In inverse problems, we often have access to data consisting of paired samples $(x,y)\sim p_{X,Y}(x,y)$ where $y$ are partial observations of a physical system, and $x$ represents the unknowns of the problem. Under these circumstances, we can employ supervised training to learn a solution $x$ and its uncertainty from the observations $y$. We refer to this problem as the "supervised" case. However, the data $y\sim p_{Y}(y)$ collected at one point could be distributed differently than observations $y'\sim p_{Y}'(y')$, relevant for a current set of problems. In the context of Bayesian inference, we propose a two-step scheme, which makes use of normalizing flows and joint data to train a conditional generator $q_{\theta}(x|y)$ to approximate the target posterior density $p_{X|Y}(x|y)$. Additionally, this preliminary phase provides a density function $q_{\theta}(x|y)$, which can be recast as a prior for the "unsupervised" problem, e.g.~when only the observations $y'\sim p_{Y}'(y')$, a likelihood model $y'|x$, and a prior on $x'$ are known. We then train another invertible generator with output density $q'_{\phi}(x|y')$ specifically for $y'$, allowing us to sample from the posterior $p_{X|Y}'(x|y')$. We present some synthetic results that demonstrate considerable training speedup when reusing the pretrained network $q_{\theta}(x|y')$ as a warm start or preconditioning for approximating $p_{X|Y}'(x|y')$, instead of learning from scratch. This training modality can be interpreted as an instance of transfer learning. This result is particularly relevant for large-scale inverse problems that employ expensive numerical simulations.

In the last decade, two paradigm shifts have reshaped the software industry - the move from boxed products to services and the widespread adoption of cloud computing. This has had a huge impact on the software development life cycle and the DevOps processes. Particularly, incident management has become critical for developing and operating large-scale services. Incidents are created to ensure timely communication of service issues and, also, their resolution. Prior work on incident management has been heavily focused on the challenges with incident triaging and de-duplication. In this work, we address the fundamental problem of structured knowledge extraction from service incidents. We have built SoftNER, a framework for unsupervised knowledge extraction from service incidents. We frame the knowledge extraction problem as a Named-entity Recognition task for extracting factual information. SoftNER leverages structural patterns like key,value pairs and tables for bootstrapping the training data. Further, we build a novel multi-task learning based BiLSTM-CRF model which leverages not just the semantic context but also the data-types for named-entity extraction. We have deployed SoftNER at Microsoft, a major cloud service provider and have evaluated it on more than 2 months of cloud incidents. We show that the unsupervised machine learning based approach has a high precision of 0.96. Our multi-task learning based deep learning model also outperforms the state of the art NER models. Lastly, using the knowledge extracted by SoftNER we are able to build significantly more accurate models for important downstream tasks like incident triaging.

Machine learning applied to generate data-driven models are lacking of transparency leading the process engineer to lose confidence in relying on the model predictions to optimize his industrial process. Bringing processes in the industry to a certain level of autonomy using data-driven models is particularly challenging as the first user of those models, is the expert in the process with often decades of experience. It is necessary to expose to the process engineer, not solely the model predictions, but also their interpretability. To that end, several approaches have been proposed in the literature. The Local Interpretable Model-agnostic Explanations (LIME) method has gained a lot of interest from the research community recently. The principle of this method is to train a linear model that is locally approximating the black-box model, by generating randomly artificial data points locally. Model-agnostic local interpretability solutions based on LIME have recently emerged to improve the original method. We present in this paper a novel approach, VAE-LIME, for local interpretability of data-driven models forecasting the temperature of the hot metal produced by a blast furnace. Such ironmaking process data is characterized by multivariate time series with high inter-correlation representing the underlying process in a blast furnace. Our contribution is to use a Variational Autoencoder (VAE) to learn the complex blast furnace process characteristics from the data. The VAE is aiming at generating optimal artificial samples to train a local interpretable model better representing the black-box model in the neighborhood of the input sample processed by the black-box model to make a prediction. In comparison with LIME, VAE-LIME is showing a significantly improved local fidelity of the local interpretable linear model with the black-box model resulting in robust model interpretability.

In this paper we study the problem of predicting clinical diagnoses from textual Electronic Health Records (EHR) data. We show the importance of this problem in medical community and present comprehensive historical review of the problem and proposed methods. As the main scientific contributions we present a modification of Bidirectional Encoder Representations from Transformers (BERT) model for sequence classification that implements a novel way of Fully-Connected (FC) layer composition and a BERT model pretrained only on domain data. To empirically validate our model, we use a large-scale Russian EHR dataset consisting of about 4 million unique patient visits. This is the largest such study for the Russian language and one of the largest globally. We performed a number of comparative experiments with other text representation models on the task of multiclass classification for 265 disease subset of ICD-10. The experiments demonstrate improved performance of our models compared to other baselines, including a fine-tuned Russian BERT (RuBERT) variant. We also show comparable performance of our model with a panel of experienced medical experts. This allows us to hope that implementation of this system will reduce misdiagnosis.

