论文检索

Nannan Wu, Zhuo Kuang, Zengqiang Yan, Li Yu

Due to escalating privacy concerns, federated learning has been recognized as a vital approach for training deep neural networks with decentralized medical data. In practice, it is challenging to ensure consistent imaging quality across various institutions, often attributed to equipment malfunctions affecting a minority of clients. This imbalance in image quality can cause the federated model to develop an inherent bias towards higher-quality images, thus posing a severe fairness issue. In this study, we pioneer the identification and formulation of this new fairness challenge within the context of the imaging quality shift. Traditional methods for promoting fairness in federated learning predominantly focus on balancing empirical risks across diverse client distributions. This strategy primarily facilitates fair optimization across different training data distributions, yet neglects the crucial aspect of generalization. To address this, we introduce a solution termed Federated learning with Inter-client Sharpness Matching (FedISM). FedISM enhances both local training and global aggregation by incorporating sharpness-awareness, aiming to harmonize the sharpness levels across clients for fair generalization. Our empirical evaluations, conducted using the widely-used ICH and ISIC 2019 datasets, establish FedISM's superiority over current state-of-the-art federated learning methods in promoting fairness. Code is available at https://github.com/wnn2000/FFL4MIA.

Xin Liu, Wei Tao, Wei Li, Dazhi Zhan, Jun Wang, Zhisong Pan

Due to its simplicity and efficiency, the first-order gradient method has been extensively employed in training neural networks. Although the optimization problem of the neural network is non-convex, recent research has proved that the first-order method is capable of attaining a global minimum during training over-parameterized neural networks, where the number of parameters is significantly larger than that of training instances. Momentum methods, including the heavy ball (HB) method and Nesterov's accelerated gradient (NAG) method, are the workhorse of first-order gradient methods owning to their accelerated convergence. In practice, NAG often exhibits superior performance than HB. However, current theoretical works fail to distinguish their convergence difference in training neural networks. To fill this gap, we consider the training problem of the two-layer ReLU neural network under over-parameterization and random initialization. Leveraging high-resolution dynamical systems and neural tangent kernel (NTK) theory, our result not only establishes tighter upper bounds of the convergence rate for both HB and NAG, but also provides the first theoretical guarantee for the acceleration of NAG over HB in training neural networks. Finally, we validate our theoretical results on three benchmark datasets.

Emanuele La Malfa, Gabriele La Malfa, Giuseppe Nicosia, Vito Latora

Deep Neural Networks (DNNs) can be represented as graphs whose links and vertices iteratively process data and solve tasks sub-optimally. Complex Network Theory (CNT), merging statistical physics with graph theory, provides a method for interpreting neural networks by analysing their weights and neuron structures. However, classic works adapt CNT metrics that only permit a topological analysis as they do not account for the effect of the input data. In addition, CNT metrics have been applied to a limited range of architectures, mainly including Fully Connected neural networks. In this work, we extend the existing CNT metrics with measures that sample from the DNNs' training distribution, shifting from a purely topological analysis to one that connects with the interpretability of deep learning. For the novel metrics, in addition to the existing ones, we provide a mathematical formalisation for Fully Connected, AutoEncoder, Convolutional and Recurrent neural networks, of which we vary the activation functions and the number of hidden layers. We show that these metrics differentiate DNNs based on the architecture, the number of hidden layers, and the activation function. Our contribution provides a method rooted in physics for interpreting DNNs that offers insights beyond the traditional input-output relationship and the CNT topological analysis.

