to select parameters that are disproportionately important to the forget set. Second, SSD induces forgetting by dampening these parameters proportional to their relative importance to the forget set with respect to the wider training data. We evaluate our method against several existing unlearning methods in a range of experiments using ResNet18 and Vision Transformer. Results show that the performance of SSD is competitive with retrain-based post hoc methods, demonstrating the viability of retrain-free post hoc unlearning approaches.

Kei Sen Fong, Mehul Motani

We introduce a conceptually simple yet effective method to create small, compact decision trees - by using splits found via Symbolic Regression (SR). Traditional decision tree (DT) algorithms partition a dataset on axis-parallel splits. When the true boundaries are not along the feature axes, DT is likely to have a complicated structure and a dense decision boundary. In this paper, we introduce SR-Enhanced DT (SREDT) - a method which utilizes SR to increase the richness of the class of possible DT splits. We evaluate SREDT on both synthetic and real-world datasets. Despite its simplicity, our method produces surprisingly small trees that outperform both DT and oblique DT (ODT) on supervised classification tasks in terms of accuracy and F-score. We show empirically that SREDTs decrease inference time (compared to DT and ODT) and argue that they allow us to obtain more explainable descriptions of the decision process. SREDT also performs competitively against state-of-the-art tabular classification methods, including tree ensembles and deep models. Finally, we introduce a local search mechanism to improve SREDT and evaluate it on 56 PMLB datasets. This mechanism shows improved performance on 77.2% of the datasets, outperforming DT and ODT. In terms of F-Score, local SREDT outperforms DT and ODT in 82.5% and 73.7% of the datasets respectively and in terms of inference time, local SREDT requires 25.8% and 26.6% less inference time than DT and ODT respectively.

Manon Flageat, Bryan Lim, Antoine Cully

Many applications in Reinforcement Learning (RL) usually have noise or stochasticity present in the environment. Beyond their impact on learning, these uncertainties lead the exact same policy to perform differently, i.e. yield different return, from one roll-out to another. Common evaluation procedures in RL summarise the consequent return distributions using solely the expected return, which does not account for the spread of the distribution. Our work defines this spread as the policy reproducibility: the ability of a policy to obtain similar performance when rolled out many times, a crucial property in some real-world applications. We highlight that existing procedures that only use the expected return are limited on two fronts: first an infinite number of return distributions with a wide range of performance-reproducibility trade-offs can have the same expected return, limiting its effectiveness when used for comparing policies; second, the expected return metric does not leave any room for practitioners to choose the best trade-off value for considered applications. In this work, we address these limitations by recommending the use of Lower Confidence Bound, a metric taken from Bayesian optimisation that provides the user with a preference parameter to choose a desired performance-reproducibility trade-off. We also formalise and quantify policy reproducibility, and demonstrate the benefit of our metrics using extensive experiments of popular RL algorithms on common uncertain RL tasks.

Dana Fisman, Noa Izsak, Swen Jacobs

The problem of learning a computational model from examples has been receiving growing attention. For the particularly challenging problem of learning models of distributed systems, existing results are restricted to models with a fixed number of interacting processes. In this work we look for the first time (to the best of our knowledge) at the problem of learning a distributed system with an arbitrary number of processes, assuming only that there exists a cutoff, i.e., a number of processes that is sufficient to produce all observable behaviors. Specifically, we consider fine broadcast protocols, these are broadcast protocols (BPs) with a finite cutoff and no hidden states. We provide a learning algorithm that can infer a correct BP from a sample that is consistent with a fine BP, and a minimal equivalent BP if the sample is sufficiently complete. On the negative side we show that (a) characteristic sets of exponential size are unavoidable, (b) the consistency problem for fine BPs is NP hard, and (c) that fine BPs are not polynomially predictable.

Stefano Fiorini, Stefano Coniglio, Michele Ciavotta, Enza Messina

We introduce QuaterGCN, a spectral Graph Convolutional Network (GCN) with quaternion-valued weights at whose core lies the Quaternionic Laplacian, a quaternion-valued Laplacian matrix by whose proposal we generalize two widely-used Laplacian matrices: the classical Laplacian (defined for undirected graphs) and the complex-valued Sign-Magnetic Laplacian (proposed within the spectral GCN SigMaNet to handle digraphs with weights of arbitrary sign). In addition to its generality, QuaterGCN is the only Laplacian to completely preserve the (di)graph topology that we are aware of, as it can handle graphs and digraphs containing antiparallel pairs of edges (digons) of different weight without reducing them to a single (directed or undirected) edge as done by other Laplacians. Experimental results show the superior performance of QuaterGCN compared to other state-of-the-art GCNs, particularly in scenarios where the information the digons carry is crucial to successfully address the task at hand.

