There is a set of data augmentation techniques that ablate parts of the input at random. These include input dropout, cutout, and random erasing. We term these techniques ablated data augmentation. Though these techniques seems similar in spirit and have shown success in improving model performance in a variety of domains, we do not yet have a mathematical understanding of the differences between these techniques like we do for other regularization techniques like L1 or L2. First, we study a formal model of mean ablated data augmentation and inverted dropout for linear regression. We prove that ablated data augmentation is equivalent to optimizing the ordinary least squares objective along with a penalty that we call the Contribution Covariance Penalty and inverted dropout, a more common implementation than dropout in popular frameworks, is equivalent to optimizing the ordinary least squares objective along with Modified L2. For deep networks, we demonstrate an empirical version of the result if we replace contributions with attributions and coefficients with average gradients, i.e., the Contribution Covariance Penalty and Modified L2 Penalty drop with the increase of the corresponding ablated data augmentation across a variety of networks.

Gradient descent has been a central training principle for artificial neural networks from the early beginnings to today's deep learning networks. The most common implementation is the backpropagation algorithm for training feed-forward neural networks in a supervised fashion. Backpropagation involves computing the gradient of a loss function, with respect to the weights of the network, to update the weights and thus minimize loss. Although the mean square error is often used as a loss function, the general stochastic gradient descent principle does not immediately connect with a specific loss function. Another drawback of backpropagation has been the search for optimal values of two important training parameters, learning rate and momentum weight, which are determined empirically in most systems. The learning rate specifies the step size towards a minimum of the loss function when following the gradient, while the momentum weight considers previous weight changes when updating current weights. Using both parameters in conjunction with each other is generally accepted as a means to improving training, although their specific values do not follow immediately from standard backpropagation theory. This paper proposes a new information-theoretical loss function motivated by neural signal processing in a synapse. The new loss function implies a specific learning rate and momentum weight, leading to empirical parameters often used in practice. The proposed framework also provides a more formal explanation of the momentum term and its smoothing effect on the training process. All results taken together show that loss, learning rate, and momentum are closely connected. To support these theoretical findings, experiments for handwritten digit recognition show the practical usefulness of the proposed loss function and training parameters.

Practitioners use feature importance to rank and eliminate weak predictors during model development in an effort to simplify models and improve generality. Unfortunately, they also routinely conflate such feature importance measures with feature impact, the isolated effect of an explanatory variable on the response variable. This can lead to real-world consequences when importance is inappropriately interpreted as impact for business or medical insight purposes. The dominant approach for computing importances is through interrogation of a fitted model, which works well for feature selection, but gives distorted measures of feature impact. The same method applied to the same data set can yield different feature importances, depending on the model, leading us to conclude that impact should be computed directly from the data. While there are nonparametric feature selection algorithms, they typically provide feature rankings, rather than measures of impact or importance. They also typically focus on single-variable associations with the response. In this paper, we give mathematical definitions of feature impact and importance, derived from partial dependence curves, that operate directly on the data. To assess quality, we show that features ranked by these definitions are competitive with existing feature selection techniques using three real data sets for predictive tasks.

Low-dimensional representations, or embeddings, of a graph's nodes facilitate data mining tasks. Known embedding methods explicitly or implicitly rely on a similarity measure among nodes. As the similarity matrix is quadratic, a tradeoff between space complexity and embedding quality arises; past research initially opted for heuristics and linear-transform factorizations, which allow for linear space but compromise on quality; recent research has proposed a quadratic-space solution as a viable option too. In this paper we observe that embedding methods effectively aim to preserve the covariance among the rows of a similarity matrix, and raise the question: is there a method that combines (i) linear space complexity, (ii) a nonlinear transform as its basis, and (iii) nontrivial quality guarantees? We answer this question in the affirmative, with FREDE(FREquent Directions Embedding), a sketching-based method that iteratively improves on quality while processing rows of the similarity matrix individually; thereby, it provides, at any iteration, column-covariance approximation guarantees that are, in due course, almost indistinguishable from those of the optimal row-covariance approximation by SVD. Our experimental evaluation on variably sized networks shows that FREDE performs as well as SVD and competitively against current state-of-the-art methods in diverse data mining tasks, even when it derives an embedding based on only 10% of node similarities.

