Pre-training by language modeling has become a popular and successful approach to NLP tasks, but we have yet to understand exactly what linguistic capacities these pre-training processes confer upon models. In this paper we introduce a suite of diagnostics drawn from human language experiments, which allow us to ask targeted questions about the information used by language models for generating predictions in context. As a case study, we apply these diagnostics to the popular BERT model, finding that it can generally distinguish good from bad completions involving shared category or role reversal, albeit with less sensitivity than humans, and it robustly retrieves noun hypernyms, but it struggles with challenging inferences and role-based event prediction -- and in particular, it shows clear insensitivity to the contextual impacts of negation.

We propose the Graph Space Embedding (GSE), a technique that maps the input into a space where interactions are implicitly encoded, with little computations required. We provide theoretical results on an optimal regime for the GSE, namely a feasibility region for its parameters, and demonstrate the experimental relevance of our findings. Next, we introduce a strategy to gain insight on which interactions are responsible for the certain predictions, paving the way for a far more transparent model. In an empirical evaluation on a real-world clinical cohort containing patients with suspected coronary artery disease, the GSE achieves far better performance than traditional algorithms.

Matrix completion aims to predict missing elements in a partially observed data matrix which in typical applications, such as collaborative filtering, is large and extremely sparsely observed. A standard solution is matrix factorization, which predicts unobserved entries as linear combinations of latent variables. We generalize to nonlinear combinations in massive-scale matrices. Bayesian approaches have been proven beneficial in linear matrix completion, but not applied in the more general nonlinear case, due to limited scalability. We introduce a Bayesian nonlinear matrix completion algorithm, which is based on a recent Bayesian formulation of Gaussian process latent variable models. To solve the challenges regarding scalability and computation, we propose a data-parallel distributed computational approach with a restricted communication scheme. We evaluate our method on challenging out-of-matrix prediction tasks using both simulated and real-world data. 1 Introduction In matrix completion--one of the most widely used approaches for collaborative filtering--the objective is to predict missing elements of a partially observed data matrix.

N-grams have been a common tool for information retrieval and machine learning applications for decades. In nearly all previous works, only a few values of $n$ are tested, with $n > 6$ being exceedingly rare. Larger values of $n$ are not tested due to computational burden or the fear of overfitting. In this work, we present a method to find the top-$k$ most frequent $n$-grams that is 60$\times$ faster for small $n$, and can tackle large $n\geq1024$. Despite the unprecedented size of $n$ considered, we show how these features still have predictive ability for malware classification tasks. More important, large $n$-grams provide benefits in producing features that are interpretable by malware analysis, and can be used to create general purpose signatures compatible with industry standard tools like Yara. Furthermore, the counts of common $n$-grams in a file may be added as features to publicly available human-engineered features that rival efficacy of professionally-developed features when used to train gradient-boosted decision tree models on the EMBER dataset.

In much of the literature on function approximation by deep networks, the function is assumed to be defined on some known domain, such as a cube or sphere. In practice, the data might not be dense on these domains, and therefore, the approximation theory results are observed to be too conservative. In manifold learning, one assumes instead that the data is sampled from an unknown manifold; i.e., the manifold is defined by the data itself. Function approximation on this unknown manifold is then a two stage procedure: first, one approximates the Laplace-Beltrami operator (and its eigen-decomposition) on this manifold using a graph Laplacian, and next, approximates the target function using the eigen-functions. In this paper, we propose a more direct approach to function approximation on unknown, data defined manifolds without computing the eigen-decomposition of some operator, and estimate the degree of approximation in terms of the manifold dimension. This leads to similar results in function approximation using deep networks where each channel evaluates a Gaussian network on a possibly unknown manifold.

Devices comprising the Internet of Things, such as sensors and small cameras, usually have small memories and limited computational power. The proliferation of such resource-constrained devices in recent years has led to the generation of large quantities of data. These data-producing devices are appealing targets for machine learning applications but struggle to run machine learning algorithms due to their limited computing capability. They typically offload input data to external computing systems (such as cloud servers) for further processing. The results of the machine learning computations are communicated back to the resource-scarce devices, but this worsens latency, leads to increased communication costs, and adds to privacy concerns. Therefore, efforts have been made to place additional computing devices at the edge of the network, i.e close to the IoT devices where the data is generated. Deploying machine learning systems on such edge devices alleviates the above issues by allowing computations to be performed close to the data sources. This survey describes major research efforts where machine learning has been deployed at the edge of computer networks.

