Search

In this work we study how diffusion-based generative models produce high-dimensional data, such as an image, by implicitly relying on a manifestation of a low-dimensional set of latent abstractions, that guide the generative process. We present a novel theoretical framework that extends NLF, and that offers a unique perspective on SDE-based generative models. The development of our theory relies on a novel formulation of the joint (state and measurement) dynamics, and an information-theoretic measure of the influence of the system state on the measurement process. According to our theory, diffusion models can be cast as a system of SDE, describing a non-linear filter in which the evolution of unobservable latent abstractions steers the dynamics of an observable measurement process (corresponding to the generative pathways). In addition, we present an empirical study to validate our theory and previous empirical results on the emergence of latent abstractions at different stages of the generative process.

Tabular data is ubiquitous in many real-life systems. In particular, time-dependent tabular data, where rows are chronologically related, is typically used for recording historical events, e.g., financial transactions, healthcare records, or stock history. Recently, hierarchical variants of the attention mechanism of transformer architectures have been used to model tabular time-series data. At first, rows (or columns) are encoded separately by computing attention between their fields. Subsequently, encoded rows (or columns) are attended to one another to model the entire tabular time-series. While efficient, this approach constrains the attention granularity and limits its ability to learn patterns at the field-level across separate rows, or columns. We take a first step to address this gap by proposing Fieldy, a fine-grained hierarchical model that contextualizes fields at both the row and column levels. We compare our proposal against state of the art models on regression and classification tasks using public tabular time-series datasets. Our results show that combining row-wise and column-wise attention improves performance without increasing model size. Code and data are available at https://github.com/raphaaal/fieldy., Comment: 9 pages; Camera Ready version

Diffusion models for Text-to-Image (T2I) conditional generation have seen tremendous success recently. Despite their success, accurately capturing user intentions with these models still requires a laborious trial and error process. This challenge is commonly identified as a model alignment problem, an issue that has attracted considerable attention by the research community. Instead of relying on fine-grained linguistic analyses of prompts, human annotation, or auxiliary vision-language models to steer image generation, in this work we present a novel method that relies on an information-theoretic alignment measure. In a nutshell, our method uses self-supervised fine-tuning and relies on point-wise mutual information between prompts and images to define a synthetic training set to induce model alignment. Our comparative analysis shows that our method is on-par or superior to the state-of-the-art, yet requires nothing but a pre-trained denoising network to estimate MI and a lightweight fine-tuning strategy.

The analysis of scientific data and complex multivariate systems requires information quantities that capture relationships among multiple random variables. Recently, new information-theoretic measures have been developed to overcome the shortcomings of classical ones, such as mutual information, that are restricted to considering pairwise interactions. Among them, the concept of information synergy and redundancy is crucial for understanding the high-order dependencies between variables. One of the most prominent and versatile measures based on this concept is O-information, which provides a clear and scalable way to quantify the synergy-redundancy balance in multivariate systems. However, its practical application is limited to simplified cases. In this work, we introduce S$\Omega$I, which allows for the first time to compute O-information without restrictive assumptions about the system. Our experiments validate our approach on synthetic data, and demonstrate the effectiveness of S$\Omega$I in the context of a real-world use case.

Data Augmentation (DA) -- enriching training data by adding synthetic samples -- is a technique widely adopted in Computer Vision (CV) and Natural Language Processing (NLP) tasks to improve models performance. Yet, DA has struggled to gain traction in networking contexts, particularly in Traffic Classification (TC) tasks. In this work, we fulfill this gap by benchmarking 18 augmentation functions applied to 3 TC datasets using packet time series as input representation and considering a variety of training conditions. Our results show that (i) DA can reap benefits previously unexplored, (ii) augmentations acting on time series sequence order and masking are better suited for TC than amplitude augmentations and (iii) basic models latent space analysis can help understanding the positive/negative effects of augmentations on classification performance., Comment: to appear at Passive and Active Measurements (PAM), 2024

Data Augmentation (DA)-augmenting training data with synthetic samples-is wildly adopted in Computer Vision (CV) to improve models performance. Conversely, DA has not been yet popularized in networking use cases, including Traffic Classification (TC). In this work, we present a preliminary study of 14 hand-crafted DAs applied on the MIRAGE19 dataset. Our results (i) show that DA can reap benefits previously unexplored in TC and (ii) foster a research agenda on the use of generative models to automate DA design., Comment: to appear at CoNEXT Student Workshop, 2023

In this work we present a new method for the estimation of Mutual Information (MI) between random variables. Our approach is based on an original interpretation of the Girsanov theorem, which allows us to use score-based diffusion models to estimate the Kullback Leibler divergence between two densities as a difference between their score functions. As a by-product, our method also enables the estimation of the entropy of random variables. Armed with such building blocks, we present a general recipe to measure MI, which unfolds in two directions: one uses conditional diffusion process, whereas the other uses joint diffusion processes that allow simultaneous modelling of two random variables. Our results, which derive from a thorough experimental protocol over all the variants of our approach, indicate that our method is more accurate than the main alternatives from the literature, especially for challenging distributions. Furthermore, our methods pass MI self-consistency tests, including data processing and additivity under independence, which instead are a pain-point of existing methods.

