Interpretable machine learning is an emerging field providing solutions on acquiring insights into machine learning models' rationale. It has been put in the map of machine learning by suggesting ways to tackle key ethical and societal issues. However, existing techniques of interpretable machine learning are far from being comprehensible and explainable to the end user. Another key issue in this field is the lack of evaluation and selection criteria, making it difficult for the end user to choose the most appropriate interpretation technique for its use. In this study, we introduce a meta-explanation methodology that will provide truthful interpretations, in terms of feature importance, to the end user through argumentation. At the same time, this methodology can be used as an evaluation or selection tool for multiple interpretation techniques based on feature importance.

We propose a new framework for reasoning about generalization in deep learning. The core idea is to couple the Real World, where optimizers take stochastic gradient steps on the empirical loss, to an Ideal World, where optimizers take steps on the population loss. This leads to an alternate decomposition of test error into: (1) the Ideal World test error plus (2) the gap between the two worlds. If the gap (2) is universally small, this reduces the problem of generalization in offline learning to the problem of optimization in online learning. We then give empirical evidence that this gap between worlds can be small in realistic deep learning settings, in particular supervised image classification. For example, CNNs generalize better than MLPs on image distributions in the Real World, but this is "because" they optimize faster on the population loss in the Ideal World. This suggests our framework is a useful tool for understanding generalization in deep learning, and lays a foundation for future research in the area.

Supervised classification techniques use training samples to find classification rules with small expected 0-1 loss. Conventional methods achieve efficient learning and out-of-sample generalization by minimizing surrogate losses over specific families of rules. This paper presents minimax risk classifiers (MRCs) that do not rely on a choice of surrogate loss and family of rules. MRCs achieve efficient learning and out-of-sample generalization by minimizing worst-case expected 0-1 loss w.r.t. uncertainty sets that are defined by linear constraints and include the true underlying distribution. In addition, MRCs' learning stage provides performance guarantees as lower and upper tight bounds for expected 0-1 loss. We also present MRCs' finite-sample generalization bounds in terms of training size and smallest minimax risk, and show their competitive classification performance w.r.t. state-of-the-art techniques using benchmark datasets.

Bilevel optimization has arisen as a powerful tool for many machine learning problems such as meta-learning, hyper-parameter optimization, reinforcement learning, etc. In this paper, we investigate the nonconvex-strongly-convex bilevel optimization problem, and propose two novel algorithms named deterBiO and stocBiO respectively for the deterministic and stochastic settings. At the core design of deterBiO is the construction of a low-cost and easy-to-implement hyper-gradient estimator via a simple back-propagation. In addition, stocBiO updates with the mini-batch data sampling rather than the existing single-sample schemes, where a sample-efficient Hessian inverse estimator is proposed. We provide the finite-time convergence guarantee for both algorithms, and show that they outperform the best known computational complexities orderwisely with respect to the condition number $\kappa$ and/or the target accuracy $\epsilon$. We further demonstrate the superior efficiency of the proposed algorithms by the experiments on meta-learning and hyper-parameter optimization.

Following these steps, we aim at showing how the latter formulation can be useful to understand and analyse Many optimisation and inference problems have been the outcome of GMMs. In particular, we exploit the shown to have an equivalent formulation in statistical cascade of phase transitions occurring during annealing physics [1, 2] that allowed a brand-new look at some longstanding procedures of the EM algorithm to build a hierarchical problems and improved the understanding of multi-scale description of a dataset. By defining an overlap complex systems [3, 4]. In particular, the identification of between the ground truth and the inferred partitions, the phase diagram of a model can bring interesting new we show on artificial datasets how it can be interpreted insights such as knowing if a given information can be as an order parameter whose value follows the sequence retrieved depending on the model's parameters and the of phase transitions.

We study contextual stochastic optimization problems, where we leverage rich auxiliary observations (e.g., product characteristics) to improve decision making with uncertain variables (e.g., demand). We show how to train forest decision policies for this problem by growing trees that choose splits to directly optimize the downstream decision quality, rather than splitting to improve prediction accuracy as in the standard random forest algorithm. We realize this seemingly computationally intractable problem by developing approximate splitting criteria that utilize optimization perturbation analysis to eschew burdensome re-optimization for every candidate split, so that our method scales to large-scale problems. We prove that our splitting criteria consistently approximate the true risk and that our method achieves asymptotic optimality. We extensively validate our method empirically, demonstrating the value of optimization-aware construction of forests and the success of our efficient approximations. We show that our approximate splitting criteria can reduce running time hundredfold, while achieving performance close to forest algorithms that exactly re-optimize for every candidate split.