Many recent methods for unsupervised or self-supervised representation learning train feature extractors by maximizing an estimate of the mutual information (MI) between different views of the data. This comes with several immediate problems: For example, MI is notoriously hard to estimate, and using it as an objective for representation learning may lead to highly entangled representations due to its invariance under arbitrary invertible transformations. Nevertheless, these methods have been repeatedly shown to excel in practice. In this paper we argue, and provide empirical evidence, that the success of these methods might be only loosely attributed to the properties of MI, and that they strongly depend on the inductive bias in both the choice of feature extractor architectures and the parametrization of the employed MI estimators. Finally, we establish a connection to deep metric learning and argue that this interpretation may be a plausible explanation for the success of the recently introduced methods.

Control policies, trained using the Deep Reinforcement Learning, have been recently shown to be vulnerable to adversarial attacks introducing even very small perturbations to the policy input. The attacks proposed so far have been designed using heuristics, and build on existing adversarial example crafting techniques used to dupe classifiers in supervised learning. In contrast, this paper investigates the problem of devising optimal attacks, depending on a well-defined attacker's objective, e.g., to minimize the main agent average reward. When the policy and the system dynamics, as well as rewards, are known to the attacker, a scenario referred to as a white-box attack, designing optimal attacks amounts to solving a Markov Decision Process. For what we call black-box attacks, where neither the policy nor the system is known, optimal attacks can be trained using Reinforcement Learning techniques. Through numerical experiments, we demonstrate the efficiency of our attacks compared to existing attacks (usually based on Gradient methods). We further quantify the potential impact of attacks and establish its connection to the smoothness of the policy under attack. Smooth policies are naturally less prone to attacks (this explains why Lipschitz policies, with respect to the state, are more resilient). Finally, we show that from the main agent perspective, the system uncertainties and the attacker can be modeled as a Partially Observable Markov Decision Process. We actually demonstrate that using Reinforcement Learning techniques tailored to POMDP (e.g. using Recurrent Neural Networks) leads to more resilient policies.

A novel framework of the fuzzy c-means distances problem based weighted distance Andy Arief Setyawan a,1,, Ahmad Ilham b,1 a Department of Information and Communication, Pemalang District Government, Pemalang, Indonesia b Department of Informatics, Universitas Muhammadiyah Semarang, Semarang 50354, Indonesia Abstract Clustering is one of the major roles in data mining that is widely application in pattern recognition and image segmentation. Fuzzy C-means (FCM) is the most used clustering algorithm that proven efficient, fast and easy to implement, however FCM uses the Euclidean distance that often leads to clustering errors, especially when handling multidimensional and noisy data. In the last few years, many distances metric have been propose by researchers to improve the performance of the FCM algorithms, and the majority of researchers propose weighted distance. In this paper, we proposed Canberra Weighted Distance to improved performance of the FCM algorithm. Experimental result using the UCI data set show the proposed method is superior to the original method and other clustering methods. Keywords: clustering, fuzzy c-means, euclidean distance, weighted distance, canberra distance 1. Introduction Cluster analysis or clustering is the process of partitioning a set of data objects into subset or clusters, where the objects in a cluster is similar to onenull This document is a collaborative effort by Intelligent Systems Research Group Indonesia and Informatics Department Universitas Muhammadiyah Semarang.

Many versions of cross-validation (CV) exist in the literature; and each version though has different variants. All are used interchangeably by many practitioners; yet, without explanation to the connection or difference among them. This article has three contributions. First, it starts by mathematical formalization of these different versions and variants that estimate the error rate and the Area Under the ROC Curve (AUC) of a classification rule, to show the connection and difference among them. Second, we prove some of their properties and prove that many variants are either redundant or "not smooth". Hence, we suggest to abandon all redundant versions and variants and only keep the leave-one-out, the $K$-fold, and the repeated $K$-fold. We show that the latter is the only among the three versions that is "smooth" and hence looks mathematically like estimating the mean performance of the classification rules. However, empirically, for the known phenomenon of "weak correlation", which we explain mathematically and experimentally, it estimates both conditional and mean performance almost with the same accuracy. Third, we conclude the article with suggesting two research points that may answer the remaining question of whether we can come up with a finalist among the three estimators: (1) a comparative study, that is much more comprehensive than those available in literature and conclude no overall winner, is needed to consider a wide range of distributions, datasets, and classifiers including complex ones obtained via the recent deep learning approach. (2) we sketch the path of deriving a rigorous method for estimating the variance of the only "smooth" version, repeated $K$-fold CV, rather than those ad-hoc methods available in the literature that ignore the covariance structure among the folds of CV.