Privacy-preserving machine learning (PPML) solutions are gaining widespread popularity. Among these, many rely on homomorphic encryption (HE) that offers confidentiality of the model and the data, but at the cost of large latency and memory requirements. Pruning neural network (NN) parameters improves latency and memory in plaintext ML but has little impact if directly applied to HE-based PPML. We introduce a framework called HE-PEx that comprises new pruning methods, on top of a packing technique called tile tensors, for reducing the latency and memory of PPML inference. HE-PEx uses permutations to prune additional ciphertexts, and expansion to recover inference loss. We demonstrate the effectiveness of our methods for pruning fully-connected and convolutional layers in NNs on PPML tasks, namely, image compression, denoising, and classification, with autoencoders, multilayer perceptrons (MLPs) and convolutional neural networks (CNNs). We implement and deploy our networks atop a framework called HElayers, which shows a 10-35% improvement in inference speed and a 17-35% decrease in memory requirement over the unpruned network, corresponding to 33-65% fewer ciphertexts, within a 2.5% degradation in inference accuracy over the unpruned network. Compared to the state-of-the-art pruning technique for PPML, our techniques generate networks with 70% fewer ciphertexts, on average, for the same degradation limit.
Percolation is an important topic in climate, physics, materials science, epidemiology, finance, and so on. Prediction of percolation thresholds with machine learning methods remains challenging. In this paper, we build a powerful graph convolutional neural network to study the percolation in both supervised and unsupervised ways. From a supervised learning perspective, the graph convolutional neural network simultaneously and correctly trains data of different lattice types, such as the square and triangular lattices. For the unsupervised perspective, combining the graph convolutional neural network and the confusion method, the percolation threshold can be obtained by the "W" shaped performance. The finding of this work opens up the possibility of building a more general framework that can probe the percolation-related phenomenon.
Trustworthy machine learning is driving a large number of ML community works in order to improve ML acceptance and adoption. The main aspect of trustworthy machine learning are the followings: fairness, uncertainty, robustness, explainability and formal guaranties. Each of these individual domains gains the ML community interest, visible by the number of related publications. However few works tackle the interconnection between these fields. In this paper we show a first link between uncertainty and explainability, by studying the relation between calibration and interpretation. As the calibration of a given model changes the way it scores samples, and interpretation approaches often rely on these scores, it seems safe to assume that the confidence-calibration of a model interacts with our ability to interpret such model. In this paper, we show, in the context of networks trained on image classification tasks, to what extent interpretations are sensitive to confidence-calibration. It leads us to suggest a simple practice to improve the interpretation outcomes: Calibrate to Interpret.
When training a machine learning classifier on data where one of the classes is intrinsically rare, the classifier will often assign too few sources to the rare class. To address this, it is common to up-weight the examples of the rare class to ensure it isn't ignored. It is also a frequent practice to train on restricted data where the balance of source types is closer to equal for the same reason. Here we show that these practices can bias the model toward over-assigning sources to the rare class. We also explore how to detect when training data bias has had a statistically significant impact on the trained model's predictions, and how to reduce the bias's impact. While the magnitude of the impact of the techniques developed here will vary with the details of the application, for most cases it should be modest. They are, however, universally applicable to every time a machine learning classification model is used, making them analogous to Bessel's correction to the sample variance.
We present two approaches to determine the dynamical stability of a hierarchical triple-star system. The first is an improvement on the Mardling-Aarseth stability formula from 2001, where we introduce a dependence on inner orbital eccentricity and improve the dependence on mutual orbital inclination. The second involves a machine learning approach, where we use a multilayer perceptron (MLP) to classify triple-star systems as `stable' and `unstable'. To achieve this, we generate a large training data set of 10^6 hierarchical triples using the N-body code MSTAR. Both our approaches perform better than previous stability criteria, with the MLP model performing the best. The improved stability formula and the machine learning model have overall classification accuracies of 93 % and 95 % respectively. Our MLP model, which accurately predicts the stability of any hierarchical triple-star system within the parameter ranges studied with almost no computation required, is publicly available on Github in the form of an easy-to-use Python script.
