There are still open questions about how the learned representations of deep models change during the training process. Understanding this process could aid in validating the training. Towards this goal, previous works analyze the training in the mutual information plane. We use a different approach and base our analysis on a method built on Reichenbach’s common cause principle. Using this method, we test whether the model utilizes information contained in human-defined features. Given such a set of features, we investigate how the relative feature usage changes throughout the training process. We analyze mul- tiple networks training on different tasks, including melanoma classifica- tion as a real-world application. We find that over the training, models concentrate on features containing information relevant to the task. This concentration is a form of representation compression. Crucially, we also find that the selected features can differ between training from-scratch and finetuning a pre-trained network.

Teleconnections that link climate processes at widely separated spatial locations form a key component of the climate system. Their analysis has traditionally been based on means, climatologies, correlations, or spectral properties, which cannot always reveal the dynamical mechanisms between different climatological processes. More recently, causal discovery methods based either on time series at grid locations or on modes of variability, estimated through dimension-reduction methods, have been introduced. A major challenge in the development of such analysis methods is a lack of ground truth benchmark datasets that have facilitated improvements in many parts of machine learning. Here, we present a simplified stochastic climate model that outputs gridded data and represents climate modes and their teleconnections through a spatially aggregated vector-autoregressive model. The model is used to construct benchmarks and evaluate a range of analysis methods. The results highlight that the model can be successfully used to benchmark different causal discovery methods for spatiotemporal data and show their strengths and weaknesses. Furthermore, we introduce a novel causal discovery method at the grid level and demonstrate that it has orders of magnitude better performance than the current approaches. Improved causal analysis tools for spatiotemporal climate data are pivotal to advance process-based understanding and climate model evaluation.

Skin cancer is the most common form of cancer, and melanoma is the leading cause of cancer related deaths. To improve the chances of survival, early detection of melanoma is crucial. Automated systems for classifying skin lesions can assist with initial analysis. However, if we expect people to entrust their well-being to an automatic classification algorithm, it is important to ensure that the algorithm makes medically sound decisions. We investigate this question by testing whether two state-of-the-art models use the features defined in the dermoscopic ABCD rule or whether they rely on biases. We use a method that frames supervised learning as a structural causal model, thus reducing the question whether a feature is used to a conditional dependence test. We show that this conditional dependence method yields meaningful results on data from the ISIC archive. Furthermore, we find that the selected models incorporate asymmetry, border and dermoscopic structures in their decisions but not color. Finally, we show that the same classifiers also use bias features such as the patient's age, skin color or the existence of colorful patches.

Bias in classifiers is a severe issue of modern deep learning methods, especially for their application in safety- and security-critical areas. Often, the bias of a classifier is a direct consequence of a bias in the training set, frequently caused by the co-occurrence of relevant features and irrelevant ones. To mitigate this issue, we require learning algorithms that prevent the propagation of known bias from the dataset into the classifier. We present a novel adversarial debiasing method, which addresses a feature of which we know that it is spuriously connected to the labels of training images but statistically independent of the labels for test images. The debiasing stops the classifier from falsly identifying this irrelevant feature as important. Irrelevant features co-occur with important features in a wide range of bias-related problems for many computer vision tasks, such as automatic skin cancer detection or driver assistance. We argue by a mathematical proof that our approach is superior to existing techniques for the abovementioned bias. Our experiments show that our approach performs better than the state-of-the-art on a well-known benchmark dataset with real-world images of cats and dogs.

Uncertainty sampling is a widely used active learning strategy to select unlabeled examples for annotation. However, previous work hints at weaknesses of uncertainty sampling when combined with deep learning, where the amount of data is even more significant. To investigate these problems, we analyze the properties of the latent statistical estimators of uncertainty sampling in simple scenarios. We prove that uncertainty sampling converges towards some decision boundary. Additionally, we show that it can be inconsistent, leading to incorrect estimates of the optimal latent boundary. The inconsistency depends on the latent class distribution, more specifically on the class overlap. Further, we empirically analyze the variance of the decision boundary and find that the performance of uncertainty sampling is also connected to the class regions overlap. We argue that our findings could be the first step towards explaining the poor performance of uncertainty sampling combined with deep models.

Deep neural networks are tremendously successful in many applications, but end-to-end trained networks often result in hard to un- derstand black-box classifiers or predictors. In this work, we present a novel method to identify whether a specific feature is relevant to a clas- sifier’s decision or not. This relevance is determined at the level of the learned mapping, instead of for a single example. The approach does neither need retraining of the network nor information on intermedi- ate results or gradients. The key idea of our approach builds upon con- cepts from causal inference. We interpret machine learning in a struc- tural causal model and use Reichenbach’s common cause principle to infer whether a feature is relevant. We demonstrate empirically that the method is able to successfully evaluate the relevance of given features on three real-life data sets, namely MS COCO, CUB200 and HAM10000.

