Probabilistic forecasting models for joint distributions of targets in irregular time series are a heavily under-researched area in machine learning with, to the best of our knowledge, only three models researched so far: GPR, the Gaussian Process Regression model [16], TACTiS, the Transformer-Attentional Copulas for Time Series [14, 2] and ProFITi [43], a multivariate normalizing flow model based on invertible attention layers. While ProFITi, thanks to using multivariate normalizing flows, is the more expressive model with a better predictive performance, we will show that it suffers from marginalization inconsistency: it does not guarantee that the marginal distributions of a subset of variables in its predictive distributions coincide with the directly predicted distributions of these variables. Also, TACTiS does not provide any guarantees for marginalization consistency. We develop a novel probabilistic irregular time series forecasting model, Marginalization Consistent Mixtures of Separable Flows (moses), that mixes several normalizing flows with (i) Gaussian Processes with full covariance matrix as source distributions and (ii) a separable invertible transformation, aiming to combine the expressivity of normalizing flows with the marginalization consistency of Gaussians. In experiments on four different datasets we show that moses outperform other state-of-the-art marginalization consistent models, perform on par with ProFITi, but different from ProFITi, guarantees marginalization consistency.

Probabilistic forecasting of irregularly sampled multivariate time series with missing values is an important problem in many fields, including health care, astronomy, and climate. State-of-the-art methods for the task estimate only marginal distributions of observations in single channels and at single timepoints, assuming a fixed-shape parametric distribution. In this work, we propose a novel model, ProFITi, for probabilistic forecasting of irregularly sampled time series with missing values using conditional normalizing flows. The model learns joint distributions over the future values of the time series conditioned on past observations and queried channels and times, without assuming any fixed shape of the underlying distribution. As model components, we introduce a novel invertible triangular attention layer and an invertible non-linear activation function on and onto the whole real line. We conduct extensive experiments on four datasets and demonstrate that the proposed model provides $4$ times higher likelihood over the previously best model.

In deep learning models, learning more with less data is becoming more important. This paper explores how neural networks with normalized Radial Basis Function (RBF) kernels can be trained to achieve better sample efficiency. Moreover, we show how this kind of output layer can find embedding spaces where the classes are compact and well-separated. In order to achieve this, we propose a two-phase method to train those type of neural networks on classification tasks. Experiments on CIFAR-10 and CIFAR-100 show that networks with normalized kernels as output layer can achieve higher sample efficiency, high compactness and well-separability through the presented method in comparison to networks with SoftMax output layer.

Parametric models, and particularly neural networks, require weight initialization as a starting point for gradient-based optimization. In most current practices, this is accomplished by using some form of random initialization. Instead, recent work shows that a specific initial parameter set can be learned from a population of tasks, i.e., dataset and target variable for supervised learning tasks. Using this initial parameter set leads to faster convergence for new tasks (model-agnostic meta-learning). Currently, methods for learning model initializations are limited to a population of tasks sharing the same schema, i.e., the same number, order, type and semantics of predictor and target variables. In this paper, we address the problem of meta-learning parameter initialization across tasks with different schemas, i.e., if the number of predictors varies across tasks, while they still share some variables. We propose Chameleon, a model that learns to align different predictor schemas to a common representation. We use permutations and masks of the predictors of the training tasks at hand. In experiments on real-life data sets, we show that Chameleon successfully can learn parameter initializations across tasks with different schemas providing a 26% lift on accuracy on average over random initialization and of 5% over a state-of-the-art method for fixed-schema learning model initializations. To the best of our knowledge, our paper is the first work on the problem of learning model initialization across tasks with different schemas.

The minimization of loss functions is the heart and soul of Machine Learning. In this paper, we propose an off-the-shelf optimization approach that can minimize virtually any non-differentiable and non-decomposable loss function (e.g. Miss-classification Rate, AUC, F1, Jaccard Index, Mathew Correlation Coefficient, etc.) seamlessly. Our strategy learns smooth relaxation versions of the true losses by approximating them through a surrogate neural network. The proposed loss networks are set-wise models which are invariant to the order of mini-batch instances. Ultimately, the surrogate losses are learned jointly with the prediction model via bilevel optimization. Empirical results on multiple datasets with diverse real-life loss functions compared with state-of-the-art baselines demonstrate the efficiency of learning surrogate losses.