One of the fundamental challenges for running machine learning algorithms on battery-powered devices is the time and energy needed for computation, as these devices have constraints on resources. There are energy-efficient classifier algorithms, but their accuracy is often sacrificed for resource efficiency. Here, we propose an ultra-low power binary classifier, SEFR, with linear time complexity, both in the training and the testing phases. The SEFR method runs by creating a hyperplane to separate two classes. The weights of this hyperplane are calculated using normalization, and then the bias is computed based on the weights. SEFR is comparable to state-of-the-art classifiers in terms of classification accuracy, but its execution time and energy consumption are 11.02% and 8.67% of the average of state-of-the-art and baseline classifiers. The energy and memory consumption of SEFR is very insignificant, and it even can perform both train and test phases on microcontrollers. We have implemented SEFR on Arduino Uno, and on a dataset with 100 records and 100 features, the training time is 195 milliseconds, and testing for 100 records with 100 features takes 0.73 milliseconds. To the best of our knowledge, this is the first multipurpose algorithm specifically devised for learning on ultra-low power devices.

Reliable yet efficient evaluation of generalisation performance of a proposed architecture is crucial to the success of neural architecture search (NAS). Traditional approaches face a variety of limitations: training each architecture to completion is prohibitively expensive, early stopping estimates may correlate poorly with fully trained performance, and model-based estimators require large training sets. Instead, motivated by recent results linking training speed and generalisation with stochastic gradient descent, we propose to estimate the final test performance based on the sum of training losses. Our estimator is inspired by the marginal likelihood, which is used for Bayesian model selection. Our model-free estimator is simple, efficient, and cheap to implement, and does not require hyperparameter-tuning or surrogate training before deployment. We demonstrate empirically that our estimator consistently outperforms other baselines and can achieve a rank correlation of 0.95 with final test accuracy on the NAS-Bench201 dataset within 50 epochs.

Machine learning algorithms with empirical risk minimization are vulnerable to distributional shifts due to the greedy adoption of all the correlations found in training data. Recently, there are robust learning methods aiming at this problem by minimizing the worst-case risk over an uncertainty set. However, they equally treat all covariates to form the uncertainty sets regardless of the stability of their correlations with the target, resulting in the overwhelmingly large set and low confidence of the learner. In this paper, we propose the Invariant Adversarial Learning (IAL) algorithm that leverages heterogeneous data sources to construct a more practical uncertainty set and conduct robustness optimization, where covariates are differentiated according to the stability of their correlations with the target. We theoretically show that our method is tractable for stochastic gradient-based optimization and provide the performance guarantees for our method.

In machine learning, it is commonly assumed that training and test data share the same population distribution. However, this assumption is often violated in practice because the sample selection bias may induce the distribution shift from training data to test data. Such a model-agnostic distribution shift usually leads to prediction instability across unknown test data. In this paper, we propose a novel balance-subsampled stable prediction (BSSP) algorithm based on the theory of fractional factorial design. It isolates the clear effect of each predictor from the confounding variables. A design-theoretic analysis shows that the proposed method can reduce the confounding effects among predictors induced by the distribution shift, hence improve both the accuracy of parameter estimation and prediction stability. Numerical experiments on both synthetic and real-world data sets demonstrate that our BSSP algorithm significantly outperforms the baseline methods for stable prediction across unknown test data.

Time series forecasting is ubiquitous in real-world applications. Examples include forecasting of demand to fill up inventories, economic growth forecasts to inform policies, and predicting stock prices to guide financial decisions. Forecasting is also a fruitful area for machine learning research, and pure and hybrid machine learning approaches have recently achieved state-of-the-art performance [1, 2]. In practice, forecasting involves a number of steps: we first need to specify, fit and select an appropriate model, and then evaluate and deploy it. There are various open-source toolboxes that help us implement these steps. However, most existing toolboxes are limited in important respects.

We estimate the Lipschitz constants of the gradient of a deep neural network and the network itself with respect to the full set of parameters. We first develop estimates for a deep feed-forward densely connected network and then, in a more general framework, for all neural networks that can be represented as solutions of controlled ordinary differential equations, where time appears as continuous depth. These estimates can be used to set the step size of stochastic gradient descent methods, which is illustrated for one example method.