Tianxiao Cao, Lu Sun, Canh Hao Nguyen, Hiroshi Mamitsuka

Compressing deep neural networks is of great importance for real-world applications on resource-constrained devices. Tensor decomposition is one promising answer that retains the functionality and most of the expressive power of the original deep models by replacing the weights with their decomposed cores. Decomposition with optimal ranks can achieve a good compression-accuracy trade-off, but it is expensive to optimize due to its discrete and combinatorial nature. A common practice is to set all ranks equal and tune one hyperparameter, but it may significantly harm the flexibility and generalization. In this paper, we propose a novel automatic rank selection method for deep model compression that allows learning model weights and decomposition ranks simultaneously. We propose to penalize the ℓ0 (quasi-)norm of the slices of decomposed tensor cores during model training. To avoid combinatorial optimization, we develop a probabilistic formulation and apply an approximate Bernoulli gate to each of the slices of tensor cores, which can be implemented in an end-to-end and scalable framework via gradient descent. It enables the automatic rank selection to be incorporated with arbitrary tensor decompositions and neural network layers such as linear layers, convolutional layers, and embedding layers. Comprehensive experiments on various tasks, including image classification, text sentiment classification, and neural machine translation, demonstrate the superior effectiveness of the proposed method over baselines.

Akhilan Boopathy, William Yue, Jaedong Hwang, Abhiram Iyer, Ila Fiete

Much research in machine learning involves finding appropriate inductive biases (e.g. convolutional neural networks, momentum-based optimizers, transformers) to promote generalization on tasks. However, quantification of the amount of inductive bias associated with these architectures and hyperparameters has been limited. We propose a novel method for efficiently computing the inductive bias required for generalization on a task with a fixed training data budget; formally, this corresponds to the amount of information required to specify well-generalizing models within a specific hypothesis space of models. Our approach involves modeling the loss distribution of random hypotheses drawn from a hypothesis space to estimate the required inductive bias for a task relative to these hypotheses. Unlike prior work, our method provides a direct estimate of inductive bias without using bounds and is applicable to diverse hypothesis spaces. Moreover, we derive approximation error bounds for our estimation approach in terms of the number of sampled hypotheses. Consistent with prior results, our empirical results demonstrate that higher dimensional tasks require greater inductive bias. We show that relative to other expressive model classes, neural networks as a model class encode large amounts of inductive bias. Furthermore, our measure quantifies the relative difference in inductive bias between different neural network architectures. Our proposed inductive bias metric provides an information-theoretic interpretation of the benefits of specific model architectures for certain tasks and provides a quantitative guide to developing tasks requiring greater inductive bias, thereby encouraging the development of more powerful inductive biases.

Qianhui Liu, Jiaqi Yan, Malu Zhang, Gang Pan, Haizhou Li

Spiking Neural Networks (SNNs) mimic the information-processing mechanisms of the human brain and are highly energy-efficient, making them well-suited for low-power edge devices. However, the pursuit of accuracy in current studies leads to large, long-timestep SNNs, conflicting with the resource constraints of these devices. In order to design lightweight and efficient SNNs, we propose a new approach named LitE-SNN that incorporates both spatial and temporal compression into the automated network design process. Spatially, we present a novel Compressive Convolution block (CompConv) to expand the search space to support pruning and mixed-precision quantization. Temporally, we are the first to propose a compressive timestep search to identify the optimal number of timesteps under specific computation cost constraints. Finally, we formulate a joint optimization to simultaneously learn the architecture parameters and spatial-temporal compression strategies to achieve high performance while minimizing memory and computation costs. Experimental results on CIFAR-10, CIFAR-100, and Google Speech Command datasets demonstrate our proposed LitE-SNNs can achieve competitive or even higher accuracy with remarkably smaller model sizes and fewer computation costs.

Zehua Liu, Jingyuan Wang, Zimeng Li, Yue He

Due to the complex and dynamic traffic contexts, the interpretability and uncertainty of traffic forecasting have gained increasing attention. Significance testing is a powerful tool in statistics used to determine whether a hypothesis is valid, facilitating the identification of pivotal features that predominantly contribute to the true relationship. However, existing works mainly regard traffic forecasting as a deterministic problem, making it challenging to perform effective significance testing. To fill this gap, we propose to conduct Full Bayesian Significance Testing for Neural Networks in Traffic Forecasting, namely ST-nFBST. A Bayesian neural network is utilized to capture the complicated traffic relationships through an optimization function resolved in the context of aleatoric uncertainty and epistemic uncertainty. Thereupon, ST-nFBST can achieve the significance testing by means of a delicate grad-based evidence value, further capturing the inherent traffic schema for better spatiotemporal modeling. Extensive experiments are conducted on METR-LA and PEMS-BAY to verify the advantages of our method in terms of uncertainty analysis and significance testing, helping the interpretability and promotion of traffic forecasting.

