Out-of-Distribution (OoD) detection is important for building safe artificial intelligence systems. However, current OoD detection methods still cannot meet the performance requirements for practical deployment. In this paper, we propose a simple yet effective algorithm based on a novel observation: in a trained neural network, OoD samples with bounded norms well concentrate in the feature space. We call the center of OoD features the Feature Space Singularity (FSS), and denote the distance of a sample feature to FSS as FSSD. Then, OoD samples can be identified by taking a threshold on the FSSD. Our analysis of the phenomenon reveals why our algorithm works. We demonstrate that our algorithm achieves state-of-the-art performance on various OoD detection benchmarks. Besides, FSSD also enjoys robustness to slight corruption in test data and can be further enhanced by ensembling. These make FSSD a promising algorithm to be employed in real world. We release our code at \url{https://github.com/megvii-research/FSSD_OoD_Detection}.

A trained ML model is deployed on another `test' dataset where target feature values (labels) are unknown. Drift is distribution change between the training and deployment data, which is concerning if model performance changes. For a cat/dog image classifier, for instance, drift during deployment could be rabbit images (new class) or cat/dog images with changed characteristics (change in distribution). We wish to detect these changes but can't measure accuracy without deployment data labels. We instead detect drift indirectly by nonparametrically testing the distribution of model prediction confidence for changes. This generalizes our method and sidesteps domain-specific feature representation. We address important statistical issues, particularly Type-1 error control in sequential testing, using Change Point Models (CPMs; see Adams and Ross 2012). We also use nonparametric outlier methods to show the user suspicious observations for model diagnosis, since the before/after change confidence distributions overlap significantly. In experiments to demonstrate robustness, we train on a subset of MNIST digit classes, then insert drift (e.g., unseen digit class) in deployment data in various settings (gradual/sudden changes in the drift proportion). A novel loss function is introduced to compare the performance (detection delay, Type-1 and 2 errors) of a drift detector under different levels of drift class contamination.

Until recently, the most accurate methods for time series classification were limited by high computational complexity. ROCKET achieves state-of-the-art accuracy with a fraction of the computational expense of most existing methods by transforming input time series using random convolutional kernels, and using the transformed features to train a linear classifier. We reformulate ROCKET into a new method, MINIROCKET, making it up to 75 times faster on larger datasets, and making it almost deterministic (and optionally, with additional computational expense, fully deterministic), while maintaining essentially the same accuracy. Using this method, it is possible to train and test a classifier on all of 109 datasets from the UCR archive to state-of-the-art accuracy in less than 10 minutes. MINIROCKET is significantly faster than any other method of comparable accuracy (including ROCKET), and significantly more accurate than any other method of even roughly-similar computational expense. As such, we suggest that MINIROCKET should now be considered and used as the default variant of ROCKET.

The recent success of Bayesian methods in neuroscience and artificial intelligence gives rise to the hypothesis that the brain is a Bayesian machine. Since logic and learning are both practices of the human brain, it leads to another hypothesis that there is a Bayesian interpretation underlying both logical reasoning and machine learning. In this paper, we introduce a generative model of logical consequence relations. It formalises the process of how the truth value of a sentence is probabilistically generated from the probability distribution over states of the world. We show that the generative model characterises a classical consequence relation, paraconsistent consequence relation and nonmonotonic consequence relation. In particular, the generative model gives a new consequence relation that outperforms them in reasoning with inconsistent knowledge. We also show that the generative model gives a new classification algorithm that outperforms several representative algorithms in predictive accuracy and complexity on the Kaggle Titanic dataset.

Cardiovascular diseases (CVDs) are the main cause of deaths all over the world. Heart murmurs are the most common abnormalities detected during the auscultation process. The two widely used publicly available phonocardiogram (PCG) datasets are from the PhysioNet/CinC (2016) and PASCAL (2011) challenges. The datasets are significantly different in terms of the tools used for data acquisition, clinical protocols, digital storages and signal qualities, making it challenging to process and analyze. In this work, we have used short-time Fourier transform (STFT) based spectrograms to learn the representative patterns of the normal and abnormal PCG signals. Spectrograms generated from both the datasets are utilized to perform three different studies: (i) train, validate and test different variants of convolutional neural network (CNN) models with PhysioNet dataset, (ii) train, validate and test the best performing CNN structure on combined PhysioNet-PASCAL dataset and (iii) finally, transfer learning technique is employed to train the best performing pre-trained network from the first study with PASCAL dataset. We propose a novel, less complex and relatively light custom CNN model for the classification of PhysioNet, combined and PASCAL datasets. The first study achieves an accuracy, sensitivity, specificity, precision and F1 score of 95.4%, 96.3%, 92.4%, 97.6% and 96.98% respectively while the second study shows accuracy, sensitivity, specificity, precision and F1 score of 94.2%, 95.5%, 90.3%, 96.8% and 96.1% respectively. Finally, the third study shows a precision of 98.29% on the noisy PASCAL dataset with transfer learning approach. All the three proposed approaches outperform most of the recent competing studies by achieving comparatively high classification accuracy and precision, which make them suitable for screening CVDs using PCG signals.