Multi-modal data-sets are ubiquitous in modern applications, and multi-modal Variational Autoencoders are a popular family of models that aim to learn a joint representation of the different modalities. However, existing approaches suffer from a coherence-quality tradeoff, where models with good generation quality lack generative coherence across modalities, and vice versa. We discuss the limitations underlying the unsatisfactory performance of existing methods, to motivate the need for a different approach. We propose a novel method that uses a set of independently trained, uni-modal, deterministic autoencoders. Individual latent variables are concatenated into a common latent space, which is fed to a masked diffusion model to enable generative modeling. We also introduce a new multi-time training method to learn the conditional score network for multi-modal diffusion. Our methodology substantially outperforms competitors in both generation quality and coherence, as shown through an extensive experimental campaign.

Generative Models (GMs) have attracted considerable attention due to their tremendous success in various domains, such as computer vision where they are capable to generate impressive realistic-looking images. Likelihood-based GMs are attractive due to the possibility to generate new data by a single model evaluation. However, they typically achieve lower sample quality compared to state-of-the-art score-based diffusion models (DMs). This paper provides a significant step in the direction of addressing this limitation. The idea is to borrow one of the strengths of score-based DMs, which is the ability to perform accurate density estimation in low-density regions and to address manifold overfitting by means of data mollification. We connect data mollification through the addition of Gaussian noise to Gaussian homotopy, which is a well-known technique to improve optimization. Data mollification can be implemented by adding one line of code in the optimization loop, and we demonstrate that this provides a boost in generation quality of likelihood-based GMs, without computational overheads. We report results on image data sets with popular likelihood-based GMs, including variants of variational autoencoders and normalizing flows, showing large improvements in FID score., Comment: NeurIPS 2023

We introduce Functional Diffusion Processes (FDPs), which generalize score-based diffusion models to infinite-dimensional function spaces. FDPs require a new mathematical framework to describe the forward and backward dynamics, and several extensions to derive practical training objectives. These include infinite-dimensional versions of Girsanov theorem, in order to be able to compute an ELBO, and of the sampling theorem, in order to guarantee that functional evaluations in a countable set of points are equivalent to infinite-dimensional functions. We use FDPs to build a new breed of generative models in function spaces, which do not require specialized network architectures, and that can work with any kind of continuous data. Our results on real data show that FDPs achieve high-quality image generation, using a simple MLP architecture with orders of magnitude fewer parameters than existing diffusion models., Comment: Under review

While the promises of Multi-Task Learning (MTL) are attractive, characterizing the conditions of its success is still an open problem in Deep Learning. Some tasks may benefit from being learned together while others may be detrimental to one another. From a task perspective, grouping cooperative tasks while separating competing tasks is paramount to reap the benefits of MTL, i.e., reducing training and inference costs. Therefore, estimating task affinity for joint learning is a key endeavor. Recent work suggests that the training conditions themselves have a significant impact on the outcomes of MTL. Yet, the literature is lacking of a benchmark to assess the effectiveness of tasks affinity estimation techniques and their relation with actual MTL performance. In this paper, we take a first step in recovering this gap by (i) defining a set of affinity scores by both revisiting contributions from previous literature as well presenting new ones and (ii) benchmarking them on the Taskonomy dataset. Our empirical campaign reveals how, even in a small-scale scenario, task affinity scoring does not correlate well with actual MTL performance. Yet, some metrics can be more indicative than others., Comment: 7 pages. AAAI'23 - 2nd International Workshop on Practical Deep Learning in the Wild

Score-based diffusion models are a class of generative models whose dynamics is described by stochastic differential equations that map noise into data. While recent works have started to lay down a theoretical foundation for these models, an analytical understanding of the role of the diffusion time T is still lacking. Current best practice advocates for a large T to ensure that the forward dynamics brings the diffusion sufficiently close to a known and simple noise distribution; however, a smaller value of T should be preferred for a better approximation of the score-matching objective and higher computational efficiency. Starting from a variational interpretation of diffusion models, in this work we quantify this trade-off, and suggest a new method to improve quality and efficiency of both training and sampling, by adopting smaller diffusion times. Indeed, we show how an auxiliary model can be used to bridge the gap between the ideal and the simulated forward dynamics, followed by a standard reverse diffusion process. Empirical results support our analysis; for image data, our method is competitive w.r.t. the state-of-the-art, according to standard sample quality metrics and log-likelihood.

