Machine Learning (cs.LG)

Abstract: Sparse mixture of expert architectures (MoEs) scale model capacity without large increases in training or inference costs. Despite their success, MoEs suffer from a number of issues: training instability, token dropping, inability to scale the number of experts, or ineffective finetuning. In this work, we proposeSoft MoE, a fully-differentiable sparse Transformer that addresses these challenges, while maintaining the benefits of MoEs. Soft MoE performs an implicit soft assignment by passing different weighted combinations of all input tokens to each expert. As in other MoE works, experts in Soft MoE only process a subset of the (combined) tokens, enabling larger model capacity at lower inference cost. In the context of visual recognition, Soft MoE greatly outperforms standard Transformers (ViTs) and popular MoE variants (Tokens Choice and Experts Choice). For example, Soft MoE-Base/16 requires 10.5x lower inference cost (5.7x lower wall-clock time) than ViT-Huge/14 while matching its performance after similar training. Soft MoE also scales well: Soft MoE Huge/14 with 128 experts in 16 MoE layers has over 40x more parameters than ViT Huge/14, while inference time cost grows by only 2%, and it performs substantially better.

Abstract: We consider the problem of multi-fidelity zeroth-order optimization, where one can evaluate a function $f$ at various approximation levels (of varying costs), and the goal is to optimize $f$ with the cheapest evaluations possible. In this paper, we study \emph{certified} algorithms, which are additionally required to output a data-driven upper bound on the optimization error. We first formalize the problem in terms of a min-max game between an algorithm and an evaluation environment. We then propose a certified variant of the MFDOO algorithm and derive a bound on its cost complexity for any Lipschitz function $f$. We also prove an $f$-dependent lower bound showing that this algorithm has a near-optimal cost complexity. We close the paper by addressing the special case of noisy (stochastic) evaluations as a direct example.

Abstract: Hierarchical reinforcement learning composites subpolicies in different hierarchies to accomplish complex tasks.Automated subpolicies discovery, which does not depend on domain knowledge, is a promising approach to generating subpolicies.However, the degradation problem is a challenge that existing methods can hardly deal with due to the lack of consideration of diversity or the employment of weak regularizers. In this paper, we propose a novel task-agnostic regularizer called the Wasserstein Diversity-Enriched Regularizer (WDER), which enlarges the diversity of subpolicies by maximizing the Wasserstein distances among action distributions. The proposed WDER can be easily incorporated into the loss function of existing methods to boost their performance further.Experimental results demonstrate that our WDER improves performance and sample efficiency in comparison with prior work without modifying hyperparameters, which indicates the applicability and robustness of the WDER.

Abstract: Time series analysis is a fundamental task in various application domains, and deep learning approaches have demonstrated remarkable performance in this area. However, many real-world time series data exhibit significant periodic or quasi-periodic dynamics that are often not adequately captured by existing deep learning-based solutions. This results in an incomplete representation of the underlying dynamic behaviors of interest. To address this gap, we propose an unsupervised method called Floss that automatically regularizes learned representations in the frequency domain. The Floss method first automatically detects major periodicities from the time series. It then employs periodic shift and spectral density similarity measures to learn meaningful representations with periodic consistency. In addition, Floss can be easily incorporated into both supervised, semi-supervised, and unsupervised learning frameworks. We conduct extensive experiments on common time series classification, forecasting, and anomaly detection tasks to demonstrate the effectiveness of Floss. We incorporate Floss into several representative deep learning solutions to justify our design choices and demonstrate that it is capable of automatically discovering periodic dynamics and improving state-of-the-art deep learning models.

Abstract: This paper discusses the system architecture design and deployment of non-stationary multi-armed bandit approaches to determine a near-optimal payment routing policy based on the recent history of transactions. We propose a Routing Service architecture using a novel Ray-based implementation for optimally scaling bandit-based payment routing to over 10000 transactions per second, adhering to the system design requirements and ecosystem constraints with Payment Card Industry Data Security Standard (PCI DSS). We first evaluate the effectiveness of multiple bandit-based payment routing algorithms on a custom simulator to benchmark multiple non-stationary bandit approaches and identify the best hyperparameters. We then conducted live experiments on the payment transaction system on a fantasy sports platform Dream11. In the live experiments, we demonstrated that our non-stationary bandit-based algorithm consistently improves the success rate of transactions by 0.92\% compared to the traditional rule-based methods over one month.

