In this paper, we study an underexplored, yet important and challenging problem: counting the number of distinct sounds in raw audio characterized by a high degree of polyphonicity. We do so by systematically proposing a novel end-to-end trainable neural network~(which we call DyDecNet, consisting of a dyadic decomposition front-end and backbone network), and quantifying the difficulty level of counting depending on sound polyphonicity. The dyadic decomposition front-end progressively decomposes the raw waveform dyadically along the frequency axis to obtain time-frequency representation in multi-stage, coarse-to-fine manner. Each intermediate waveform convolved by a parent filter is further processed by a pair of child filters that evenly split the parent filter's carried frequency response, with the higher-half child filter encoding the detail and lower-half child filter encoding the approximation. We further introduce an energy gain normalization to normalize sound loudness variance and spectrum overlap, and apply it to each intermediate parent waveform before feeding it to the two child filters. To better quantify sound counting difficulty level, we further design three polyphony-aware metrics: polyphony ratio, max polyphony and mean polyphony. We test DyDecNet on various datasets to show its superiority, and we further show dyadic decomposition network can be used as a general front-end to tackle other acoustic tasks.
论文检索
输入标题、作者或关键词，从 2,866 篇学术成果中精准定位
Most current long-tailed classification approaches assume the cost-agnostic scenario, where the training distribution of classes is long-tailed while the testing distribution of classes is balanced. Meanwhile, the misclassification costs of all instances are the same. On the other hand, in many real-world applications, it is more proper to assume that the training and testing distributions of classes are the same, while the misclassification cost of tail-class instances is varied. In this work, we model such a scenario as cost-aware long-tailed classification, in which the identification of high-cost tail instances and focusing learning on them thereafter is essential. In consequence, we propose the learning strategy of augmenting new instances based on adaptive region partition in the feature space. We conduct theoretical analysis to show that under the assumption that the feature-space distance and the misclassification cost are correlated, the identification of high-cost tail instances can be realized by building region partitions with a low variance of risk within each region. The resulting AugARP approach could significantly outperform baseline approaches on both benchmark datasets and real-world product sales datasets.
Detecting out-of-distribution (OOD) data is essential to ensure the reliability of machine learning models when deployed in real-world scenarios. Different from most previous test-time OOD detection methods that focus on designing OOD scores, we delve into the challenges in OOD detection from the perspective of typicality and regard the feature’s high-probability region as the feature’s typical set. However, the existing typical-feature-based OOD detection method implies an assumption: the proportion of typical feature sets for each channel is fixed. According to our experimental analysis, each channel contributes differently to OOD detection. Adopting a fixed proportion for all channels results in several channels losing too many typical features or incorporating too many abnormal features, resulting in low performance. Therefore, exploring the channel-aware typical features is crucial to better-separating ID and OOD data. Driven by this insight, we propose expLoring channel-Aware tyPical featureS (LAPS). Firstly, LAPS obtains the channel-aware typical set by calibrating the channel-level typical set with the global typical set from the mean and standard deviation. Then, LAPS rectifies the features into channel-aware typical sets to obtain channel-aware typical features. Finally, LAPS leverages the channel-aware typical features to calculate the energy score for OOD detection. Theoretical and visual analyses verify that LAPS achieves a better bias-variance trade-off. Experiments verify the effectiveness and generalization of LAPS under different architectures and OOD scores.
Label distribution learning (LDL) is an effective learning paradigm for handling label ambiguity. When applying LDL, it typically requires datasets annotated with label distributions. However, obtaining supervised data for LDL is a challenging task. Due to the randomness of label annotation, the annotator can produce inaccurate annotation results for the instance, affecting the accuracy and generalization ability of the LDL model. To address this problem, we propose a generative approach to calibrate the inaccurate annotation for LDL using variational inference techniques. Specifically, we assume that instances with similar features share latent similar label distributions. The feature vectors and label distributions are generated by Gaussian mixture and Dirichlet mixture, respectively. The relationship between them is established through a shared categorical variable, which effectively utilizes the label distribution of instances with similar features, and achieves a more accurate label distribution through the generative approach. Furthermore, we use a confusion matrix to model the factors that contribute to the inaccuracy during the annotation process, which captures the relationship between label distributions and inaccurate label distributions. Finally, the label distribution is used to calibrate the available information in the noisy dataset to obtain the ground-truth label distribution.