Lu Jiang, Yanan Xiao, Xinxin Zhao, Yuanbo Xu, Shuli Hu, Pengyang Wang, Minghao Yin

With the widespread popularity of massive open online courses, personalized course recommendation has become increasingly important due to enhancing users' learning efficiency. While achieving promising performances, current works suffering from the vary across the users and other MOOC entities. To address this problem, we propose hierarchical reinforcement learning with a multi-channel hypergraphs neural network for course recommendation(called HHCoR). Specifically, we first construct an online course hypergraph as the environment to capture the complex relationships and historical information by considering all entities. Then, we design a multi-channel propagation mechanism to aggregate embeddings in the online course hypergraph and extract user interest through an attention layer. Besides, we employ two-level decision-making: the low-level focuses on the rating courses, while the high-level integrates these considerations to finalize the decision. Furthermore, in co-optimization, we design a joint reward function to improve the policy of two-layer agents. Finally, we conducted extensive experiments on two real-world datasets and the quantitative results have demonstrated the effectiveness of the proposed method.

Wei Wang, Hanyang Li, Ke Shi, Chao Huang, Yang Cao, Cong Wang, Xiaochun Cao

Domain adaptation aims to achieve label transfer from a labeled source domain to an unlabeled target domain, where the two domains exhibit different distributions. Existing methods primarily concentrate on designing a feature extractor to learn better domain-invariant features, along with developing an effective classifier for reliable predictions. In this paper, we introduce optimal graph learning to generate a cross-domain graph that effectively connects the two domains, and two domain-specific graphs to capture domain-specific structures. On the one hand, we incorporate the three graphs into the label propagation (LP) classifier to enhance its robustness to distribution difference. On the other hand, we leverage the three graphs to introduce graph embedding losses, promoting the learning of locally discriminative and domain-invariant features. Furthermore, we maximize the nuclear norm of predictions in LP to enhance class diversity, thereby improving its robustness to class imbalance problem. Correspondingly, we develop an efficient algorithm to solve the associated optimization problem. Finally, we integrate the proposed LP and graph embedding losses into a deep neural network, resulting in our proposed deep cross-domain robust LP. Extensive experiments conducted on three cross-domain benchmark datasets demonstrate that our proposed approach could outperform existing state-of-the-art domain adaptation methods.

Tianyu Song, Guiyue Jin, Pengpeng Li, Kui Jiang, Xiang Chen, Jiyu Jin

Recently, spiking neural networks (SNNs) have demonstrated substantial potential in computer vision tasks. In this paper, we present an Efficient Spiking Deraining Network, called ESDNet. Our work is motivated by the observation that rain pixel values will lead to a more pronounced intensity of spike signals in SNNs. However, directly applying deep SNNs to image deraining task still remains a significant challenge. This is attributed to the information loss and training difficulties that arise from discrete binary activation and complex spatiotemporal dynamics. To this end, we develop a spiking residual block to convert the input into spike signals, then adaptively optimize the membrane potential by introducing attention weights to adjust spike responses in a data-driven manner, alleviating information loss caused by discrete binary activation. By this way, our ESDNet can effectively detect and analyze the characteristics of rain streaks by learning their fluctuations. This also enables better guidance for the deraining process and facilitates high-quality image reconstruction. Instead of relying on the ANN-SNN conversion strategy, we introduce a gradient proxy strategy to directly train the model for overcoming the challenge of training. Experimental results show that our approach gains comparable performance against ANN-based methods while reducing energy consumption by 54%. The code source is available at https://github.com/MingTian99/ESDNet.

