Continual learning experiments used in current deep learning papers do not faithfully assess fundamental challenges of learning continually, masking weak-points of the suggested approaches instead. We study gaps in such existing evaluations, proposing essential experimental evaluations that are more representative of continual learning's challenges, and suggest a re-prioritization of research efforts in the field. We show that current approaches fail with our new evaluations and, to analyse these failures, we propose a variational loss which unifies many existing solutions to continual learning under a Bayesian framing, as either 'prior-focused' or 'likelihood-focused'. We show that while prior-focused approaches such as EWC and VCL perform well on existing evaluations, they perform dramatically worse when compared to likelihood-focused approaches on other simple tasks.

In our previous study, we introduced stable specification search for cross-sectional data (S3C). It is an exploratory causal method that combines stability selection concept and multi-objective optimization to search for stable and parsimonious causal structures across the entire range of model complexities. In this study, we extended S3C to S3C-Latent, to model causal relations between latent variables. We evaluated S3C-Latent on simulated data and compared the results to those of PC-MIMBuild, an extension of the PC algorithm, the state-of-the-art causal discovery method. The comparison showed that S3C-Latent achieved better performance. We also applied S3C-Latent to real-world data of children with attention deficit/hyperactivity disorder and data about measuring mental abilities among pupils. The results are consistent with those of previous studies.

This paper considers the learning of Boolean rules in either disjunctive normal form (DNF, OR-of-ANDs, equivalent to decision rule sets) or conjunctive normal form (CNF, AND-of-ORs) as an interpretable model for classification. An integer program is formulated to optimally trade classification accuracy for rule simplicity. Column generation (CG) is used to efficiently search over an exponential number of candidate clauses (conjunctions or disjunctions) without the need for heuristic rule mining. This approach also bounds the gap between the selected rule set and the best possible rule set on the training data. To handle large datasets, we propose an approximate CG algorithm using randomization. Compared to three recently proposed alternatives, the CG algorithm dominates the accuracy-simplicity trade-off in 7 out of 15 datasets. When maximized for accuracy, CG is competitive with rule learners designed for this purpose, sometimes finding significantly simpler solutions that are no less accurate.

The paper introduces concentric Echo State Network, an approach to design reservoir topologies that tries to bridge the gap between deterministically constructed simple cycle models and deep reservoir computing approaches. We show how to modularize the reservoir into simple unidirectional and concentric cycles with pairwise bidirectional jump connections between adjacent loops. We provide a preliminary experimental assessment showing how concentric reservoirs yield to superior predictive accuracy and memory capacity with respect to single cycle reservoirs and deep reservoir models.

Agents trained in simulation may make errors in the real world due to mismatches between training and execution environments. These mistakes can be dangerous and difficult to discover because the agent cannot predict them a priori. We propose using oracle feedback to learn a predictive model of these blind spots to reduce costly errors in real-world applications. We focus on blind spots in reinforcement learning (RL) that occur due to incomplete state representation: The agent does not have the appropriate features to represent the true state of the world and thus cannot distinguish among numerous states. We formalize the problem of discovering blind spots in RL as a noisy supervised learning problem with class imbalance. We learn models to predict blind spots in unseen regions of the state space by combining techniques for label aggregation, calibration, and supervised learning. The models take into consideration noise emerging from different forms of oracle feedback, including demonstrations and corrections. We evaluate our approach on two domains and show that it achieves higher predictive performance than baseline methods, and that the learned model can be used to selectively query an oracle at execution time to prevent errors. We also empirically analyze the biases of various feedback types and how they influence the discovery of blind spots.