We propose a novel method to explain trained deep neural networks (DNNs), by distilling them into surrogate models using unsupervised clustering. Our method can be applied flexibly to any subset of layers of a DNN architecture and can incorporate low-level and high-level information. On image datasets given pre-trained DNNs, we demonstrate the strength of our method in finding similar training samples, and shedding light on the concepts the DNNs base their decisions on. Via user studies, we show that our model can improve the user trust in model's prediction.

Rescaling a vector $\vec{\delta} \in \mathbb{R}^n$ to a desired length is a common operation in many areas such as data science and machine learning. When the rescaled perturbation $\eta \vec{\delta}$ is added to a starting point $\vec{x} \in D$ (where $D$ is the data domain, e.g. $D = [0, 1]^n$), the resulting vector $\vec{v} = \vec{x} + \eta \vec{\delta}$ will in general not be in $D$. To enforce that the perturbed vector $v$ is in $D$, the values of $\vec{v}$ can be clipped to $D$. This subsequent element-wise clipping to the data domain does however reduce the effective perturbation size and thus interferes with the rescaling of $\vec{\delta}$. The optimal rescaling $\eta$ to obtain a perturbation with the desired norm after the clipping can be iteratively approximated using a binary search. However, such an iterative approach is slow and non-differentiable. Here we show that the optimal rescaling can be found analytically using a fast and differentiable algorithm. Our algorithm works for any p-norm and can be used to train neural networks on inputs with normalized perturbations. We provide native implementations for PyTorch, TensorFlow, JAX, and NumPy based on EagerPy.

Recently there was an increasing interest in applications of graph neural networks in non-Euclidean geometry; however, are non-Euclidean representations always useful for graph learning tasks? For different problems such as node classification and link prediction we compute hyperbolic embeddings and conclude that for tasks that require global prediction consistency it might be useful to use non-Euclidean embeddings, while for other tasks Euclidean models are superior. To do so we first fix an issue of the existing models associated with the optimization process at zero curvature. Current hyperbolic models deal with gradients at the origin in ad-hoc manner, which is inefficient and can lead to numerical instabilities. We solve the instabilities of kappa-Stereographic model at zero curvature cases and evaluate the approach of embedding graphs into the manifold in several graph representation learning tasks.

Information technology has spread widely, and extraordinarily large amounts of data have been made accessible to users, which has made it challenging to select data that are in accordance with user needs. For the resolution of the above issue, recommender systems have emerged, which much help users go through the process of decision-making and selecting relevant data. A recommender system predicts users behavior to be capable of detecting their interests and needs, and it often uses the classification technique for this purpose. It may not be sufficiently accurate to employ individual classification, where not all cases can be examined, which makes the method inappropriate to specific problems. In this research, group classification and the ensemble learning technique were used for increasing prediction accuracy in recommender systems. Another issue that is raised here concerns user analysis. Given the large size of the data and a large number of users, the process of user needs analysis and prediction (using a graph in most cases, representing the relations between users and their selected items) is complicated and cumbersome in recommender systems. Graph embedding was also proposed for resolution of this issue, where all or part of user behavior can be simulated through the generation of several vectors, resolving the problem of user behavior analysis to a large extent while maintaining high efficiency. In this research, individuals most similar to the target user were classified using ensemble learning, fuzzy rules, and the decision tree, and relevant recommendations were then made to each user with a heterogeneous knowledge graph and embedding vectors. This study was performed on the MovieLens datasets, and the obtained results indicated the high efficiency of the presented method.

While neural networks excel at predicting a single-point estimate of a given target for complex machine learning problems, an open research question is the design of neural generative models able to output a predictive distribution, that can capture the inherent variability of the observations or the model's level of confidence in its predictions. The main motivation behind uncertainty quantification is the design of AI systems for critical applications that are safe, and are mitigating risks while automatizing decision-making. On one hand, Bayesian statistics offer a mathematically grounded framework to reason about uncertainty; however, such models generally require prohibitive computational costs, which make them not widely used in practice. On the other hand, frequentist methods and metrics have been developed for confidence estimation in neural networks, in particular in the classification setting [Brosse et al., 2020], [Corbière et al., 2019].

As large scale cloud computing centers become more popular than individual servers, predicting future resource demand need has become an important problem. Forecasting resource need allows public cloud providers to proactively allocate or deallocate resources for cloud services. This work seeks to predict one resource, CPU usage, over both a short term and long term time scale. To gain insight into the model characteristics that best support specific tasks, we consider two vastly different architectures: the historically relevant SARIMA model and the more modern neural network, LSTM model. We apply these models to Azure data resampled to 20 minutes per data point with the goal of predicting usage over the next hour for the short-term task and for the next three days for the long-term task. The SARIMA model outperformed the LSTM for the long term prediction task, but performed poorer on the short term task. Furthermore, the LSTM model was more robust, whereas the SARIMA model relied on the data meeting certain assumptions about seasonality.