When learning to act in a stochastic, partially observable environment, an intelligent agent should be prepared to anticipate a change in its belief of the environment state, and be capable of adapting its actions on-the-fly to changing conditions. As humans, we are able to form contingency plans when learning a task with the explicit aim of being able to correct errors in the initial control, and hence prove useful if ever there is a sudden change in our perception of the environment which requires immediate corrective action. This is especially the case for autonomous vehicles (AVs) navigating real-world situations where safety is paramount, and a strong ability to react to a changing belief about the environment is truly needed. In this paper we explore an end-to-end approach, from training to execution, for learning robust contingency plans and combining them with a hierarchical planner to obtain a robust agent policy in an autonomous navigation task where other vehicles' behaviours are unknown, and the agent's belief about these behaviours is subject to sudden, last-second change. We show that our approach results in robust, safe behaviour in a partially observable, stochastic environment, generalizing well over environment dynamics not seen during training., Comment: To appear in Generalizable Policy Learning in the Physical World workshop (ICLR 2022)

When learning to behave in a stochastic environment where safety is critical, such as driving a vehicle in traffic, it is natural for human drivers to plan fallback strategies as a backup to use if ever there is an unexpected change in the environment. Knowing to expect the unexpected, and planning for such outcomes, increases our capability for being robust to unseen scenarios and may help prevent catastrophic failures. Control of Autonomous Vehicles (AVs) has a particular interest in knowing when and how to use fallback strategies in the interest of safety. Due to imperfect information available to an AV about its environment, it is important to have alternate strategies at the ready which might not have been deduced from the original training data distribution. In this paper we present a principled approach for a model-free Reinforcement Learning (RL) agent to capture multiple modes of behaviour in an environment. We introduce an extra pseudo-reward term to the reward model, to encourage exploration to areas of state-space different from areas privileged by the optimal policy. We base this reward term on a distance metric between the trajectories of agents, in order to force policies to focus on different areas of state-space than the initial exploring agent. Throughout the paper, we refer to this particular training paradigm as learning fallback strategies. We apply this method to an autonomous driving scenario, and show that we are able to learn useful policies that would have otherwise been missed out on during training, and unavailable to use when executing the control algorithm., Comment: To appear in proceedings of International Conference on Machine Learning Technologies (ICMLT) 2022

Anomaly detection in time series is a complex task that has been widely studied. In recent years, the ability of unsupervised anomaly detection algorithms has received much attention. This trend has led researchers to compare only learning-based methods in their articles, abandoning some more conventional approaches. As a result, the community in this field has been encouraged to propose increasingly complex learning-based models mainly based on deep neural networks. To our knowledge, there are no comparative studies between conventional, machine learning-based and, deep neural network methods for the detection of anomalies in multivariate time series. In this work, we study the anomaly detection performance of sixteen conventional, machine learning-based and, deep neural network approaches on five real-world open datasets. By analyzing and comparing the performance of each of the sixteen methods, we show that no family of methods outperforms the others. Therefore, we encourage the community to reincorporate the three categories of methods in the anomaly detection in multivariate time series benchmarks.

In Multi-Task Learning (MTL), it is a common practice to train multi-task networks by optimizing an objective function, which is a weighted average of the task-specific objective functions. Although the computational advantages of this strategy are clear, the complexity of the resulting loss landscape has not been studied in the literature. Arguably, its optimization may be more difficult than a separate optimization of the constituting task-specific objectives. In this work, we investigate the benefits of such an alternative, by alternating independent gradient descent steps on the different task-specific objective functions and we formulate a novel way to combine this approach with state-of-the-art optimizers. As the separation of task-specific objectives comes at the cost of increased computational time, we propose a random task grouping as a trade-off between better optimization and computational efficiency. Experimental results over three well-known visual MTL datasets show better overall absolute performance on losses and standard metrics compared to an averaged objective function and other state-of-the-art MTL methods. In particular, our method shows the most benefits when dealing with tasks of different nature and it enables a wider exploration of the shared parameter space. We also show that our random grouping strategy allows to trade-off between these benefits and computational efficiency.

We revisit the theoretical properties of Hamiltonian stochastic differential equations (SDES) for Bayesian posterior sampling, and we study the two types of errors that arise from numerical SDE simulation: the discretization error and the error due to noisy gradient estimates in the context of data subsampling. Our main result is a novel analysis for the effect of mini-batches through the lens of differential operator splitting, revising previous literature results. The stochastic component of a Hamiltonian SDE is decoupled from the gradient noise, for which we make no normality assumptions. This leads to the identification of a convergence bottleneck: when considering mini-batches, the best achievable error rate is $\mathcal{O}(\eta^2)$, with $\eta$ being the integrator step size. Our theoretical results are supported by an empirical study on a variety of regression and classification tasks for Bayesian neural networks.

