In this game, Nature selects a prior over distributions that generate labeled data consisting of features and an associated outcome, and the Predictor observes data sampled from a distribution drawn from this prior. The Predictor's objective is to learn a function that maps from a new feature to an estimate of the associated outcome. We establish that, under reasonable conditions, the Predictor has an optimal strategy that is equivariant to shifts and rescalings of the outcome and is invariant to permutations of the observations and to shifts, rescalings, and permutations of the features. We introduce a neural network architecture that satisfies these properties. The proposed strategy performs favorably compared to standard practice in both parametric and nonparametric experiments.

Adversarial training based on the minimax formulation is necessary for obtaining adversarial robustness of trained models. However, it is conservative or even pessimistic so that it sometimes hurts the natural generalization. In this paper, we raise a fundamental question---do we have to trade off natural generalization for adversarial robustness? We argue that adversarial training is to employ confident adversarial data for updating the current model. We propose a novel approach of friendly adversarial training (FAT): rather than employing most adversarial data maximizing the loss, we search for least adversarial (i.e., friendly adversarial) data minimizing the loss, among the adversarial data that are confidently misclassified. Our novel formulation is easy to implement by just stopping the most adversarial data searching algorithms such as PGD (projected gradient descent) early, which we call early-stopped PGD. Theoretically, FAT is justified by an upper bound of the adversarial risk. Empirically, early-stopped PGD allows us to answer the earlier question negatively---adversarial robustness can indeed be achieved without compromising the natural generalization.

The time it takes for a classifier to make an accurate prediction can be crucial in many behaviour recognition problems. For example, an autonomous vehicle should detect hazardous pedestrian behaviour early enough for it to take appropriate measures. In this context, we compare the switching linear dynamical system (SLDS) and a three-layered bi-directional long short-term memory (LSTM) neural network, which are applied to infer pedestrian behaviour from motion tracks. We show that, though the neural network model achieves an accuracy of 80%, it requires long sequences to achieve this (100 samples or more). The SLDS, has a lower accuracy of 74%, but it achieves this result with short sequences (10 samples). To our knowledge, such a comparison on sequence length has not been considered in the literature before. The results provide a key intuition of the suitability of the models in time-critical problems.

The proliferation of mobile phones, wearables and edge devices has led to an unprecedented growth in the generation of user interaction data. Systems which tap into the power of this rich data while respecting the privacy of users are geared to lead the next generation of intelligent applications and devices. The leading distributed learning framework in this setting is federated learning [30]. In federated learning, a number of client devices with privacy-sensitive data collaboratively learn a machine learning model under the orchestration of a central server, while keeping their data decentralized. This is achieved by pushing the actual computation to the devices while the server coordinates with the devices for aggregation of model updates.

Diabetes is a disease-causing high level of blood sugar. In type 1 Diabetes, body doesn't produce insulin, but if injected from external sources, will use it and in type 2, the body doesn't produce as well as use insulin. It is estimated that 30.3 million people of all ages in the US are suffering from Diabetes as of 2015, out of which 7.2 million are unaware[1]. As of 2016, it is ranked seventh in the list of global causes of mortality. Diabetes can be an underlying cause for many cardiovascular diseases, retinopathy, and nephropathy leading to frequent readmission in the hospital. The Centers for Medicare and Medicaid Services(CMS) labeled a 30-day readmission rate as a measure of healthcare quality offered by the hospital in order to provide the best inpatient care and improve the healthcare quality. Hospitals with high readmission rates will be penalized as per the Patient Protection and Affordable Care Act(ACA) of 2010[2]. During the recent studies[19], it was observed that a 30-day readmission rate for patients with Diabetes ranges between 14.4%-22.7%,

To make weather/climate modeling computationally affordable, small-scale processes are usually represented in terms of the large-scale, explicitly-resolved processes using physics-based or semi-empirical parameterization schemes. Another approach, computationally more demanding but often more accurate, is super-parameterization (SP), which involves integrating the equations of small-scale processes on high-resolution grids embedded within the low-resolution grids of large-scale processes. Recently, studies have used machine learning (ML) to develop data-driven parameterization (DD-P) schemes. Here, we propose a new approach, data-driven SP (DD-SP), in which the equations of the small-scale processes are integrated data-drivenly using ML methods such as recurrent neural networks. Employing multi-scale Lorenz 96 systems as testbed, we compare the cost and accuracy (in terms of both short-term prediction and long-term statistics) of parameterized low-resolution (LR), SP, DD-P, and DD-SP models. We show that with the same computational cost, DD-SP substantially outperforms LR, and is better than DD-P, particularly when scale separation is lacking. DD-SP is much cheaper than SP, yet its accuracy is the same in reproducing long-term statistics and often comparable in short-term forecasting. We also investigate generalization, finding that when models trained on data from one system are applied to a system with different forcing (e.g., more chaotic), the models often do not generalize, particularly when the short-term prediction accuracy is examined. But we show that transfer-learning, which involves re-training the data-driven model with a small amount of data from the new system, significantly improves generalization. Potential applications of DD-SP and transfer-learning in climate/weather modeling and the expected challenges are discussed.