Existing semi-supervised domain adaptation (SSDA) models have exhibited impressive performance on the target domain by effectively utilizing few labeled target samples per class (e.g., 3 samples per class). To guarantee an equal number of labeled target samples for each class, however, they require domain experts to manually recognize a considerable amount of the unlabeled target data. Moreover, as the target samples are not equally informative for shaping the decision boundaries of the learning models, it is crucial to select the most informative target samples for labeling, which is, however, impossible for human selectors. As a remedy, we propose an EFfective Target Labeling (EFTL) framework that harnesses active learning and pseudo-labeling strategies to automatically select some informative target samples to annotate. Concretely, we introduce a novel sample query strategy, called non-maximal degree node suppression (NDNS), that iteratively performs maximal degree node query and non-maximal degree node removal to select representative and diverse target samples for labeling. To learn target-specific characteristics, we propose a novel pseudo-labeling strategy that attempts to label low-confidence target samples accurately via clustering consistency (CC), and then inject information of the model uncertainty into our query process. CC enhances the utilization of the annotation budget and increases the number of “labeled” target samples while requiring no additional manual effort. Our proposed EFTL framework can be easily coupled with existing SSDA models, showing significant improvements on three benchmarks
Not All Tasks Are Equally Difficult: Multi-Task Deep Reinforcement Learning with Dynamic Depth Routing
PDF ↗Multi-task reinforcement learning endeavors to accomplish a set of different tasks with a single policy. To enhance data efficiency by sharing parameters across multiple tasks, a common practice segments the network into distinct modules and trains a routing network to recombine these modules into task-specific policies. However, existing routing approaches employ a fixed number of modules for all tasks, neglecting that tasks with varying difficulties commonly require varying amounts of knowledge. This work presents a Dynamic Depth Routing (D2R) framework, which learns strategic skipping of certain intermediate modules, thereby flexibly choosing different numbers of modules for each task. Under this framework, we further introduce a ResRouting method to address the issue of disparate routing paths between behavior and target policies during off-policy training. In addition, we design an automatic route-balancing mechanism to encourage continued routing exploration for unmastered tasks without disturbing the routing of mastered ones. We conduct extensive experiments on various robotics manipulation tasks in the Meta-World benchmark, where D2R achieves state-of-the-art performance with significantly improved learning efficiency.
Among existing Neural Architecture Search methods, DARTS is known for its efficiency and simplicity. This approach applies continuous relaxation of network representation to construct a weight-sharing supernet and enables the identification of excellent subnets in just a few GPU days. However, performance collapse in DARTS results in deteriorating architectures filled with parameter-free operations and remains a great challenge to the robustness. To resolve this problem, we reveal that the fundamental reason is the biased estimation of the candidate importance in the search space through theoretical and experimental analysis, and more precisely select operations via information-based measurements. Furthermore, we demonstrate that the excessive concern over the supernet and inefficient utilization of data in bi-level optimization also account for suboptimal results. We adopt a more realistic objective focusing on the performance of subnets and simplify it with the help of the informationbased measurements. Finally, we explain theoretically why progressively shrinking the width of the supernet is necessary and reduce the approximation error of optimal weights in DARTS. Our proposed method, named IS-DARTS, comprehensively improves DARTS and resolves the aforementioned problems. Extensive experiments on NAS-Bench-201 and DARTS-based search space demonstrate the effectiveness of IS-DARTS.
Offline meta-reinforcement learning (meta-RL) methods, which adapt to unseen target tasks with prior experience, are essential in robot control tasks. Current methods typically utilize task contexts and skills as prior experience, where task contexts are related to the information within each task and skills represent a set of temporally extended actions for solving subtasks. However, these methods still suffer from limited performance when adapting to unseen target tasks, mainly because the learned prior experience lacks generalization, i.e., they are unable to extract effective prior experience from meta-training tasks by exploration and learning of continuous latent spaces. We propose a framework called decoupled meta-reinforcement learning (DCMRL), which (1) contrastively restricts the learning of task contexts through pulling in similar task contexts within the same task and pushing away different task contexts of different tasks, and (2) utilizes a Gaussian quantization variational autoencoder (GQ-VAE) for clustering the Gaussian distributions of the task contexts and skills respectively, and decoupling the exploration and learning processes of their spaces. These cluster centers which serve as representative and discrete distributions of task context and skill are stored in task context codebook and skill codebook, respectively. DCMRL can acquire generalizable prior experience and achieve effective adaptation to unseen target tasks during the meta-testing phase. Experiments in the navigation and robot manipulation continuous control tasks show that DCMRL is more effective than previous meta-RL methods with more generalizable prior experience.
