Abstract--In dynamic selection (DS) techniques, only the most competent classifiers, for the classification of a specific test sample are selected to predict the sample's class labels. The more important step in DES techniques is estimating the competence of the base classifiers for the classification of each specific test sample. The classifiers' competence is usually estimated using the neighborhood of the test sample defined on the validation samples, called the region of competence. Thus, the performance of DS techniques is sensitive to the distribution of the validation set. In this paper, we evaluate six prototype selection techniques that work by editing the validation data in order to remove noise and redundant instances. Experiments conducted using several state-of-the-art DS techniques over 30 classification problems demonstrate that by using prototype selection techniques we can improve the classification accuracy of DS techniques and also significantly reduce the computational cost involved. Multiple Classifier Systems (MCS) aim to combine classifiers in order to increase the recognition accuracy in pattern recognition systems [1], [2]. MCS are composed of three phases [3]: (1) Generation, (2) Selection, and (3) Integration.

We present a method that "meta" classifies whether segments (objects) predicted by a semantic segmentation neural network intersect with the ground truth. To this end, we employ measures of dispersion for predicted pixel-wise class probability distributions, like classification entropy, that yield heat maps of the input scene's size. We aggregate these dispersion measures segment-wise and derive metrics that are well-correlated with the segment-wise $\mathit{IoU}$ of prediction and ground truth. In our tests, we use two publicly available DeepLabv3+ networks (pre-trained on the Cityscapes data set) and analyze the predictive power of different metrics and different sets of metrics. To this end, we compute logistic LASSO regression fits for the task of classifying $\mathit{IoU}=0$ vs. $\mathit{IoU} > 0$ per segment and obtain classification rates of up to $81.91\%$ and AUROC values of up to $87.71\%$ without the incorporation of advanced techniques like Monte-Carlo dropout. We complement these tests with linear regression fits to predict the segment-wise $\mathit{IoU}$ and obtain prediction standard deviations of down to $0.130$ as well as $R^2$ values of up to $81.48\%$. We show that these results clearly outperform single-metric baseline approaches.

Deep learning models have become state of the art for natural language processing (NLP) tasks, however deploying these models in production system poses significant memory constraints. Existing compression methods are either lossy or introduce significant latency. We propose a compression method that leverages low rank matrix factorization during training,to compress the word embedding layer which represents the size bottleneck for most NLP models. Our models are trained, compressed and then further re-trained on the downstream task to recover accuracy while maintaining the reduced size. Empirically, we show that the proposed method can achieve 90% compression with minimal impact in accuracy for sentence classification tasks, and outperforms alternative methods like fixed-point quantization or offline word embedding compression. We also analyze the inference time and storage space for our method through FLOP calculations, showing that we can compress DNN models by a configurable ratio and regain accuracy loss without introducing additional latency compared to fixed point quantization. Finally, we introduce a novel learning rate schedule, the Cyclically Annealed Learning Rate (CALR), which we empirically demonstrate to outperform other popular adaptive learning rate algorithms on a sentence classification benchmark.

Due to its high computational speed and accuracy compared to ab-initio quantum chemistry and forcefield modeling, the prediction of molecular properties using machine learning has received great attention in the fields of materials design and drug discovery. A main ingredient required for machine learning is a training dataset consisting of molecular features\textemdash for example fingerprint bits, chemical descriptors, etc. that adequately characterize the corresponding molecules. However, choosing features for any application is highly non-trivial. No "universal" method for feature selection exists. In this work, we propose a data fusion framework that uses Independent Vector Analysis to exploit underlying complementary information contained in different molecular featurization methods, bringing us a step closer to automated feature generation. Our approach takes an arbitrary number of individual feature vectors and automatically generates a single, compact (low dimensional) set of molecular features that can be used to enhance the prediction performance of regression models. At the same time our methodology retains the possibility of interpreting the generated features to discover relationships between molecular structures and properties. We demonstrate this on the QM7b dataset for the prediction of several properties such as atomization energy, polarizability, frontier orbital eigenvalues, ionization potential, electron affinity, and excitation energies. In addition, we show how our method helps improve the prediction of experimental binding affinities for a set of human BACE-1 inhibitors. To encourage more widespread use of IVA we have developed the PyIVA Python package, an open source code which is available for download on Github.