Abstract: In the problem of out-of-distribution (OOD) detection, the usage of auxiliary data as outlier data for fine-tuning has demonstrated encouraging performance. However, previous methods have suffered from a trade-off between classification accuracy (ACC) and OOD detection performance (AUROC, FPR, AUPR). To improve this trade-off, we make three contributions: (i) Incorporating a self-knowledge distillation loss can enhance the accuracy of the network; (ii) Sampling semi-hard outlier data for training can improve OOD detection performance with minimal impact on accuracy; (iii) The introduction of our novel supervised contrastive learning can simultaneously improve OOD detection performance and the accuracy of the network. By incorporating all three factors, our approach enhances both accuracy and OOD detection performance by addressing the trade-off between classification and OOD detection. Our method achieves improvements over previous approaches in both performance metrics.

Abstract: We provide an implementation to compute the flat metric in any dimension. The flat metric, also called dual bounded Lipschitz distance, generalizes the well-known Wasserstein distance W1 to the case that the distributions are of unequal total mass. This is of particular interest for unbalanced optimal transport tasks and for the analysis of data distributions where the sample size is important or normalization is not possible. The core of the method is based on a neural network to determine on optimal test function realizing the distance between two given measures. Special focus was put on achieving comparability of pairwise computed distances from independently trained networks. We tested the quality of the output in several experiments where ground truth was available as well as with simulated data.

Abstract: Graph anomaly detection (GAD) has achieved success and has been widely applied in various domains, such as fraud detection, cybersecurity, finance security, and biochemistry. However, existing graph anomaly detection algorithms focus on distinguishing individual entities (nodes or graphs) and overlook the possibility of anomalous groups within the graph. To address this limitation, this paper introduces a novel unsupervised framework for a new task called Group-level Graph Anomaly Detection (Gr-GAD). The proposed framework first employs a variant of Graph AutoEncoder (GAE) to locate anchor nodes that belong to potential anomaly groups by capturing long-range inconsistencies. Subsequently, group sampling is employed to sample candidate groups, which are then fed into the proposed Topology Pattern-based Graph Contrastive Learning (TPGCL) method. TPGCL utilizes the topology patterns of groups as clues to generate embeddings for each candidate group and thus distinct anomaly groups. The experimental results on both real-world and synthetic datasets demonstrate that the proposed framework shows superior performance in identifying and localizing anomaly groups, highlighting it as a promising solution for Gr-GAD. Datasets and codes of the proposed framework are at the github repository https://anonymous.4open.science/r/Topology-Pattern-Enhanced-Unsupervised-Group-level-Graph-Anomaly-Detection.

Abstract: The principle of boosting in supervised learning involves combining multiple weak classifiers to obtain a stronger classifier. AdaBoost has the reputation to be a perfect example of this approach. We have previously shown that AdaBoost is not truly an optimization algorithm. This paper shows that AdaBoost is an algorithm in name only, as the resulting combination of weak classifiers can be explicitly calculated using a truth table. This study is carried out by considering a problem with two classes and is illustrated by the particular case of three binary classifiers and presents results in comparison with those from the implementation of AdaBoost algorithm of the Python library scikit-learn.