Within the cultural heritage sector, there has been a growing and concerted effort to consider a critical sociotechnical lens when applying machine learning techniques to digital collections. Though the cultural heritage community has collectively developed an emerging body of work detailing responsible operations for machine learning in libraries and other cultural heritage institutions at the organizational level, there remains a paucity of guidelines created specifically for practitioners embarking on machine learning projects. The manifold stakes and sensitivities involved in applying machine learning to cultural heritage underscore the importance of developing such guidelines. This paper contributes to this need by formulating a detailed checklist with guiding questions and practices that can be employed while developing a machine learning project that utilizes cultural heritage data. I call the resulting checklist the "Collections as ML Data" checklist, which, when completed, can be published with the deliverables of the project. By surveying existing projects, including my own project, Newspaper Navigator, I justify the "Collections as ML Data" checklist and demonstrate how the formulated guiding questions can be employed and operationalized.
Machine learning has significantly enhanced the abilities of robots, enabling them to perform a wide range of tasks in human environments and adapt to our uncertain real world. Recent works in various machine learning domains have highlighted the importance of accounting for fairness to ensure that these algorithms do not reproduce human biases and consequently lead to discriminatory outcomes. With robot learning systems increasingly performing more and more tasks in our everyday lives, it is crucial to understand the influence of such biases to prevent unintended behavior toward certain groups of people. In this work, we present the first survey on fairness in robot learning from an interdisciplinary perspective spanning technical, ethical, and legal challenges. We propose a taxonomy for sources of bias and the resulting types of discrimination due to them. Using examples from different robot learning domains, we examine scenarios of unfair outcomes and strategies to mitigate them. We present early advances in the field by covering different fairness definitions, ethical and legal considerations, and methods for fair robot learning. With this work, we aim to pave the road for groundbreaking developments in fair robot learning.
To extract essential information from complex data, computer scientists have been developing machine learning models that learn low-dimensional representation mode. From such advances in machine learning research, not only computer scientists but also social scientists have benefited and advanced their research because human behavior or social phenomena lies in complex data. However, this emerging trend is not well documented because different social science fields rarely cover each other's work, resulting in fragmented knowledge in the literature. To document this emerging trend, we survey recent studies that apply word embedding techniques to human behavior mining. We built a taxonomy to illustrate the methods and procedures used in the surveyed papers, aiding social science researchers in contextualizing their research within the literature on word embedding applications. This survey also conducts a simple experiment to warn that common similarity measurements used in the literature could yield different results even if they return consistent results at an aggregate level.
We study trends in model size of notable machine learning systems over time using a curated dataset. From 1950 to 2018, model size in language models increased steadily by seven orders of magnitude. The trend then accelerated, with model size increasing by another five orders of magnitude in just 4 years from 2018 to 2022. Vision models grew at a more constant pace, totaling 7 orders of magnitude of growth between 1950 and 2022. We also identify that, since 2020, there have been many language models below 20B parameters, many models above 70B parameters, but a scarcity of models in the 20-70B parameter range. We refer to that scarcity as the parameter gap. We provide some stylized facts about the parameter gap and propose a few hypotheses to explain it. The explanations we favor are: (a) increasing model size beyond 20B parameters requires adopting different parallelism techniques, which makes mid-sized models less cost-effective, (b) GPT-3 was one order of magnitude larger than previous language models, and researchers afterwards primarily experimented with bigger models to outperform it. While these dynamics likely exist, and we believe they play some role in generating the gap, we don't have high confidence that there are no other, more important dynamics at play.
A tensor network is a type of decomposition used to express and approximate large arrays of data. A given data-set, quantum state or higher dimensional multi-linear map is factored and approximated by a composition of smaller multi-linear maps. This is reminiscent to how a Boolean function might be decomposed into a gate array: this represents a special case of tensor decomposition, in which the tensor entries are replaced by 0, 1 and the factorisation becomes exact. The collection of associated techniques are called, tensor network methods: the subject developed independently in several distinct fields of study, which have more recently become interrelated through the language of tensor networks. The tantamount questions in the field relate to expressability of tensor networks and the reduction of computational overheads. A merger of tensor networks with machine learning is natural. On the one hand, machine learning can aid in determining a factorization of a tensor network approximating a data set. On the other hand, a given tensor network structure can be viewed as a machine learning model. Herein the tensor network parameters are adjusted to learn or classify a data-set. In this survey we recover the basics of tensor networks and explain the ongoing effort to develop the theory of tensor networks in machine learning.