Resistive crossbars have attracted significant interest in the design of Deep Neural Network (DNN) accelerators due to their ability to natively execute massively parallel vector-matrix multiplications within dense memory arrays. However, crossbar-based computations face a major challenge due to a variety of device and circuit-level non-idealities, which manifest as errors in the vector-matrix multiplications and eventually degrade DNN accuracy. To address this challenge, there is a need for tools that can model the functional impact of non-idealities on DNN training and inference. Existing efforts towards this goal are either limited to inference, or are too slow to be used for large-scale DNN training. We propose TxSim, a fast and customizable modeling framework to functionally evaluate DNN training on crossbar-based hardware considering the impact of non-idealities. The key features of TxSim that differentiate it from prior efforts are: (i) It comprehensively models non-idealities during all training operations (forward propagation, backward propagation, and weight update) and (ii) it achieves computational efficiency by mapping crossbar evaluations to well-optimized BLAS routines and incorporates speedup techniques to further reduce simulation time with minimal impact on accuracy. TxSim achieves orders-of-magnitude improvement in simulation speed over prior works, and thereby makes it feasible to evaluate training of large-scale DNNs on crossbars. Our experiments using TxSim reveal that the accuracy degradation in DNN training due to non-idealities can be substantial (3%-10%) for large-scale DNNs, underscoring the need for further research in mitigation techniques. We also analyze the impact of various device and circuit-level parameters and the associated non-idealities to provide key insights that can guide the design of crossbar-based DNN training accelerators.

Deep neural networks have demonstrated state-of-the-art performance on many classification tasks. However, they have no inherent capability to recognize when their predictions are wrong. There have been several efforts in the recent past to detect natural errors but the suggested mechanisms pose additional energy requirements. To address this issue, we propose an ensemble of classifiers at hidden layers to enable energy efficient detection of natural errors. In particular, we append Relevant-features based Auxiliary Cells (RACs) which are class specific binary linear classifiers trained on relevant features. The consensus of RACs is used to detect natural errors. Based on combined confidence of RACs, classification can be terminated early, thereby resulting in energy efficient detection. We demonstrate the effectiveness of our technique on various image classification datasets such as CIFAR-10, CIFAR-100 and Tiny-ImageNet.

Extraneous variables are variables that are irrelevant for a certain task, but heavily affect the distribution of the available data. In this work, we show that the presence of such variables can degrade the performance of deep-learning models. We study three datasets where there is a strong influence of known extraneous variables: classification of upper-body movements in stroke patients, annotation of surgical activities, and recognition of corrupted images. Models trained with batch normalization learn features that are highly dependent on the extraneous variables. In batch normalization, the statistics used to normalize the features are learned from the training set and fixed at test time, which produces a mismatch in the presence of varying extraneous variables. We demonstrate that estimating the feature statistics adaptively during inference, as in instance normalization, addresses this issue, producing normalized features that are more robust to changes in the extraneous variables. This results in a significant gain in performance for different network architectures and choices of feature statistics.

The communication of gradients is costly for training deep neural networks with multiple devices in computer vision applications. In particular, the growing size of deep learning models leads to higher communication overheads that defy the ideal linear training speedup regarding the number of devices. Gradient quantization is one of the common methods to reduce communication costs. However, it can lead to quantization error in the training and result in model performance degradation. In this work, we deduce the optimal condition of both the binary and multi-level gradient quantization for \textbf{ANY} gradient distribution. Based on the optimal condition, we develop two novel quantization schemes: biased BinGrad and unbiased ORQ for binary and multi-level gradient quantization respectively, which dynamically determine the optimal quantization levels. Extensive experimental results on CIFAR and ImageNet datasets with several popular convolutional neural networks show the superiority of our proposed methods.