Abstract: Time Series Classification (TSC) has received much attention in the past two decades and is still a crucial and challenging problem in data science and knowledge engineering. Indeed, along with the increasing availability of time series data, many TSC algorithms have been suggested by the research community in the literature. Besides state-of-the-art methods based on similarity measures, intervals, shapelets, dictionaries, deep learning methods or hybrid ensemble methods, several tools for extracting unsupervised informative summary statistics, aka features, from time series have been designed in the recent years. Originally designed for descriptive analysis and visualization of time series with informative and interpretable features, very few of these feature engineering tools have been benchmarked for TSC problems and compared with state-of-the-art TSC algorithms in terms of predictive performance. In this article, we aim at filling this gap and propose a simple TSC process to evaluate the potential predictive performance of the feature sets obtained with existing feature engineering tools. Thus, we present an empirical study of 11 feature engineering tools branched with 9 supervised classifiers over 112 time series data sets. The analysis of the results of more than 10000 learning experiments indicate that feature-based methods perform as accurately as current state-of-the-art TSC algorithms, and thus should rightfully be considered further in the TSC literature.

Abstract: We present a framework for learning Hamiltonian systems using data. This work is based on the lifting hypothesis, which posits that nonlinear Hamiltonian systems can be written as nonlinear systems with cubic Hamiltonians. By leveraging this, we obtain quadratic dynamics that are Hamiltonian in a transformed coordinate system. To that end, for given generalized position and momentum data, we propose a methodology to learn quadratic dynamical systems, enforcing the Hamiltonian structure in combination with a symplectic auto-encoder. The enforced Hamiltonian structure exhibits long-term stability of the system, while the cubic Hamiltonian function provides relatively low model complexity. For low-dimensional data, we determine a higher-order transformed coordinate system, whereas, for high-dimensional data, we find a lower-order coordinate system with the desired properties. We demonstrate the proposed methodology by means of both low-dimensional and high-dimensional nonlinear Hamiltonian systems.

Abstract: Spectrum analysis systems in online water quality testing are designed to detect types and concentrations of pollutants and enable regulatory agencies to respond promptly to pollution incidents. However, spectral data-based testing devices suffer from complex noise patterns when deployed in non-laboratory environments. To make the analysis model applicable to more environments, we propose a noise patterns transferring model, which takes the spectrum of standard water samples in different environments as cases and learns the differences in their noise patterns, thus enabling noise patterns to transfer to unknown samples. Unfortunately, the inevitable sample-level baseline noise makes the model unable to obtain the paired data that only differ in dataset-level environmental noise. To address the problem, we generate a sample-to-sample case-base to exclude the interference of sample-level noise on dataset-level noise learning, enhancing the system's learning performance. Experiments on spectral data with different background noises demonstrate the good noise-transferring ability of the proposed method against baseline systems ranging from wavelet denoising, deep neural networks, and generative models. From this research, we posit that our method can enhance the performance of DL models by generating high-quality cases. The source code is made publicly available online at https://github.com/Magnomic/CNST.

Abstract: This paper proposes a locally differentially private federated learning algorithm for strongly convex but possibly nonsmooth problems that protects the gradients of each worker against an honest but curious server. The proposed algorithm adds artificial noise to the shared information to ensure privacy and dynamically allocates the time-varying noise variance to minimize an upper bound of the optimization error subject to a predefined privacy budget constraint. This allows for an arbitrarily large but finite number of iterations to achieve both privacy protection and utility up to a neighborhood of the optimal solution, removing the need for tuning the number of iterations. Numerical results show the superiority of the proposed algorithm over state-of-the-art methods.

Abstract: In contemporary self-supervised contrastive algorithms like SimCLR, MoCo, etc., the task of balancing attraction between two semantically similar samples and repulsion between two samples from different classes is primarily affected by the presence of hard negative samples. While the InfoNCE loss has been shown to impose penalties based on hardness, the temperature hyper-parameter is the key to regulating the penalties and the trade-off between uniformity and tolerance. In this work, we focus our attention to improve the performance of InfoNCE loss in SSL by studying the effect of temperature hyper-parameter values. We propose a cosine similarity-dependent temperature scaling function to effectively optimize the distribution of the samples in the feature space. We further analyze the uniformity and tolerance metrics to investigate the optimal regions in the cosine similarity space for better optimization. Additionally, we offer a comprehensive examination of the behavior of local and global structures in the feature space throughout the pre-training phase, as the temperature varies. Experimental evidence shows that the proposed framework outperforms or is at par with the contrastive loss-based SSL algorithms. We believe our work (DySTreSS) on temperature scaling in SSL provides a foundation for future research in contrastive learning.