Jan Finkbeiner, Thomas Gmeinder, Mark Pupilli, Alexander Titterton, Emre Neftci

Current AI training infrastructure is dominated by single instruction multiple data (SIMD) and systolic array architectures, such as Graphics Processing Units (GPUs) and Tensor Processing Units (TPUs), that excel at accelerating parallel workloads and dense vector matrix multiplications. Potentially more efficient neural network models utilizing sparsity and recurrence cannot leverage the full power of SIMD processor and are thus at a severe disadvantage compared to today's prominent parallel architectures like Transformers and CNNs, thereby hindering the path towards more sustainable AI. To overcome this limitation, we explore sparse and recurrent model training on a massively parallel multiple instruction multiple data (MIMD) architecture with distributed local memory. We implement a training routine based on backpropagation though time (BPTT) for the brain-inspired class of Spiking Neural Networks (SNNs) that feature binary sparse activations. We observe a massive advantage in using sparse activation tensors with a MIMD processor, the Intelligence Processing Unit (IPU) compared to GPUs. On training workloads, our results demonstrate 5-10x throughput gains compared to A100 GPUs and up to 38x gains for higher levels of activation sparsity, without a significant slowdown in training convergence or reduction in final model performance. Furthermore, our results show highly promising trends for both single and multi IPU configurations as we scale up to larger model sizes. Our work paves the way towards more efficient, non-standard models via AI training hardware beyond GPUs, and competitive large scale SNN models.

Wei Feng, Guoshuai Sheng, Qianqian Wang, Quanxue Gao, Zhiqiang Tao, Bo Dong

Partial multi-view clustering is a challenging and practical research problem for data analysis in real-world applications, due to the potential data missing issue in different views. However, most existing methods have not fully explored the correlation information among various incomplete views. In addition, these existing clustering methods always ignore discovering discriminative features inside the data itself in this unsupervised task. To tackle these challenges, we propose Partial Multi-View Clustering via Self-Supervised \textbf{N}etwork (PVC-SSN) in this paper. Specifically, we employ contrastive learning to obtain a more discriminative and consistent subspace representation, which is guided by a self-supervised module. Self-supervised learning can exploit effective cluster information through the data itself to guide the learning process of clustering tasks. Thus, it can pull together embedding features from the same cluster and push apart these from different clusters. Extensive experiments on several benchmark datasets show that the proposed PVC-SCN method outperforms several state-of-the-art clustering methods.

Shibo Feng, Chunyan Miao, Zhong Zhang, Peilin Zhao

The probability prediction of multivariate time series is a notoriously challenging but practical task. This research proposes to condense high-dimensional multivariate time series forecasting into a problem of latent space time series generation, to improve the expressiveness of each timestamp and make forecasting more manageable. To solve the problem that the existing work is hard to extend to high-dimensional multivariate time series, we present a latent multivariate time series diffusion framework called Latent Diffusion Transformer (LDT), which consists of a symmetric statistics-aware autoencoder and a diffusion-based conditional generator, to implement this idea. Through careful design, the time series autoencoder can compress multivariate timestamp patterns into a concise latent representation by considering dynamic statistics. Then, the diffusion-based conditional generator is able to efficiently generate realistic multivariate timestamp values on a continuous latent space under a novel self-conditioning guidance which is modeled in a non-autoregressive way. Extensive experiments demonstrate that our model achieves state-of-the-art performance on many popular high-dimensional multivariate time series datasets.