We develop a novel method for carrying out model selection for Bayesian autoencoders (BAEs) by means of prior hyper-parameter optimization. Inspired by the common practice of type-II maximum likelihood optimization and its equivalence to Kullback-Leibler divergence minimization, we propose to optimize the distributional sliced-Wasserstein distance (DSWD) between the output of the autoencoder and the empirical data distribution. The advantages of this formulation are that we can estimate the DSWD based on samples and handle high-dimensional problems. We carry out posterior estimation of the BAE parameters via stochastic gradient Hamiltonian Monte Carlo and turn our BAE into a generative model by fitting a flexible Dirichlet mixture model in the latent space. Consequently, we obtain a powerful alternative to variational autoencoders, which are the preferred choice in modern applications of autoencoders for representation learning with uncertainty. We evaluate our approach qualitatively and quantitatively using a vast experimental campaign on a number of unsupervised learning tasks and show that, in small-data regimes where priors matter, our approach provides state-of-the-art results, outperforming multiple competitive baselines.

Approximations to Gaussian processes based on inducing variables, combined with variational inference techniques, enable state-of-the-art sparse approaches to infer GPs at scale through mini batch-based learning. In this work, we address one limitation of sparse GPs, which is due to the challenge in dealing with a large number of inducing variables without imposing a special structure on the inducing inputs. In particular, we introduce a novel hierarchical prior, which imposes sparsity on the set of inducing variables. We treat our model variationally, and we experimentally show considerable computational gains compared to standard sparse GPs when sparsity on the inducing variables is realized considering the nearest inducing inputs of a random mini-batch of the data. We perform an extensive experimental validation that demonstrates the effectiveness of our approach compared to the state-of-the-art. Our approach enables the possibility to use sparse GPs using a large number of inducing points without incurring a prohibitive computational cost., Comment: 10 pages

A large part of the literature on learning disentangled representations focuses on variational autoencoders (VAE). Recent developments demonstrate that disentanglement cannot be obtained in a fully unsupervised setting without inductive biases on models and data. However, Khemakhem et al., AISTATS, 2020 suggest that employing a particular form of factorized prior, conditionally dependent on auxiliary variables complementing input observations, can be one such bias, resulting in an identifiable model with guarantees on disentanglement. Working along this line, we propose a novel VAE-based generative model with theoretical guarantees on identifiability. We obtain our conditional prior over the latents by learning an optimal representation, which imposes an additional strength on their regularization. We also extend our method to semi-supervised settings. Experimental results indicate superior performance with respect to state-of-the-art approaches, according to several established metrics proposed in the literature on disentanglement.

Multi-task learning has gained popularity due to the advantages it provides with respect to resource usage and performance. Nonetheless, the joint optimization of parameters with respect to multiple tasks remains an active research topic. Sub-partitioning the parameters between different tasks has proven to be an efficient way to relax the optimization constraints over the shared weights, may the partitions be disjoint or overlapping. However, one drawback of this approach is that it can weaken the inductive bias generally set up by the joint task optimization. In this work, we present a novel way to partition the parameter space without weakening the inductive bias. Specifically, we propose Maximum Roaming, a method inspired by dropout that randomly varies the parameter partitioning, while forcing them to visit as many tasks as possible at a regulated frequency, so that the network fully adapts to each update. We study the properties of our method through experiments on a variety of visual multi-task data sets. Experimental results suggest that the regularization brought by roaming has more impact on performance than usual partitioning optimization strategies. The overall method is flexible, easily applicable, provides superior regularization and consistently achieves improved performances compared to recent multi-task learning formulations., Comment: Accepted at the 35th AAAI Conference on Artificial Intelligence (AAAI 2021)

In this work we define a unified mathematical framework to deepen our understanding of the role of stochastic gradient (SG) noise on the behavior of Markov chain Monte Carlo sampling (SGMCMC) algorithms. Our formulation unlocks the design of a novel, practical approach to posterior sampling, which makes the SG noise isotropic using a fixed learning rate that we determine analytically, and that requires weaker assumptions than existing algorithms. In contrast, the common traits of existing \sgmcmc algorithms is to approximate the isotropy condition either by drowning the gradients in additive noise (annealing the learning rate) or by making restrictive assumptions on the \sg noise covariance and the geometry of the loss landscape. Extensive experimental validations indicate that our proposal is competitive with the state-of-the-art on \sgmcmc, while being much more practical to use.