State-of-the-art machine learning methods, especially deep neural networks, have reached impressive results in many prediction and classification tasks. Rising complexity and automatic feature selection make the resulting learned models hard to interpret and turns them into black boxes. Advances into feature visualization have mitigated this problem but some shortcomings still exist. For example, methods only work locally, meaning they only explain the behavior for single inputs, and they only identify important parts of the input. In this work, we propose a method that is also able to decide whether a feature calculated from the input to an estimator is globally useful. Since the question about explanatory power is a causal one, we frame this approach with causal inference methods.

Climate projections are limited by the arising uncertainties associated with not well-known physical processes in climate change. In every new generation, climate models improve several aspects of projections, while others remain in the same uncertainty range, especially those regarding equilibrium climate sensitivity (ECS) and climate feedbacks. Emergent constraints, defined as a 'physically explainable empirical relationships between characteristics of the current climate and long-term climate prediction that emerge in collections of climate model simulations' [1], is a promising novel approach that can shed light on climate change uncertainties and improve climate models. Since the first emergent constraint was proposed to constrain the Surface Albedo Feedback in 2006 by Hall & Qu [2], several of them have been proposed for constraining feedbacks and uncertainties of climate models, e.g., ECS, low-level cloud optical depth or tropical primary production. Emergent constraints have already been a prolific approach to improve our climate models. The typical approach to identify emergent constraints comes from expert knowledge when this is used to explore climate data and select those emergent constraints that are physically explainable. Caldwell et al. [3] presented a work where a new approach was suggested. In that paper, they attempt to identify quantities in the current climate which are skillful predictors of ECS yet can be constrained by observations. One of the main conclusions of this work was that the development of data mining methods for identifying emergent constraints should be aware of spurious emergent relations that could arise by chance. This becomes especially relevant in the next phase of the CMIP Project (6th). In the present work, we discuss simple spatiotemporal climate ("toy") models to analyze and evaluate methodologies to identify predictors for emergent constraints. Such models are simple enough to be analyzed not only empirically but also analytically, and at the same time incorporate relevant aspects of the complexity of a nonlinear dynamical spatiotemporal system. Consequently, they can be used to study assumptions and pitfalls of data mining methods for emergent constraints and guide the development of future approaches. [1]: Klein, S. A., & Hall, A. (2015). Emergent constraints for cloud feedbacks. Current Climate Change Reports, 1(4), 276-287. [2]: Hall, A., & Qu, X. (2006). Using the current seasonal cycle to constrain snow albedo feedback in future climate change. Geophysical Research Letters, 33(3). [3]: Caldwell, P. M., Bretherton, C. S., Zelinka, M. D., Klein, S. A., Santer, B. D., & Sanderson, B. M. (2014). Statistical significance of climate sensitivity predictors obtained by data mining. Geophysical Research Letters, 41(5), 1803-1808.

Exploring and understanding spatio-temporal patterns on Earth system datasets is one of the principal goals of the climate and geo-science communities. In this direction, Empirical Orthogonal Functions (EOFs) have been used to characterize phenomena such as the El Nino Southern Oscillation, the Arctic jet stream or the Indian Monsoon. However, EOF analysis has several limitations, for example, it can only identify linear and orthogonal patterns. We present a framework that makes use of a convolutional variational autoencoder (VAE) as a learnable feature function to extract spatio-temporal dynamics via PCA. The VAE encodes the information in an abstract space of higher order features representing different patterns. Over this space, PCA is performed to obtain a spatial representation of related temporal dynamics. We have used three datasets, two artificial datasets where the dynamics are ruled by a hidden spatially varying parameter and an observational reanalysis dataset of monthly sea surface temperature from 1898 to 2014. The artificial datasets have chaotic and, chaotic and stochastic dynamics depending on the spatial hidden parameter. As baseline methods, EOF analysis and Kernel PCA were performed over the original spaces. For the two artificial datasets, we found a high correlation between some of the first Principal Components on the feature space and the spatial hidden parameter. This correlation was not found using baseline methods in the original space. In the reanalysis dataset, the method was able to find known modes, such as ENSO, as well as other patterns that baseline methods did not reveal that might have inmmediate effect on how we understand the earth system after expert interpretation. These results provide a proof of concept: SupernoVAE is not only able to extract well-known climate patterns previously characterized with linear EOF analysis, but also allows to extract non-linear and non-orthogonal patterns that can help in analyzing Earth system dynamics that could not be characterized before.

It is a constant challenge to better understand the underlying dynamics and forces driving the Earth system. Advances in the field of deep learning allow for unprecedented results, but use of these methods in Earth system science is still very limited. We present a framework that makes use of a convolutional variational autoencoder as a learnable kernel from which to extract spatio-temporal dynamics via PCA. The method promises the ability of deep learning to digest highly complex spatio-temporal datasets while allowing expert interpretability. Preliminary results over two artificial datasets, with chaotic and stochastic temporal dynamics, show that the method can recover a latent driver parameter while baseline approaches cannot. While further testing on the limitations of the method is needed and experiments on real Earth datasets are in order, the present approach may contribute to further the understanding of Earth datasets that are highly non-linear.