Qianhan Feng, Lujing Xie, Shijie Fang, Tong Lin

Semi-supervised Learning (SSL) reduces the need for extensive annotations in deep learning, but the more realistic challenge of imbalanced data distribution in SSL remains largely unexplored. In Class Imbalanced Semi-supervised Learning (CISSL), the bias introduced by unreliable pseudo-labels can be exacerbated by imbalanced data distributions. Most existing methods address this issue at instance-level through reweighting or resampling, but the performance is heavily limited by their reliance on biased backbone representation. Some other methods do perform feature-level adjustments like feature blending but might introduce unfavorable noise. In this paper, we discuss the bonus of a more balanced feature distribution for the CISSL problem, and further propose a Balanced Feature-Level Contrastive Learning method (BaCon). Our method directly regularizes the distribution of instances' representations in a well-designed contrastive manner. Specifically, class-wise feature centers are computed as the positive anchors, while negative anchors are selected by a straightforward yet effective mechanism. A distribution-related temperature adjustment is leveraged to control the class-wise contrastive degrees dynamically. Our method demonstrates its effectiveness through comprehensive experiments on the CIFAR10-LT, CIFAR100-LT, STL10-LT, and SVHN-LT datasets across various settings. For example, BaCon surpasses instance-level method FixMatch-based ABC on CIFAR10-LT with a 1.21% accuracy improvement, and outperforms state-of-the-art feature-level method CoSSL on CIFAR100-LT with a 0.63% accuracy improvement. When encountering more extreme imbalance degree, BaCon also shows better robustness than other methods.

Jiaheng Feng, Mingxiao Feng, Haolin Song, Wengang Zhou, Houqiang Li

Offline-to-online reinforcement learning (RL) provides a promising solution to improving suboptimal offline pre-trained policies through online fine-tuning. However, one efficient method, unconstrained fine-tuning, often suffers from severe policy collapse due to excessive distribution shift. To ensure stability, existing methods retain offline constraints and employ additional techniques during fine-tuning, which hurts efficiency. In this work, we introduce a novel perspective: eliminating the policy collapse without imposing constraints. We observe that such policy collapse arises from the mismatch between unconstrained fine-tuning and the conventional RL training framework. To this end, we propose Stabilized Unconstrained Fine-tuning (SUF), a streamlined framework that benefits from the efficiency of unconstrained fine-tuning while ensuring stability by modifying the Update-To-Data ratio. With just a few lines of code adjustments, SUF demonstrates remarkable adaptability to diverse backbones and superior performance over state-of-the-art baselines.

Yujie Fang, Xin Li, Qianyu Chen, Mingzhong Wang

The widespread adoption of Graph Neural Networks (GNNs) has led to an increasing focus on their reliability. To address the issue of underconfidence in GNNs, various calibration methods have been developed to gain notable reductions in calibration error. However, we observe that existing approaches generally fail to enhance consistently, and in some cases even deteriorate, GNNs' ability to discriminate between correct and incorrect predictions. In this study, we advocate the significance of discriminative ability and the inclusion of relevant evaluation metrics. Our rationale is twofold: 1) Overlooking discriminative ability can inadvertently compromise the overall quality of the model; 2) Leveraging discriminative ability can significantly inform and improve calibration outcomes. Therefore, we thoroughly explore the reasons why existing calibration methods have ineffectiveness and even degradation regarding the discriminative ability of GNNs. Building upon these insights, we conduct GNN calibration experiments across multiple datasets using a straightforward example model, denoted as DC(GNN). Its excellent performance confirms the potential of integrating discriminative ability as a key consideration in the calibration of GNNs, thereby establishing a pathway toward more effective and reliable network calibration.

Junpeng Fang, Gongduo Zhang, Qing Cui, Caizhi Tang, Lihong Gu, Longfei Li, Jinjie Gu, Jun Zhou

Accurate prediction of coupon usage is crucial for promoting user consumption through targeted coupon recommendations. However, in real-world coupon recommendations, the coupon allocation process is not solely determined by the model trained with the history interaction data but is also interfered with by marketing tactics desired to fulfill specific commercial goals.This interference creates an imbalance in the interactions, which causes the data to deviate from the user's natural preferences. We refer to this deviation as the matching bias. Such biased interaction data affects the efficacy of the model, and thus it is necessary to employ debiasing techniques to prevent any negative impact. We investigate the mitigation of matching bias in coupon recommendations from a causal-effect perspective. By treating the attributes of users and coupons associated with marketing tactics as confounders, we find the confounders open the backdoor path between user-coupon matching and the conversion, which introduces spurious correlation. To remove the bad effect, we propose a novel training paradigm named Backdoor Adjustment via Group Adaptation (BAGA) for debiased coupon recommendations, which performs intervened training and inference, i.e., separately modeling each user-coupon group pair. However, modeling all possible group pairs greatly increases the computational complexity and cost. To address the efficiency challenge, we further present a simple but effective dual-tower multi-task framework and leverage the Customized Gate Control (CGC) model architecture, which separately models each user and coupon group with a separate expert module. We instantiate BAGA on five representative models: FM, DNN, NCF, MASKNET, and DEEPFM, and conduct comprehensive offline and online experiments to demonstrate the efficacy of our proposed paradigm.