Transfer learning has become a cornerstone of computer vision with the advent of ImageNet features, yet little work has been done to evaluate the performance of ImageNet architectures across different datasets. An implicit hypothesis in modern computer vision research is that models that perform better on ImageNet necessarily perform better on other vision tasks. However, this hypothesis has never been systematically tested. Here, we compare the performance of 13 classification models on 12 image classification tasks in three settings: as fixed feature extractors, fine-tuned, and trained from random initialization. We find that, when networks are used as fixed feature extractors, ImageNet accuracy is only weakly predictive of accuracy on other tasks ($r^2=0.24$). In this setting, ResNets consistently outperform networks that achieve higher accuracy on ImageNet. When networks are fine-tuned, we observe a substantially stronger correlation ($r^2 = 0.86$). We achieve state-of-the-art performance on eight image classification tasks simply by fine-tuning state-of-the-art ImageNet architectures, outperforming previous results based on specialized methods for transfer learning. Finally, we observe that, on three small fine-grained image classification datasets, networks trained from random initialization perform similarly to ImageNet-pretrained networks. Together, our results show that ImageNet architectures generalize well across datasets, with small improvements in ImageNet accuracy producing improvements across other tasks, but ImageNet features are less general than previously suggested.

Background and objective: Stacking is an ensemble machine learning method that averages predictions from multiple other algorithms, such as generalized linear models and regression trees. A recent iteration of stacking, called super learning, has been developed as a general approach to black box supervised learning and has seen frequent usage, in part due to the availability of an R package. I develop super learning in the SAS software system using a new macro, and demonstrate its performance relative to the R package. Methods: I follow closely previous work using the R SuperLearner package and assess the performance of super learning in a number of domains. I compare the R package with the new SAS macro in a small set of simulations assessing curve fitting in a prediction model, a set of 14 publicly available datasets to assess cross-validated, expected loss, and data from a randomized trial of job seekers' training to assess the utility of super learning in causal inference using inverse probability weighting. Results: Across the simulated data and the publicly available data, the macro performed similarly to the R package, even with a different set of potential algorithms available natively in R and SAS. The example with inverse probability weighting demonstrated the ability of the SAS macro to include algorithms developed in R. Conclusions: The super learner macro performs as well as the R package at a number of tasks. Further, by extending the macro to include the use of R packages, the macro can leverage both the robust, enterprise oriented procedures in SAS and the nimble, cutting edge packages in R. In the spirit of ensemble learning, this macro extends the potential library of algorithms beyond a single software system and provides a simple avenue into machine learning in SAS.

We study the quantification of uncertainty of Convolutional Neural Networks (CNNs) based on gradient metrics. Unlike the classical softmax entropy, such metrics gather information from all layers of the CNN. We show for the (E)MNIST data set that for several such metrics we achieve the same meta classification accuracy -- i.e. the task of classifying correctly predicted labels as correct and incorrectly predicted ones as incorrect without knowing the actual label -- as for entropy thresholding. Meta classification rates for out of sample images can be increased when using entropy together with several gradient based metrics as input quantities for a meta-classifier. This proves that our gradient based metrics do not contain the same information as the entropy. We also apply meta classification to concepts not used during training: EMNIST/Omniglot letters, CIFAR10 and noise. Meta classifiers only trained on the uncertainty metrics of classes available during training usually do not perform equally well for all the unknown concepts letters, CIFAR10 and uniform noise. If we however allow the meta classifier to be trained on uncertainty metrics including some samples of some or all of the categories, meta classification for concepts remote from MNIST digits can be improved considerably.

The primary goal of supervised machine learning is accurate prediction. We want our ML model to be as accurate as possible when predicting on new data (for which the target variable is unknown). Said in a different way, we want our models, which have been built from some training data, to generalize well to new data. That way, when we deploy the model in production, we can be assured that the predictions generated are of high quality. Therefore, when we evaluate the performance of a model, we want to determine how well that model will perform on new data.

We begin this journey through our favorite articles of the month by discovering and reading one of the most influential paper of natural language processing (NLP). Papers can be intimidating to read. So I liked the author's idea of combining screenshots from the original paper, explanations in plain English and code snippets with an actual implementation of the paper in Python. The article can be read at different levels: from the presentation of encoders, decoders and the attention function to real-world examples including the use of regularization and GPU training. One of the world-famous universities of California -- Berkleley -- launches its new data science program.