Graph Neural Networks (GNNs) have received widespread attention and applications due to their excellent performance in graph representation learning. Most existing GNNs can only aggregate 1-hop neighbors in a GNN layer, so they usually stack multiple GNN layers to obtain more information from larger neighborhoods. However, many studies have shown that model performance experiences a significant degradation with the increase of GNN layers. In this paper, we first introduce the concept of distinguishability of class to indirectly evaluate the learned node representations, and verify the positive correlation between distinguishability of class and model performance. Then, we propose a Graph Neural Network guided by Distinguishability of class (Disc-GNN) to monitor the representation learning, so as to learn better node representations and improve model performance. Specifically, we first perform inter-layer filtering and initial compensation based on Local Distinguishability of Class (LDC) in each layer, so that the learned node representations have the ability to distinguish different classes. Furthermore, we add a regularization term based on Global Distinguishability of Class (GDC) to achieve global optimization of model performance. Extensive experiments on six real-world datasets have shown that the competitive performance of Disc-GNN to the state-of-the-art methods on node classification and node clustering tasks.
Graph contrastive learning (GCL) has attracted considerable attention because it can self-supervisedly extract low-dimensional representation of graph data. InfoNCE-based loss function is widely used in graph contrastive learning, which pulls the representations of positive pairs close to each other and pulls the representations of negative pairs away from each other. Recent works mainly focus on designing new augmentation methods or sampling strategies. However, we argue that the widely used InfoNCE-based methods may contain an implicit conflict which seriously confuses models when learning from negative pairs. This conflict is engendered by the encoder's message-passing mechanism and the InfoNCE loss function. As a result, the learned representations between negative samples cannot be far away from each other, compromising the model performance. To our best knowledge, this is the first time to report and analysis this conflict of GCL. To address this problem, we propose a simple but effective method called Partial ignored Graph Contrastive Learning (PiGCL). Specifically, PiGCL first dynamically captures the conflicts during training by detecting the gradient of representation similarities. It then enables the loss function to ignore the conflict, allowing the encoder to adaptively learn the ignored information without self-supervised samples. Extensive experiments demonstrate the effectiveness of our method.
Fairness under Covariate Shift: Improving Fairness-Accuracy Tradeoff with Few Unlabeled Test Samples
PDF ↗Covariate shift in the test data is a common practical phenomena that can significantly downgrade both the accuracy and the fairness performance of the model. Ensuring fairness across different sensitive groups under covariate shift is of paramount importance due to societal implications like criminal justice. We operate in the unsupervised regime where only a small set of unlabeled test samples along with a labeled training set is available. Towards improving fairness under this highly challenging yet realistic scenario, we make three contributions. First is a novel composite weighted entropy based objective for prediction accuracy which is optimized along with a representation matching loss for fairness. We experimentally verify that optimizing with our loss formulation outperforms a number of state-of-the-art baselines in the pareto sense with respect to the fairness-accuracy tradeoff on several standard datasets. Our second contribution is a new setting we term Asymmetric Covariate Shift that, to the best of our knowledge, has not been studied before. Asymmetric covariate shift occurs when distribution of covariates of one group shifts significantly compared to the other groups and this happens when a dominant group is over-represented. While this setting is extremely challenging for current baselines, We show that our proposed method significantly outperforms them. Our third contribution is theoretical, where we show that our weighted entropy term along with prediction loss on the training set approximates test loss under covariate shift. Empirically and through formal sample complexity bounds, we show that this approximation to the unseen test loss does not depend on importance sampling variance which affects many other baselines.
In light of the advances in big data, high-dimensional datasets are often encountered. Incorporating them into data-driven models can enhance performance; however, this comes at the cost of high computation and the risk of overfitting, particularly due to abundant redundant features. Identifying an informative subset of the features helps in reducing the dimensionality and enhancing model interpretability. In this paper, we propose a novel framework for unsupervised feature selection, called Selective Deep Auto-Encoder (SDAE). It aims to reduce the number of features used in unlabeled datasets without compromising the quality of information obtained. It achieves this by selecting sufficient features - from the original feature set - capable of representing the entire feature space and reconstructing them. Architecturally, it leverages the use of highly nonlinear latent representations in deep Autoencoders and intrinsically learns, in an unsupervised fashion, the relevant and globally representative subset of features through a customized Selective Layer. Extensive experimental results on three high-dimensional public datasets have shown promising feature selection performance by SDAE in comparison to other existing state-of-the-art unsupervised feature selection methods.