Xinrun Wang, Chang Yang, Shuxin Li, Pengdeng Li, Xiao Huang, Hau Chan, Bo An

Nash Equilibrium (NE) is the canonical solution concept of game theory, which provides an elegant tool to understand the rationalities. Though mixed strategy NE exists in any game with finite players and actions, computing NE in two- or multi-player general-sum games is PPAD-Complete. Various alternative solutions, e.g., Correlated Equilibrium (CE), and learning methods, e.g., fictitious play (FP), are proposed to approximate NE. For convenience, we call these methods as ``inexact solvers'', or ``solvers'' for short. However, the alternative solutions differ from NE and the learning methods generally fail to converge to NE. Therefore, in this work, we propose REinforcement Nash Equilibrium Solver (RENES), which trains a single policy to modify the games with different sizes and applies the solvers on the modified games where the obtained solution is evaluated on the original games. Specifically, our contributions are threefold. i) We represent the games as alpha-rank response graphs and leverage graph neural network (GNN) to handle the games with different sizes as inputs; ii) We use tensor decomposition, e.g., canonical polyadic (CP), to make the dimension of modifying actions fixed for games with different sizes; iii) We train the modifying strategy for games with the widely-used proximal policy optimization (PPO) and apply the solvers to solve the modified games, where the obtained solution is evaluated on original games. Extensive experiments on large-scale normal-form games show that our method can further improve the approximation of NE of different solvers, i.e., alpha-rank, CE, FP and PRD, and can be generalized to unseen games.

Adrien Benamira, Thomas Peyrin, Trevor Yap, Tristan Guérand, Bryan Hooi

We introduce "Truth Table net"' (TTnet), a novel Deep Neural Network (DNN) architecture designed to provide excellent scalability/compactness trade-offs among DNNs, allowing in turn to tackle the DNN challenge of fast formal verification. TTnet is constructed using Learning Truth Table (LTT) filters, analogous to how a Deep Convolutional Neural Network (DCNN) is built upon convolutional filters. The differentiable LTT filters are unique by their dual form: they are both a neural network-based function and a small-sized truth table that can be computed within a practical time frame. This characteristic guarantees, by design and independently of the overall architecture, the ability to practically extract an efficient (in terms of the number of logical gates) and functionally equivalent Conjunctive Normal Form (CNF) Boolean logic gate implementation. This CNF circuit is even optimal when the LTT truth table's input bit size n < 12. In particular, TTnet architecture is the first differentiable DNN with as dual form a compact logic gate representation that can scale to datasets larger than CIFAR-10: we achieve an accuracy of 41% on the ImageNet dataset while ensuring that each LTT filter truth table is fully computable within 2^{16} operations. We further compare the compactness and scalability performances of TTnet Boolean logic circuit representation to state-of-the-art differentiable logic DNNs across tabular, MNIST, and CIFAR-10 datasets. We emphasize that TTnet is the first solution to the open problem of designing differentiable convolutional neural networks with an exact dual logic gate circuit representation, bridging the gap between symbolic AI and trainable DCNNs. Finally, as improving DNNs compactness in Boolean logic circuit form reduces the complexity of their formal verification, we demonstrate TTnet effectiveness in exact sound and complete formal verification. Notably, our model achieves robustness verification in 10ms vs 100s for traditional state-of-the-art DNNs solvers.

William St-Arnaud, Margarida Carvalho, Golnoosh Farnadi

Generative Flow Networks, known as GFlowNets, have been introduced in recent times, presenting an exciting possibility for neural networks to model distributions across various data structures. In this paper, we broaden their applicability to encompass scenarios where the data structures are optimal solutions of a combinatorial problem. Concretely, we propose the use of GFlowNets to learn the distribution of optimal solutions for kidney exchange problems (KEPs), a generalized form of matching problems involving cycles.