Abstract: Off-policy learning enables a reinforcement learning (RL) agent to reason counterfactually about policies that are not executed and is one of the most important ideas in RL. It, however, can lead to instability when combined with function approximation and bootstrapping, two arguably indispensable ingredients for large-scale reinforcement learning. This is the notorious deadly triad. Gradient Temporal Difference (GTD) is one powerful tool to solve the deadly triad. Its success results from solving a doubling sampling issue indirectly with weight duplication or Fenchel duality. In this paper, we instead propose a direct method to solve the double sampling issue by simply using two samples in a Markovian data stream with an increasing gap. The resulting algorithm is as computationally efficient as GTD but gets rid of GTD's extra weights. The only price we pay is a logarithmically increasing memory as time progresses. We provide both asymptotic and finite sample analysis, where the convergence rate is on-par with the canonical on-policy temporal difference learning. Key to our analysis is a novel refined discretization of limiting ODEs.

Abstract: Our Visual Analytics (VA) tool ScrutinAI supports human analysts to investigate interactively model performanceand data sets. Model performance depends on labeling quality to a large extent. In particular in medical settings, generation of high quality labels requires in depth expert knowledge and is very costly. Often, data sets are labeled by collecting opinions of groups of experts. We use our VA tool to analyse the influence of label variations between different experts on the model performance. ScrutinAI facilitates to perform a root cause analysis that distinguishes weaknesses of deep neural network (DNN) models caused by varying or missing labeling quality from true weaknesses. We scrutinize the overall detection of intracranial hemorrhages and the more subtle differentiation between subtypes in a publicly available data set.

Abstract: Calibrating deep neural models plays an important role in building reliable, robust AI systems in safety-critical applications. Recent work has shown that modern neural networks that possess high predictive capability are poorly calibrated and produce unreliable model predictions. Though deep learning models achieve remarkable performance on various benchmarks, the study of model calibration and reliability is relatively underexplored. Ideal deep models should have not only high predictive performance but also be well calibrated. There have been some recent methods proposed to calibrate deep models by using different mechanisms. In this survey, we review the state-of-the-art calibration methods and provide an understanding of their principles for performing model calibration. First, we start with the definition of model calibration and explain the root causes of model miscalibration. Then we introduce the key metrics that can measure this aspect. It is followed by a summary of calibration methods that we roughly classified into four categories: post-hoc calibration, regularization methods, uncertainty estimation, and composition methods. We also covered some recent advancements in calibrating large models, particularly large language models (LLMs). Finally, we discuss some open issues, challenges, and potential directions.

Abstract: In this paper we present a practical Bayesian self-supervised learning method with Cyclical Stochastic Gradient Hamiltonian Monte Carlo (cSGHMC). Within this framework, we place a prior over the parameters of a self-supervised learning model and use cSGHMC to approximate the high dimensional and multimodal posterior distribution over the embeddings. By exploring an expressive posterior over the embeddings, Bayesian self-supervised learning produces interpretable and diverse representations. Marginalizing over these representations yields a significant gain in performance, calibration and out-of-distribution detection on a variety of downstream classification tasks. We provide experimental results on multiple classification tasks on four challenging datasets. Moreover, we demonstrate the effectiveness of the proposed method in out-of-distribution detection using the SVHN and CIFAR-10 datasets.

Abstract: We present Lode Encoder, a gamified mixed-initiative level creation system for the classic platform-puzzle game Lode Runner. The system is built around several autoencoders which are trained on sets of Lode Runner levels. When fed with the user's design, each autoencoder produces a version of that design which is closer in style to the levels that it was trained on. The Lode Encoder interface allows the user to build and edit levels through 'painting' from the suggestions provided by the autoencoders. Crucially, in order to encourage designers to explore new possibilities, the system does not include more traditional editing tools. We report on the system design and training procedure, as well as on the evolution of the system itself and user tests.