Building reliable machine learning systems requires that we correctly understand their level of confidence. Calibration focuses on measuring the degree of accuracy in a model's confidence and most research in calibration focuses on techniques to improve an empirical estimate of calibration error, ECE_bin. Using simulation, we show that ECE_bin can systematically underestimate or overestimate the true calibration error depending on the nature of model miscalibration, the size of the evaluation data set, and the number of bins. Critically, ECE_bin is more strongly biased for perfectly calibrated models. We propose a simple alternative calibration error metric, ECE_sweep, in which the number of bins is chosen to be as large as possible while preserving monotonicity in the calibration function. Evaluating our measure on distributions fit to neural network confidence scores on CIFAR-10, CIFAR-100, and ImageNet, we show that ECE_sweep produces a less biased estimator of calibration error and therefore should be used by any researcher wishing to evaluate the calibration of models trained on similar datasets.

Analysis of water and environmental data is an important aspect of many intelligent water and environmental system applications where inference from such analysis plays a significant role in decision making. Quite often these data that are collected through sensible sensors can be anomalous due to different reasons such as systems breakdown, malfunctioning of sensor detectors, and more. Regardless of their root causes, such data severely affect the results of the subsequent analysis. This paper demonstrates data cleaning and preparation for time-series data and further proposes cost-sensitive machine learning algorithms as a solution to detect anomalous data points in time-series data. The following models: Logistic Regression, Random Forest, Support Vector Machines have been modified to support the cost-sensitive learning which penalizes misclassified samples thereby minimizing the total misclassification cost. Our results showed that Random Forest outperformed the rest of the models at predicting the positive class (i.e anomalies). Applying predictive model improvement techniques like data oversampling seems to provide little or no improvement to the Random Forest model. Interestingly, with recursive feature elimination, we achieved a better model performance thereby reducing the dimensions in the data. Finally, with Influxdb and Kapacitor the data was ingested and streamed to generate new data points to further evaluate the model performance on unseen data, this will allow for early recognition of undesirable changes in the drinking water quality and will enable the water supply companies to rectify on a timely basis whatever undesirable changes abound.

It's easy to do bad things with Facebook data. From targeting ads for bizarrely specific T-shirts to manipulating an electorate, the questionable purposes to which the social media behemoth can be put are numerous. But there are also some people out there trying to use Facebook for good--or, at least, to improve the diagnosis of mental illness. On December 3, a group of researchers reported that they had managed to predict psychiatric diagnoses with Facebook data--using messages sent up to 18 months before a user received an official diagnosis. The team worked with 223 volunteers, who all gave the researchers access to their personal Facebook messages.

When formulated as an unsupervised learning problem, anomaly detection often requires a model to learn the distribution of normal data. Previous works apply Generative Adversarial Networks (GANs) to anomaly detection tasks and show good performances from these models. Motivated by the observation that GAN ensembles often outperform single GANs in generation tasks, we propose to construct GAN ensembles for anomaly detection. In the proposed method, a group of generators and a group of discriminators are trained together, so every generator gets feedback from multiple discriminators, and vice versa. Compared to a single GAN, a GAN ensemble can better model the distribution of normal data and thus better detect anomalies. Our theoretical analysis of GANs and GAN ensembles explains the role of a GAN discriminator in anomaly detection. In the empirical study, we evaluate ensembles constructed from four types of base models, and the results show that these ensembles clearly outperform single models in a series of tasks of anomaly detection.

We present a machine learning framework and a new test bed for data mining from the Slurm Workload Manager for high-performance computing (HPC) clusters. The focus was to find a method for selecting features to support decisions: helping users decide whether to resubmit failed jobs with boosted CPU and memory allocations or migrate them to a computing cloud. This task was cast as both supervised classification and regression learning, specifically, sequential problem solving suitable for reinforcement learning. Selecting relevant features can improve training accuracy, reduce training time, and produce a more comprehensible model, with an intelligent system that can explain predictions and inferences. We present a supervised learning model trained on a Simple Linux Utility for Resource Management (Slurm) data set of HPC jobs using three different techniques for selecting features: linear regression, lasso, and ridge regression. Our data set represented both HPC jobs that failed and those that succeeded, so our model was reliable, less likely to overfit, and generalizable. Our model achieved an R^2 of 95\% with 99\% accuracy. We identified five predictors for both CPU and memory properties.