In this paper, we employ variational arguments to establish a connection between ensemble methods for Neural Networks and Bayesian inference. We consider an ensemble-based scheme where each model/particle corresponds to a perturbation of the data by means of parametric bootstrap and a perturbation of the prior. We derive conditions under which any optimization steps of the particles makes the associated distribution reduce its divergence to the posterior over model parameters. Such conditions do not require any particular form for the approximation and they are purely geometrical, giving insights on the behavior of the ensemble on a number of interesting models such as Neural Networks with ReLU activations. Experiments confirm that ensemble methods can be a valid alternative to approximate Bayesian inference; the theoretical developments in the paper seek to explain this behavior.

Accurate travel products price forecasting is a highly desired feature that allows customers to take informed decisions about purchases, and companies to build and offer attractive tour packages. Thanks to machine learning (ML), it is now relatively cheap to develop highly accurate statistical models for price time-series forecasting. However, once models are deployed in production, it is their monitoring, maintenance and improvement which carry most of the costs and difficulties over time. We introduce a data-driven framework to continuously monitor and maintain deployed time-series forecasting models' performance, to guarantee stable performance of travel products price forecasting models. Under a supervised learning approach, we predict the errors of time-series forecasting models over time, and use this predicted performance measure to achieve both model monitoring and maintenance. We validate the proposed method on a dataset of 18K time-series from flight and hotel prices collected over two years and on two public benchmarks.

We present a novel method - LIBRE - to learn an interpretable classifier, which materializes as a set of Boolean rules. LIBRE uses an ensemble of bottom-up weak learners operating on a random subset of features, which allows for the learning of rules that generalize well on unseen data even in imbalanced settings. Weak learners are combined with a simple union so that the final ensemble is also interpretable. Experimental results indicate that LIBRE efficiently strikes the right balance between prediction accuracy, which is competitive with black box methods, and interpretability, which is often superior to alternative methods from the literature.

Large scale machine learning is increasingly relying on distributed optimization, whereby several machines contribute to the training process of a statistical model. In this work we study the performance of asynchronous, distributed settings, when applying sparsification, a technique used to reduce communication overheads. In particular, for the first time in an asynchronous, non-convex setting, we theoretically prove that, in presence of staleness, sparsification does not harm SGD performance: the ergodic convergence rate matches the known result of standard SGD, that is $\mathcal{O} \left( 1/\sqrt{T} \right)$. We also carry out an empirical study to complement our theory, and confirm that the effects of sparsification on the convergence rate are negligible, when compared to 'vanilla' SGD, even in the challenging scenario of an asynchronous, distributed system.

The identification of anomalies in temporal data is a core component of numerous research areas such as intrusion detection, fault prevention, genomics and fraud detection. This article provides an experimental comparison of the novelty detection problem applied to discrete sequences. The objective of this study is to identify which state-of-the-art methods are efficient and appropriate candidates for a given use case. These recommendations rely on extensive novelty detection experiments based on a variety of public datasets in addition to novel industrial datasets. We also perform thorough scalability and memory usage tests resulting in new supplementary insights of the methods' performance, key selection criterion to solve problems relying on large volumes of data and to meet the expectations of applications subject to strict response time constraints., Comment: Submitted to Artificial Intelligence Review journal; 24 pages, 4 tables, 11 figures

Stochastic variational inference is an established way to carry out approximate Bayesian inference for deep models. While there have been effective proposals for good initializations for loss minimization in deep learning, far less attention has been devoted to the issue of initialization of stochastic variational inference. We address this by proposing a novel layer-wise initialization strategy based on Bayesian linear models. The proposed method is extensively validated on regression and classification tasks, including Bayesian DeepNets and ConvNets, showing faster and better convergence compared to alternatives inspired by the literature on initializations for loss minimization., Comment: 8 pages of main paper (+3 for references and +6 of supplement material)

Nowadays, data-centers are largely under-utilized because resource allocation is based on reservation mechanisms which ignore actual resource utilization. Indeed, it is common to reserve resources for peak demand, which may occur only for a small portion of the application life time. As a consequence, cluster resources often go under-utilized. In this work, we propose a mechanism that improves cluster utilization, thus decreasing the average turnaround time, while preventing application failures due to contention in accessing finite resources such as RAM. Our approach monitors resource utilization and employs a data-driven approach to resource demand forecasting, featuring quantification of uncertainty in the predictions. Using demand forecast and its confidence, our mechanism modulates cluster resources assigned to running applications, and reduces the turnaround time by more than one order of magnitude while keeping application failures under control. Thus, tenants enjoy a responsive system and providers benefit from an efficient cluster utilization.