Datasets play a central role in the training and evaluation of machine learning (ML) models. But they are also the root cause of many undesired model behaviors, such as biased predictions. To overcome this situation, the ML community is proposing a data-centric cultural shift where data issues are given the attention they deserve, and more standard practices around the gathering and processing of datasets start to be discussed and established. So far, these proposals are mostly high-level guidelines described in natural language and, as such, they are difficult to formalize and apply to particular datasets. In this sense, and inspired by these proposals, we define a new domain-specific language (DSL) to precisely describe machine learning datasets in terms of their structure, data provenance, and social concerns. We believe this DSL will facilitate any ML initiative to leverage and benefit from this data-centric shift in ML (e.g., selecting the most appropriate dataset for a new project or better replicating other ML results). The DSL is implemented as a Visual Studio Code plugin, and it has been published under an open source license.
Adversarial attacks on state-of-the-art machine learning models pose a significant threat to the safety and security of mission-critical autonomous systems. This paper considers the additional vulnerability of machine learning models when attackers can measure the power consumption of their underlying hardware platform. In particular, we explore the utility of power consumption information for adversarial attacks on non-volatile memory crossbar-based single-layer neural networks. Our results from experiments with MNIST and CIFAR-10 datasets show that power consumption can reveal important information about the neural network's weight matrix, such as the 1-norm of its columns. That information can be used to infer the sensitivity of the network's loss with respect to different inputs. We also find that surrogate-based black box attacks that utilize crossbar power information can lead to improved attack efficiency.
Machine learning (ML) models are typically optimized for their accuracy on a given dataset. However, this predictive criterion rarely captures all desirable properties of a model, in particular how well it matches a domain expert's understanding of a task. Underspecification refers to the existence of multiple models that are indistinguishable in their in-domain accuracy, even though they differ in other desirable properties such as out-of-distribution (OOD) performance. Identifying these situations is critical for assessing the reliability of ML models. We formalize the concept of underspecification and propose a method to identify and partially address it. We train multiple models with an independence constraint that forces them to implement different functions. They discover predictive features that are otherwise ignored by standard empirical risk minimization (ERM), which we then distill into a global model with superior OOD performance. Importantly, we constrain the models to align with the data manifold to ensure that they discover meaningful features. We demonstrate the method on multiple datasets in computer vision (collages, WILDS-Camelyon17, GQA) and discuss general implications of underspecification. Most notably, in-domain performance cannot serve for OOD model selection without additional assumptions.
Nowadays, systems based on machine learning (ML) are widely used in different domains. Given their popularity, ML models have become targets for various attacks. As a result, research at the intersection of security/privacy and ML has flourished. Typically such work has focused on individual types of security/privacy concerns and mitigations thereof. However, in real-life deployments, an ML model will need to be protected against several concerns simultaneously. A protection mechanism optimal for one security or privacy concern may interact negatively with mechanisms intended to address other concerns. Despite its practical relevance, the potential for such conflicts has not been studied adequately. We first provide a framework for analyzing such "conflicting interactions". We then focus on systematically analyzing pairwise interactions between protection mechanisms for one concern, model and data ownership verification, with two other classes of ML protection mechanisms: differentially private training, and robustness against model evasion. We find that several pairwise interactions result in conflicts. We explore potential approaches for avoiding such conflicts. First, we study the effect of hyperparameter relaxations, finding that there is no sweet spot balancing the performance of both protection mechanisms. Second, we explore if modifying one type of protection mechanism (ownership verification) so as to decouple it from factors that may be impacted by a conflicting mechanism (differentially private training or robustness to model evasion) can avoid conflict. We show that this approach can avoid the conflict between ownership verification mechanisms when combined with differentially private training, but has no effect on robustness to model evasion. Finally, we identify the gaps in the landscape of studying interactions between other types of ML protection mechanisms.