Seid Miad Zandavi

A new method is proposed to increase the accuracy of the state-of-the-art single image super-resolution (SISR) using novel training procedure. The proposed method, named post-trained convolutional neural network (CNN), is carried out stochastic dual simplex algorithm (SDSA) in the last reconstruction layer. The method utilizes contextual information to update the last reconstruction layer of CNN. The extracted contextual information is projected to the last reconstructed layer by optimized weights and the bias is managed through SDSA. Post-trained CNN is applied to the very deep super-resolution (VDSR) method to show its performance. The quantitative and visual results demonstrate that the proposed post-trained VDSR (PTVDSR) exhibits excellent and competitive performance when compared with the VDSR and other super-resolution methods.

Md Shahriar Iqbal, Jianhai Su, Lars Kotthoff, Pooyan Jamshidi

The design of machine learning systems often requires trading off different objectives, for example, prediction error and energy consumption for deep neural networks (DNNs). Typically, no single design performs well in all objectives; therefore, finding Pareto-optimal designs is of interest. The search for Pareto-optimal designs involves evaluating designs in an iterative process, and the measurements are used to evaluate an acquisition function that guides the search process. However, measuring different objectives incurs different costs. For example, the cost of measuring the prediction error of DNNs is orders of magnitude higher than that of measuring the energy consumption of a pre-trained DNN as it requires re-training the DNN. Current state-of-the-art methods do not consider this difference in objective evaluation cost, potentially incurring expensive evaluations of objective functions in the optimization process. In this paper, we develop a novel decoupled and cost-aware multi-objective optimization algorithm, which we call Flexible Multi-Objective Bayesian Optimization (FlexiBO) to address this issue. For evaluating each design, FlexiBO selects the objective with higher relative gain by weighting the improvement of the hypervolume of the Pareto region with the measurement cost of each objective. This strategy, therefore, balances the expense of collecting new information with the knowledge gained through objective evaluations, preventing FlexiBO from performing expensive measurements for little to no gain. We evaluate FlexiBO on seven state-of-the-art DNNs for image recognition, natural language processing (NLP), and speech-to-text translation. Our results indicate that, given the same total experimental budget, FlexiBO discovers designs with 4.8% to 12.4% lower hypervolume error than the best method in state-of-the-art multi-objective optimization.

Mingrui Liu

The current analysis of federated optimization algorithms for training deep neural networks assumes that the data is non-sequential (e.g., images), which incurs a smooth loss objective. In contrast, edge devices generate lots of sequential data every day, where these sequences exhibit significant sequential correlation at different time stamps (e.g., text messages). In order to learn from such sequential data, people typically use a class of neural networks that is inherently nonsmooth, with a potentially unbounded smoothness parameter. Examples include recurrent neural networks, long-short-term memory networks, and transformers. It remains unclear how to design provably efficient algorithms for training these neural networks to learn from sequential data. My goal is to lay the algorithmic foundation of federated learning with sequential data, which contributes novel algorithms for learning from a range of real-world sequential data (e.g., natural language, electronic health record, transportation, time series, etc.) using state-of-the-art deep neural networks. In this talk, I will first motivate the problem by showing that the transformer, which is widely used for sequential data learning, has an unbounded smooth landscape. Then, I will introduce provably efficient federated deep learning algorithms in the presence of unbounded smoothness. In particular, I will introduce a few efficient algorithms for various settings of federated learning, including homogeneous data, heterogeneous data, and partial client participation. The main result is twofold. First, we show that the designed algorithms provably small computational and communication complexities. Second, we establish fundamental hardness results in the unbounded smoothness setting. Ultimately, I will discuss the future challenges of extending our research framework from small-scale neural networks to large language models.

Wei Hu

Deep learning has exhibited a number of surprising generalization phenomena that are not captured by classical statistical learning theory. This talk will survey some of my work on the theoretical characterizations of several such intriguing phenomena: (1) Implicit regularization: A major mystery in deep learning is that deep neural networks can often generalize well despite their excessive expressive capacity. Towards explaining this mystery, it has been suggested that commonly used gradient-based optimization algorithms enforce certain implicit regularization which effectively constrains the model capacity. (2) Benign overfitting: In certain scenarios, a model can perfectly fit noisily labeled training data, but still archives near-optimal test error at the same time, which is very different from the classical notion of overfitting. (3) Grokking: In certain scenarios, a model initially achieves perfect training accuracy but no generalization (i.e. no better than a random predictor), and upon further training, transitions to almost perfect generalization. Theoretically establishing these properties often involves making appropriate high-dimensional assumptions on the problem as well as a careful analysis of the training dynamics.