Modern sensors generate large amounts of timestamped measurement data. These data sets are critical in a wide range of applications including traffic flow prediction, transportation management, GPS navigation, and city planning. Machine learning-based prediction algorithms typically adjust their parameters automatically based on the data, but also require users to set additional parameters, known as hyperparameters. For example, in a kernel-based regression model, the (ordinary) parameters are the regression weights, whereas the hyperparameters include the kernel scales and regularization constants. These hyperparameters have a strong influence on the prediction accuracy. Often, their values are set based on past experience or through time-consuming grid searches. In applications where the characteristics of the data change, such as unusual traffic pattern due to upcoming concert events, these hyperparameters have to be adjusted dynamically in order to maintain prediction quality. In this paper, we use the term hyperparameter learning, hyperparameter optimization, and hyperparameter selection/tuning interchangeably, referring to the process of configuring the model specification before model fitting.

Deep structured models are widely used for tasks like semantic segmentation, where explicit correlations between variables provide important prior information which generally helps to reduce the data needs of deep nets. However, current deep structured models are restricted by oftentimes very local neighborhood structure, which cannot be increased for computational complexity reasons, and by the fact that the output configuration, or a representation thereof, cannot be transformed further. Very recent approaches which address those issues include graphical model inference inside deep nets so as to permit subsequent non-linear output space transformations. However, optimization of those formulations is challenging and not well understood. Here, we develop a novel model which generalizes existing approaches, such as structured prediction energy networks, and discuss a formulation which maintains applicability of existing inference techniques.

Adversarial examples are a pervasive phenomenon of machine learning models where seemingly imperceptible perturbations to the input lead to misclassifications for otherwise statistically accurate models. We propose a geometric framework, drawing on tools from the manifold reconstruction literature, to analyze the high-dimensional geometry of adversarial examples. In particular, we highlight the importance of codimension: for low-dimensional data manifolds embedded in high-dimensional space there are many directions off the manifold in which to construct adversarial examples. Adversarial examples are a natural consequence of learning a decision boundary that classifies the low-dimensional data manifold well, but classifies points near the manifold incorrectly. Using our geometric framework we prove (1) a tradeoff between robustness under different norms, (2) that adversarial training in balls around the data is sample inefficient, and (3) sufficient sampling conditions under which nearest neighbor classifiers and ball-based adversarial training are robust.

Bayesian modelling of dynamic systems must achieve a compromise between providing a complete mechanistic specification of the process while retaining the flexibility to handle those situations in which data is sparse relative to model complexity, or a full specification is hard to motivate. Latent force models achieve this dual aim by specifying a parsimonious linear evolution equation which an additive latent Gaussian process (GP) forcing term. In this work we extend the latent force framework to allow for multiplicative interactions between the GP and the latent states leading to more control over the geometry of the trajectories. Unfortunately inference is no longer straightforward and so we introduce an approximation based on the method of successive approximations and examine its performance using a simulation study.

To cluster data that are not linearly separable in the original feature space, $k$-means clustering was extended to the kernel version. However, the performance of kernel $k$-means clustering largely depends on the choice of kernel function. To mitigate this problem, multiple kernel learning has been introduced into the $k$-means clustering to obtain an optimal kernel combination for clustering. Despite the success of multiple kernel $k$-means clustering in various scenarios, few of the existing work update the combination coefficients based on the diversity of kernels, which leads to the result that the selected kernels contain high redundancy and would degrade the clustering performance and efficiency. In this paper, we propose a simple but efficient strategy that selects a diverse subset from the pre-specified kernels as the representative kernels, and then incorporate the subset selection process into the framework of multiple $k$-means clustering. The representative kernels can be indicated as the significant combination weights. Due to the non-convexity of the obtained objective function, we develop an alternating minimization method to optimize the combination coefficients of the selected kernels and the cluster membership alternatively. We evaluate the proposed approach on several benchmark and real-world datasets. The experimental results demonstrate the competitiveness of our approach in comparison with the state-of-the-art methods.

Sparse regression such as Lasso has achieved great success in dealing with high dimensional data for several decades. However, there are few methods applicable to missing data, which often occurs in high dimensional data. Recently, CoCoLasso was proposed to deal with high dimensional missing data, but it still suffers from highly missing data. In this paper, we propose a novel Lasso-type regression technique for Highly Missing data, called `HMLasso'. We use the mean imputed covariance matrix, which is notorious in general due to its estimation bias for missing data. However, we effectively incorporate it into Lasso, by using a useful connection with the pairwise covariance matrix. The resulting optimization problem can be seen as a weighted modification of CoCoLasso with the missing ratios, and is quite effective for highly missing data. To the best of our knowledge, this is the first method that can efficiently deal with both high dimensional and highly missing data. We show that the proposed method is beneficial with regards to non-asymptotic properties of the covariance matrix. Numerical experiments show that the proposed method is highly advantageous in terms of estimation error and generalization error.