Interaction between pharmacological agents can trigger unexpected adverse events. Capturing richer and more comprehensive information about drug-drug interactions (DDI) is one of the key tasks in public health and drug development. Recently, several knowledge graph embedding approaches have received increasing attention in the DDI domain due to their capability of projecting drugs and interactions into a low-dimensional feature space for predicting links and classifying triplets. However, existing methods only apply a uniformly random mode to construct negative samples. As a consequence, these samples are often too simplistic to train an effective model. In this paper, we propose a new knowledge graph embedding framework by introducing adversarial autoencoders (AAE) based on Wasserstein distances and Gumbel-Softmax relaxation for drug-drug interactions tasks. In our framework, the autoencoder is employed to generate high-quality negative samples and the hidden vector of the autoencoder is regarded as a plausible drug candidate. Afterwards, the discriminator learns the embeddings of drugs and interactions based on both positive and negative triplets. Meanwhile, in order to solve vanishing gradient problems on the discrete representation--an inherent flaw in traditional generative models--we utilize the Gumbel-Softmax relaxation and the Wasserstein distance to train the embedding model steadily. We empirically evaluate our method on two tasks, link prediction and DDI classification. The experimental results show that our framework can attain significant improvements and noticeably outperform competitive baselines.

Self-supervised learning has emerged as a strategy to reduce the reliance on costly supervised signal by pretraining representations only using unlabeled data. These methods combine heuristic proxy classification tasks with data augmentations and have achieved significant success, but our theoretical understanding of this success remains limited. In this paper we analyze self-supervised representation learning using a causal framework. We show how data augmentations can be more effectively utilized through explicit invariance constraints on the proxy classifiers employed during pretraining. Based on this, we propose a novel self-supervised objective, Representation Learning via Invariant Causal Mechanisms (ReLIC), that enforces invariant prediction of proxy targets across augmentations through an invariance regularizer which yields improved generalization guarantees. Further, using causality we generalize contrastive learning, a particular kind of self-supervised method, and provide an alternative theoretical explanation for the success of these methods. Empirically, ReLIC significantly outperforms competing methods in terms of robustness and out-of-distribution generalization on ImageNet, while also significantly outperforming these methods on Atari achieving above human-level performance on $51$ out of $57$ games.

Learning the cumulative distribution function (CDF) of an outcome variable conditional on a set of features remains challenging, especially in high-dimensional settings. Conditional transformation models provide a semi-parametric approach that allows to model a large class of conditional CDFs without an explicit parametric distribution assumption and with only a few parameters. Existing estimation approaches within the class of transformation models are, however, either limited in their complexity and applicability to unstructured data sources such as images or text, or can incorporate complex effects of different features but lack interpretability. We close this gap by introducing the class of deep conditional transformation models which unify existing approaches and allow to learn both interpretable (non-)linear model terms and more complex predictors in one holistic neural network. To this end we propose a novel network architecture, provide details on different model definitions and derive suitable constraints and derive suitable network regularization terms. We demonstrate the efficacy of our approach through numerical experiments and applications.

The k-Nearest Neighbors (kNN) classifier is a fundamental non-parametric machine learning algorithm. However, it is well known that it suffers from the curse of dimensionality, which is why in practice one often applies a kNN classifier on top of a (pre-trained) feature transformation. From a theoretical perspective, most, if not all theoretical results aimed at understanding the kNN classifier are derived for the raw feature space. This leads to an emerging gap between our theoretical understanding of kNN and its practical applications. In this paper, we take a first step towards bridging this gap. We provide a novel analysis on the convergence rates of a kNN classifier over transformed features. This analysis requires in-depth understanding of the properties that connect both the transformed space and the raw feature space. More precisely, we build our convergence bound upon two key properties of the transformed space: (1) safety -- how well can one recover the raw posterior from the transformed space, and (2) smoothness -- how complex this recovery function is. Based on our result, we are able to explain why some (pre-trained) feature transformations are better suited for a kNN classifier than other. We empirically validate that both properties have an impact on the kNN convergence on 30 feature transformations with 6 benchmark datasets spanning from the vision to the text domain.