Federated Learning (FL) involves training a model over a dataset distributed among clients, with the constraint that each client’s dataset is localized and possibly heterogeneous. In FL, small and noisy datasets are common, highlighting the need for well-calibrated models that represent the uncertainty of predictions. The closest FL techniques to achieving such goals are the Bayesian FL methods which collect parameter samples from local posteriors, and aggregate them to approximate the global posterior. To improve scalability for larger models, one common Bayesian approach is to approximate the global predictive posterior by multiplying local predictive posteriors. In this work, we demonstrate that this method gives systematically overconfident predictions, and we remedy this by proposing β-Predictive Bayes, a Bayesian FL algorithm that interpolates between a mixture and product of the predictive posteriors, using a tunable parameter β. This parameter is tuned to improve the global ensemble’s calibration, before it is distilled to a single model. Our method is evaluated on a variety of regression and classification datasets to demonstrate its superiority in calibration to other baselines, even as data heterogeneity increases. Code available at https://github.com/hasanmohsin/betaPredBayesFL. Our paper's full version is at https://arxiv.org/abs/2312.09817.
AlphaZero and MuZero have achieved state-of-the-art (SOTA) performance in a wide range of domains, including board games and robotics, with discrete and continuous action spaces. However, to obtain an improved policy, they often require an excessively large number of simulations, especially for domains with large action spaces. As the simulation budget decreases, their performance drops significantly. In addition, many important real-world applications have combinatorial (or exponential) action spaces, making it infeasible to search directly over all possible actions. In this paper, we extend AlphaZero and MuZero to learn and plan in more complex multiagent (MA) Markov decision processes, where the action spaces increase exponentially with the number of agents. Our new algorithms, MA Gumbel AlphaZero and MA Gumbel MuZero, respectively without and with model learning, achieve superior performance on cooperative multiagent control problems, while reducing the number of environmental interactions by up to an order of magnitude compared to model-free approaches. In particular, we significantly improve prior performance when planning with much fewer simulation budgets. The code and appendix are available at https://github.com/tjuHaoXiaotian/MA-MuZero.
Double-Layer Hybrid-Label Identification Feature Selection for Multi-View Multi-Label Learning
PDF ↗Multi-view multi-label feature selection aims to select informative features where the data are collected from multiple sources with multiple interdependent class labels. For fully exploiting multi-view information, most prior works mainly focus on the common part in the ideal circumstance. However, the inconsistent part hidden in each view, including noises and specific elements, may affect the quality of mapping between labels and feature representations. Meanwhile, ignoring the specific part might lead to a suboptimal result, as each label is supposed to possess specific characteristics of its own. To deal with the double problems in multi-view multi-label feature selection, we propose a unified loss function which is a totally splitting structure for observed labels as hybrid labels that is, common labels, view-to-all specific labels and noisy labels, and the view-to-all specific labels further splits into several specific labels of each view. The proposed method simultaneously considers the consistency and complementarity of different views. Through exploring the feature weights of hybrid labels, the mapping relationships between labels and features can be established sequentially based on their attributes. Additionally, the interrelatedness among hybrid labels is also investigated and injected into the function. Specific to the specific labels of each view, we construct the novel regularization paradigm incorporating logic operations. Finally, the convergence of the result is proved after applying the multiplicative update rules. Experiments on six datasets demonstrate the effectiveness and superiority of our method compared with the state-of-the-art methods.