In this paper, we study the problem of deriving fast and accurate classification algorithms with uncertainty quantification. Gaussian process classification provides a principled approach, but the corresponding computational burden is hardly sustainable in large-scale problems and devising efficient alternatives is a challenge. In this work, we investigate if and how Gaussian process regression directly applied to the classification labels can be used to tackle this question. While in this case training time is remarkably faster, predictions need be calibrated for classification and uncertainty estimation. To this aim, we propose a novel approach based on interpreting the labels as the output of a Dirichlet distribution. Extensive experimental results show that the proposed approach provides essentially the same accuracy and uncertainty quantification of Gaussian process classification while requiring only a fraction of computational resources.

The wide adoption of Convolutional Neural Networks (CNNs) in applications where decision-making under uncertainty is fundamental, has brought a great deal of attention to the ability of these models to accurately quantify the uncertainty in their predictions. Previous work on combining CNNs with Gaussian processes (GPs) has been developed under the assumption that the predictive probabilities of these models are well-calibrated. In this paper we show that, in fact, current combinations of CNNs and GPs are miscalibrated. We proposes a novel combination that considerably outperforms previous approaches on this aspect, while achieving state-of-the-art performance on image classification tasks., Comment: 12 pages

In modern large-scale distributed systems, analytics jobs submitted by various users often share similar work, for example scanning and processing the same subset of data. Instead of optimizing jobs independently, which may result in redundant and wasteful processing, multi-query optimization techniques can be employed to save a considerable amount of cluster resources. In this work, we introduce a novel method combining in-memory cache primitives and multi-query optimization, to improve the efficiency of data-intensive, scalable computing frameworks. By careful selection and exploitation of common (sub)expressions, while satisfying memory constraints, our method transforms a batch of queries into a new, more efficient one which avoids unnecessary recomputations. To find feasible and efficient execution plans, our method uses a cost-based optimization formulation akin to the multiple-choice knapsack problem. Extensive experiments on a prototype implementation of our system show significant benefits of worksharing for both TPC-DS workloads and detailed micro-benchmarks., Comment: 12 pages + references, extended version

We consider elastic resource provisioning in the cloud, focusing on in-memory key-value stores used as caches. Our goal is to dynamically scale resources to the traffic pattern minimizing the overall cost, which includes not only the storage cost, but also the cost due to misses. In fact, a small variation on the cache miss ratio may have a significant impact on user perceived performance in modern web services, which in turn has an impact on the overall revenues for the content provider that uses those services. We propose and study a dynamic algorithm for TTL caches, which is able to obtain close-to-minimal costs. Since high-throughput caches require low complexity operations, we discuss a practical implementation of such a scheme requiring constant overhead per request independently from the cache size. We evaluate our solution with real-world traces collected from Akamai, and show that we are able to obtain a 17% decrease in the overall cost compared to a baseline static configuration.

We present Stocator, a high performance object store connector for Apache Spark, that takes advantage of object store semantics. Previous connectors have assumed file system semantics, in particular, achieving fault tolerance and allowing speculative execution by creating temporary files to avoid interference between worker threads executing the same task and then renaming these files. Rename is not a native object store operation; not only is it not atomic, but it is implemented using a costly copy operation and a delete. Instead our connector leverages the inherent atomicity of object creation, and by avoiding the rename paradigm it greatly decreases the number of operations on the object store as well as enabling a much simpler approach to dealing with the eventually consistent semantics typical of object stores. We have implemented Stocator and shared it in open source. Performance testing shows that it is as much as 18 times faster for write intensive workloads and performs as much as 30 times fewer operations on the object store than the legacy Hadoop connectors, reducing costs both for the client and the object storage service provider.

This work addresses the problem of scheduling user-defined analytic applications, which we define as high-level compositions of frameworks, their components, and the logic necessary to carry out work. The key idea in our application definition, is to distinguish classes of components, including rigid and elastic types: the first being required for an application to make progress, the latter contributing to reduced execution times. We show that the problem of scheduling such applications poses new challenges, which existing approaches address inefficiently. Thus, we present the design and evaluation of a novel, flexible heuristic to schedule analytic applications, that aims at high system responsiveness, by allocating resources efficiently. Our algorithm is evaluated using trace-driven simulations, with large-scale real system traces: our flexible scheduler outperforms a baseline approach across a variety of metrics, including application turnaround times, and resource allocation efficiency. We also present the design and evaluation of a full-fledged system, which we have called Zoe, that incorporates the ideas presented in this paper, and report concrete improvements in terms of efficiency and performance, with respect to prior generations of our system.