An increasing body of work has recognized the importance of exploiting machine learning (ML) advancements to address the need for efficient automation in extracting access control attributes, policy mining, policy verification, access decisions, etc. In this work, we survey and summarize various ML approaches to solve different access control problems. We propose a novel taxonomy of the ML model's application in the access control domain. We highlight current limitations and open challenges such as lack of public real-world datasets, administration of ML-based access control systems, understanding a black-box ML model's decision, etc., and enumerate future research directions.
Machine learning based decision making systems applied in safety critical areas require reliable high certainty predictions. For this purpose, the system can be extended by an reject option which allows the system to reject inputs where only a prediction with an unacceptably low certainty would be possible. While being able to reject uncertain samples is important, it is also of importance to be able to explain why a particular sample was rejected. With the ongoing rise of eXplainable AI (XAI), a lot of explanation methodologies for machine learning based systems have been developed -- explaining reject options, however, is still a novel field where only very little prior work exists. In this work, we propose to explain rejects by semifactual explanations, an instance of example-based explanation methods, which them self have not been widely considered in the XAI community yet. We propose a conceptual modeling of semifactual explanations for arbitrary reject options and empirically evaluate a specific implementation on a conformal prediction based reject option.
We present an open-source differentiable acoustic simulator, j-Wave, which can solve time-varying and time-harmonic acoustic problems. It supports automatic differentiation, which is a program transformation technique that has many applications, especially in machine learning and scientific computing. j-Wave is composed of modular components that can be easily customized and reused. At the same time, it is compatible with some of the most popular machine learning libraries, such as JAX and TensorFlow. The accuracy of the simulation results for known configurations is evaluated against the widely used k-Wave toolbox and a cohort of acoustic simulation software. j-Wave is available from https://github.com/ucl-bug/jwave.
When training a machine learning model, there is likely to be a tradeoff between accuracy and the diversity of the dataset. Previous research has shown that if we train a model to detect one specific malware family, we generally obtain stronger results as compared to a case where we train a single model on multiple diverse families. However, during the detection phase, it would be more efficient to have a single model that can reliably detect multiple families, rather than having to score each sample against multiple models. In this research, we conduct experiments based on byte $n$-gram features to quantify the relationship between the generality of the training dataset and the accuracy of the corresponding machine learning models, all within the context of the malware detection problem. We find that neighborhood-based algorithms generalize surprisingly well, far outperforming the other machine learning techniques considered.
In the growing world of machine learning and data analytics, scholars are finding new and innovative ways to solve real-world problems. One solution comes by way of an intersection between healthcare, sports statistics, and data sciences. Within the realm of Major League Baseball (MLB), pitchers are regarded as the most important roster position. They often are among the highest paid players and are crucial to a franchise's success, but they are more at risk to suffer an injury that sidelines them for over a complete season. The ulnar collateral ligament (UCL) is a small ligament in the elbow that controls the strength and stability of a pitcher's throwing arm. Due to repetitive strain, it is not uncommon for pitchers to tear it partially or completely during their careers. Repairing this injury requires UCL reconstruction surgery, as known informally as Tommy John surgery. In this podium abstract, we want to investigate whether we can use machine learning techniques to predict UCL injury by analyzing online pitcher data.
Equilibrium systems are a powerful way to express neural computations. As special cases, they include models of great current interest in both neuroscience and machine learning, such as deep neural networks, equilibrium recurrent neural networks, deep equilibrium models, or meta-learning. Here, we present a new principle for learning such systems with a temporally- and spatially-local rule. Our principle casts learning as a least-control problem, where we first introduce an optimal controller to lead the system towards a solution state, and then define learning as reducing the amount of control needed to reach such a state. We show that incorporating learning signals within a dynamics as an optimal control enables transmitting activity-dependent credit assignment information, avoids storing intermediate states in memory, and does not rely on infinitesimal learning signals. In practice, our principle leads to strong performance matching that of leading gradient-based learning methods when applied to an array of problems involving recurrent neural networks and meta-learning. Our results shed light on how the brain might learn and offer new ways of approaching a broad class of machine learning problems.