Active learning (AL) aims to improve model performance within a fixed labeling budget by choosing the most informative data points to label. Existing AL focuses on the single-domain setting, where all data come from the same domain (e.g., the same dataset). However, many real-world tasks often involve multiple domains. For example, in visual recognition, it is often desirable to train an image classifier that works across different environments (e.g., different backgrounds), where images from each environment constitute one domain. Such a multi-domain AL setting is challenging for prior methods because they (1) ignore the similarity among different domains when assigning labeling budget and (2) fail to handle distribution shift of data across different domains. In this paper, we propose the first general method, dubbed composite active learning (CAL), for multi-domain AL. Our approach explicitly considers the domain-level and instance-level information in the problem; CAL first assigns domain-level budgets according to domain-level importance, which is estimated by optimizing an upper error bound that we develop; with the domain-level budgets, CAL then leverages a certain instance-level query strategy to select samples to label from each domain. Our theoretical analysis shows that our method achieves a better error bound compared to current AL methods. Our empirical results demonstrate that our approach significantly outperforms the state-of-the-art AL methods on both synthetic and real-world multi-domain datasets. Code is available at https://github.com/Wang-ML-Lab/multi-domain-active-learning.
Complexity of Neural Network Training and ETR: Extensions with Effectively Continuous Functions
PDF ↗The training problem of neural networks (NNs) is known to be ER-complete with respect to ReLU and linear activation functions. We show that the training problem for NNs equipped with arbitrary activation functions is polynomial-time bireducible to the existential theory of the reals extended with the corresponding activation functions. For effectively continuous activation functions (e.g., the sigmoid function), we obtain an inclusion to low levels of the arithmetical hierarchy. Consequently, the sigmoid activation function leads to the existential theory of the reals with the exponential function, and hence the decidability of training NNs using the sigmoid activation function is equivalent to the decidability of the existential theory of the reals with the exponential function, a long-standing open problem. In contrast, we obtain that the training problem is undecidable if sinusoidal activation functions are considered.
The multi-armed bandit(MAB) is a classical sequential decision problem. Most work requires assumptions about the reward distribution (e.g., bounded), while practitioners may have difficulty obtaining information about these distributions to design models for their problems, especially in non-stationary MAB problems. This paper aims to design a multi-armed bandit algorithm that can be implemented without using information about the reward distribution while still achieving substantial regret upper bounds. To this end, we propose a novel algorithm alternating between greedy rule and forced exploration. Our method can be applied to Gaussian, Bernoulli and other subgaussian distributions, and its implementation does not require additional information. We employ a unified analysis method for different forced exploration strategies and provide problem-dependent regret upper bounds for stationary and piecewise-stationary settings. Furthermore, we compare our algorithm with popular bandit algorithms on different reward distributions.
Fine-tuning pre-trained neural network models has become a widely adopted approach across various domains. However, it can lead to the distortion of pre-trained feature extractors that already possess strong generalization capabilities. Mitigating feature distortion during adaptation to new target domains is crucial. Recent studies have shown promising results in handling feature distortion by aligning the head layer on in-distribution datasets before performing fine-tuning. Nonetheless, a significant limitation arises from the treatment of batch normalization layers during fine-tuning, leading to suboptimal performance. In this paper, we propose Domain-Aware Fine-Tuning (DAFT), a novel approach that incorporates batch normalization conversion and the integration of linear probing and fine-tuning. Our batch normalization conversion method effectively mitigates feature distortion by reducing modifications to the neural network during fine-tuning. Additionally, we introduce the integration of linear probing and fine-tuning to optimize the head layer with gradual adaptation of the feature extractor. By leveraging batch normalization layers and integrating linear probing and fine-tuning, our DAFT significantly mitigates feature distortion and achieves improved model performance on both in-distribution and out-of-distribution datasets. Extensive experiments demonstrate that our method outperforms other baseline methods, demonstrating its effectiveness in not only improving performance but also mitigating feature distortion.
In this paper, we introduce a fresh perspective on the challenges of credit assignment and policy evaluation. First, we delve into the nuances of eligibility traces and explore instances where their updates may result in unexpected credit assignment to preceding states. From this investigation emerges the concept of a novel value function, which we refer to as the ????????????? ????? ????????. Unlike traditional state value functions, bidirectional value functions account for both future expected returns (rewards anticipated from the current state onward) and past expected returns (cumulative rewards from the episode's start to the present). We derive principled update equations to learn this value function and, through experimentation, demonstrate its efficacy in enhancing the process of policy evaluation. In particular, our results indicate that the proposed learning approach can, in certain challenging contexts, perform policy evaluation more rapidly than TD(λ)–a method that learns forward value functions, v^π, ????????. Overall, our findings present a new perspective on eligibility traces and potential advantages associated with the novel value function it inspires, especially for policy evaluation.