The composition of multiple Gaussian Processes as a Deep Gaussian Process (DGP) enables a deep probabilistic nonparametric approach to flexibly tackle complex machine learning problems with sound quantification of uncertainty. Existing inference approaches for DGP models have limited scalability and are notoriously cumbersome to construct. In this work, we introduce a novel formulation of DGPs based on random feature expansions that we train using stochastic variational inference. This yields a practical learning framework which significantly advances the state-of-the-art in inference for DGPs, and enables accurate quantification of uncertainty. We extensively showcase the scalability and performance of our proposal on several datasets with up to 8 million observations, and various DGP architectures with up to 30 hidden layers.

In this paper we address the problem of rule-based stream data cleaning, which sets stringent requirements on latency, rule dynamics and ability to cope with the unbounded nature of data streams. We design a system, called Bleach, which achieves real-time violation detection and data repair on a dirty data stream. Bleach relies on efficient, compact and distributed data structures to maintain the necessary state to repair data, using an incremental version of the equivalence class algorithm. Additionally, it supports rule dynamics and uses a "cumulative" sliding window operation to improve cleaning accuracy. We evaluate a prototype of Bleach using a TPC-DS derived dirty data stream and observe its high throughput, low latency and high cleaning accuracy, even with rule dynamics. Experimental results indicate superior performance of Bleach compared to a baseline system built on the micro-batch streaming paradigm.

As data sets grow in size, analytics applications struggle to get instant insight into large datasets. Modern applications involve heavy batch processing jobs over large volumes of data and at the same time require efficient ad-hoc interactive analytics on temporary data. Existing solutions, however, typically focus on one of these two aspects, largely ignoring the need for synergy between the two. Consequently, interactive queries need to re-iterate costly passes through the entire dataset (e.g., data loading) that may provide meaningful return on investment only when data is queried over a long period of time. In this paper, we propose DiNoDB, an interactive-speed query engine for ad-hoc queries on temporary data. DiNoDB avoids the expensive loading and transformation phase that characterizes both traditional RDBMSs and current interactive analytics solutions. It is tailored to modern workflows found in machine learning and data exploration use cases, which often involve iterations of cycles of batch and interactive analytics on data that is typically useful for a narrow processing window. The key innovation of DiNoDB is to piggyback on the batch processing phase the creation of metadata that DiNoDB exploits to expedite the interactive queries. Our experimental analysis demonstrates that DiNoDB achieves very good performance for a wide range of ad-hoc queries compared to alternatives %such as Hive, Stado, SparkSQL and Impala.

An increasing number of Analytics-as-a-Service solutions has recently seen the light, in the landscape of cloud-based services. These services allow flexible composition of compute and storage components, that create powerful data ingestion and processing pipelines. This work is a first attempt at an experimental evaluation of analytic application performance executed using a wide range of storage service configurations. We present an intuitive notion of data locality, that we use as a proxy to rank different service compositions in terms of expected performance. Through an empirical analysis, we dissect the performance achieved by analytic workloads and unveil problems due to the impedance mismatch that arise in some configurations. Our work paves the way to a better understanding of modern cloud-based analytic services and their performance, both for its end-users and their providers., Comment: Longer version of the paper in Submission at IEEE CLOUD'16

In this paper we propose an online approximate k-nn graph building algorithm, which is able to quickly update a k-nn graph using a flow of data points. One very important step of the algorithm consists in using the current distributed graph to search for the neighbors of a new node. Hence we also propose a distributed partitioning method based on balanced k-medoids clustering, that we use to optimize the distributed search process. Finally, we present the improved sequential search procedure that is used inside each partition. We also perform an experimental evaluation of the different algorithms, where we study the influence of the parameters and compare the result of our algorithms to existing state of the art. This experimental evaluation confirms that the fast online k-nn graph building algorithm produces a graph that is highly similar to the graph produced by an offline exhaustive algorithm, while it requires less similarity computations., Comment: Submitted to ACM SIGKDD 2016

By executing jobs serially rather than in parallel, size-based scheduling policies can shorten time needed to complete jobs; however, major obstacles to their applicability are fairness guarantees and the fact that job sizes are rarely known exactly a-priori. Here, we introduce the Pri family of size-based scheduling policies; Pri simulates any reference scheduler and executes jobs in the order of their simulated completion: we show that these schedulers give strong fairness guarantees, since no job completes later in Pri than in the reference policy. In addition, we introduce PSBS, a practical implementation of such a scheduler: it works online (i.e., without needing knowledge of jobs submitted in the future), it has an efficient O(log n) implementation and it allows setting priorities to jobs. Most importantly, unlike earlier size-based policies, the performance of PSBS degrades gracefully with errors, leading to performances that are close to optimal in a variety of realistic use cases.