Doseok Jang, Larry Yan, Lucas Spangher, Costas J. Spanos

Reinforcement learning (RL) is a powerful tool for optimal control that has found great success in Atari games, the game of Go, robotic control, and building optimization. RL is also very brittle; agents often overfit to their training environment and fail to generalize to new settings. Unsupervised environment design (UED) has been proposed as a solution to this problem, in which the agent trains in environments that have been specially selected to help it learn. Previous UED algorithms focus on trying to train an RL agent that generalizes across a large distribution of environments. This is not necessarily desirable when we wish to prioritize performance in one environment over others. In this work, we will be examining the setting of robust RL building control, where we wish to train an RL agent that prioritizes performing well in normal weather while still being robust to extreme weather conditions. We demonstrate a novel UED algorithm, ActivePLR, that uses uncertainty-aware neural network architectures to generate new training environments at the limit of the RL agent's ability while being able to prioritize performance in a desired base environment. We show that ActivePLR is able to outperform state-of-the-art UED algorithms in minimizing energy usage while maximizing occupant comfort in the setting of building control.

Hyunmin Choi, Simon S. Woo, Hyoungshick Kim

Fingerprint authentication is a popular security mechanism for smartphones and laptops. However, its adoption in web and cloud environments has been limited due to privacy concerns over storing and processing biometric data on servers. This paper introduces Blind-Touch, a novel machine learning-based fingerprint authentication system leveraging homomorphic encryption to address these privacy concerns. Homomorphic encryption allows computations on encrypted data without decrypting. Thus, Blind-Touch can keep fingerprint data encrypted on the server while performing machine learning operations. Blind-Touch combines three strategies to efficiently utilize homomorphic encryption in machine learning: (1) It optimizes the feature vector for a distributed architecture, processing the first fully connected layer (FC-16) in plaintext on the client side and the subsequent layer (FC-1) post-encryption on the server, thereby minimizing encrypted computations; (2) It employs a homomorphic encryption-compatible data compression technique capable of handling 8,192 authentication results concurrently; and (3) It utilizes a clustered server architecture to simultaneously process authentication results, thereby enhancing scalability with increasing user numbers. Blind-Touch achieves high accuracy on two benchmark fingerprint datasets, with a 93.6% F1- score for the PolyU dataset and a 98.2% F1-score for the SOKOTO dataset. Moreover, Blind-Touch can match a fingerprint among 5,000 in about 0.65 seconds. With its privacy-focused design, high accuracy, and efficiency, Blind-Touch is a promising alternative to conventional fingerprint authentication for web and cloud applications.

Yue Zhao, Congyi Li, Kai Chen

Recent advances in backdoor attacks, like leveraging complex triggers or stealthy implanting techniques, have introduced new challenges in backdoor scanning, limiting the usability of Deep Neural Networks (DNNs) in various scenarios. In this paper, we propose Unlearning-based Model Ablation (UMA), a novel approach to facilitate backdoor scanning and defend against advanced backdoor attacks. UMA filters out backdoor-irrelevant features by ablating the inherent features of the target class within the model and subsequently reveals the backdoor through dynamic trigger optimization. We evaluate our method on 1700 models (700 benign and 1000 trojaned) with 6 model structures, 7 different backdoor attacks and 4 datasets. Our results demonstrate that the proposed methodology effectively detect these advanced backdoors. Specifically, our method can achieve 91% AUC-ROC and 86.6% detection accuracy on average, which outperforms the baselines, including Neural Cleanse, ABS, K-Arm and MNTD.