Size-based schedulers have very desirable performance properties: optimal or near-optimal response time can be coupled with strong fairness guarantees. Despite this, such systems are very rarely implemented in practical settings, because they require knowing a priori the amount of work needed to complete jobs: this assumption is very difficult to satisfy in concrete systems. It is definitely more likely to inform the system with an estimate of the job sizes, but existing studies point to somewhat pessimistic results if existing scheduler policies are used based on imprecise job size estimations. We take the goal of designing scheduling policies that are explicitly designed to deal with inexact job sizes: first, we show that existing size-based schedulers can have bad performance with inexact job size information when job sizes are heavily skewed; we show that this issue, and the pessimistic results shown in the literature, are due to problematic behavior when large jobs are underestimated. Once the problem is identified, it is possible to amend existing size-based schedulers to solve the issue. We generalize FSP -- a fair and efficient size-based scheduling policy -- in order to solve the problem highlighted above; in addition, our solution deals with different job weights (that can be assigned to a job independently from its size). We provide an efficient implementation of the resulting protocol, which we call Practical Size-Based Scheduler (PSBS). Through simulations evaluated on synthetic and real workloads, we show that PSBS has near-optimal performance in a large variety of cases with inaccurate size information, that it performs fairly and it handles correctly job weights. We believe that this work shows that PSBS is indeed pratical, and we maintain that it could inspire the design of schedulers in a wide array of real-world use cases., Comment: arXiv admin note: substantial text overlap with arXiv:1403.5996

User connectivity patterns in network applications are known to be heterogeneous, and to follow periodic (daily and weekly) patterns. In many cases, the regularity and the correlation of those patterns is problematic: for network applications, many connected users create peaks of demand; in contrast, in peer-to-peer scenarios, having few users online results in a scarcity of available resources. On the other hand, since connectivity patterns exhibit a periodic behavior, they are to some extent predictable. This work shows how this can be exploited to anticipate future user connectivity and to have applications proactively responding to it. We evaluate the probability that any given user will be online at any given time, and assess the prediction on six-month availability traces from three different Internet applications. Building upon this, we show how our probabilistic approach makes it easy to evaluate and optimize the performance in a number of diverse network application models, and to use them to optimize systems. In particular, we show how this approach can be used in distributed hash tables, friend-to-friend storage, and cache pre-loading for social networks, resulting in substantial gains in data availability and system efficiency at negligible costs., Comment: Accepted for publication in IEEE/ACM Transactions on Networking

We study size-based schedulers, and focus on the impact of inaccurate job size information on response time and fairness. Our intent is to revisit previous results, which allude to performance degradation for even small errors on job size estimates, thus limiting the applicability of size-based schedulers. We show that scheduling performance is tightly connected to workload characteristics: in the absence of large skew in the job size distribution, even extremely imprecise estimates suffice to outperform size-oblivious disciplines. Instead, when job sizes are heavily skewed, known size-based disciplines suffer. In this context, we show -- for the first time -- the dichotomy of over-estimation versus under-estimation. The former is, in general, less problematic than the latter, as its effects are localized to individual jobs. Instead, under-estimation leads to severe problems that may affect a large number of jobs. We present an approach to mitigate these problems: our technique requires no complex modifications to original scheduling policies and performs very well. To support our claim, we proceed with a simulation-based evaluation that covers an unprecedented large parameter space, which takes into account a variety of synthetic and real workloads. As a consequence, we show that size-based scheduling is practical and outperforms alternatives in a wide array of use-cases, even in presence of inaccurate size information., Comment: To be published in the proceedings of IEEE MASCOTS 2014

This work introduces a new task preemption primitive for Hadoop, that allows tasks to be suspended and resumed exploiting existing memory management mechanisms readily available in modern operating systems. Our technique fills the gap that exists between the two extremes cases of killing tasks (which waste work) or waiting for their completion (which introduces latency): experimental results indicate superior performance and very small overheads when compared to existing alternatives.

We present the Hadoop Fair Sojourn Protocol (HFSP) scheduler, which implements a size-based scheduling discipline for Hadoop. The benefits of size-based scheduling disciplines are well recognized in a variety of contexts (computer networks, operating systems, etc...), yet, their practical implementation for a system such as Hadoop raises a number of important challenges. With HFSP, which is available as an open-source project, we address issues related to job size estimation, resource management and study the effects of a variety of preemption strategies. Although the architecture underlying HFSP is suitable for any size-based scheduling discipline, in this work we revisit and extend the Fair Sojourn Protocol, which solves problems related to job starvation that affect FIFO, Processor Sharing and a range of size-based disciplines. Our experiments, in which we compare HFSP to standard Hadoop schedulers, pinpoint at a significant decrease in average job sojourn times - a metric that accounts for the total time a job spends in the system, including waiting and serving times - for realistic workloads that we generate according to production traces available in literature., Comment: 12 pages, 8 figures