We are concerned with learning models that generalize well to different \emph{unseen} domains. consider a worst-case formulation over data distributions near the source domain in feature space. Only using training from single distribution, we propose an iterative procedure augments dataset examples fictitious target is "hard" under current model. show our scheme adaptive augmentation method where append adversarial at each iteration. For softmax losses, data-dependent regularization behaves...

Recent works showed that Generative Adversarial Networks (GANs) can be successfully applied in unsupervised domain adaptation, where, given a labeled source dataset and an unlabeled target dataset, the goal is to train powerful classifiers for samples. In particular, it was shown GAN objective function used learn features indistinguishable from ones. this work, we extend framework by (i) forcing learned feature extractor domain-invariant, (ii) training through data augmentation space, namely...

We are concerned with the vulnerability of computer vision models to distributional shifts. formulate a combinatorial optimization problem that allows evaluating regions in image space where given model is more vulnerable, terms transformations applied input, and face it standard search algorithms. further embed this idea training procedure, we define new data augmentation rules according current most vulnerable to, over iterations. An empirical evaluation on classification semantic...

The potential of conjugated polymers to compete with inorganic materials in the field semiconductor is conditional on fine-tuning charge carriers mobility. latter closely related material morphology, and various studies have shown that bottleneck for transport connectivity between well-ordered crystallites, a high degree $\ensuremath{\pi}\text{\ensuremath{-}}\ensuremath{\pi}$ stacking, dispersed into disordered matrix. However, at this time there lack theoretical descriptions accounting link...

Most standard learning approaches lead to fragile models which are prone drift when sequentially trained on samples of a different nature—the well-known catastrophic forgetting issue. In particular, model consecutively learns from visual domains, it tends forget the past domains in favor most recent ones. this context, we show that one way learn inherently more robust against is domain randomization—for vision tasks, randomizing current domain’s distribution with heavy image manipulations....

We investigate and characterize the inherent resilience of conditional Generative Adversarial Networks (cGANs) against noise in their conditioning labels, exploit this fact context Unsupervised Domain Adaptation (UDA). In UDA, a classifier trained on labelled source set can be used to infer pseudo-labels unlabelled target set. However, will result significant amount misclassified examples (due well-known domain shift issue), which interpreted as injection ground-truth labels for show that...

In this paper, we present the first study that compares different models of Bayesian neural networks (BNNs) to predict posterior distribution cosmological parameters directly from cosmic microwave background (CMB) temperature and polarization maps. We focus our analysis on four methods sample weights network during training: Dropout, DropConnect, Reparameterization Trick (RT), Flipout. find Flipout outperforms all other regardless architecture used, provides tighter constraints for...

Dropout is a very effective way of regularizing neural networks. Stochastically “dropping out” units with certain probability discourages over-specific co-adaptations feature detectors, preventing overfitting and improving network generalization. Besides, can be interpreted as an approximate model aggregation technique, where exponential number smaller networks are averaged in order to get more powerful ensemble. In this paper, we show that using fixed dropout during training suboptimal...

Conventionally, AI models are thought to trade off explainability for lower accuracy. We develop a training strategy that not only leads more explainable system object classification, but as consequence, suffers no perceptible accuracy degradation. Explanations defined regions of visual evidence upon which deep classification network makes decision. This is represented in the form saliency map conveying how much each pixel contributed network's Our enforces periodic saliency-based feedback...

We propose a new problem formulation and corresponding evaluation framework to advance research on unsupervised domain adaptation for semantic image segmentation. The overall goal is fostering the development of adaptive learning systems that will continuously learn, without supervision, in ever-changing environments. Typical protocols study algorithms segmentation models are limited few domains, happens offline, human intervention generally required, at least annotate data hyperparameter...

We present a study of mobility field and temperature dependence for C60 with Kinetic Monte Carlo simulations. propose new scheme to take into account polarization effects in organic materials through atomic induced dipoles on nearby molecules. This leads an energy correction the single site energies external reorganization happening after each hopping. The inclusion allows us obtain good agreement experiments both dependence.

We study, within Marcus theory, the possibility of charge-transfer (CT) state splitting at organic interfaces and a subsequent transport free charge carriers to electrodes. As case study we analyze model anthracene-C60 interfaces. Kinetic Monte Carlo (KMC) simulations on cold CT were performed range applied electric fields, with fields angles interface simulate action field in bulk heterojunction (BHJ) interface. The results show that inclusion polarization our increases dissociation...

Upcoming experiments such as Hydrogen Epoch of Reionization Array(HERA) and the Square Kilometre Array (SKA) are intended to measure 21 cm signal over a wide range redshifts, representing an incredible opportunity in advancing our understanding about nature cosmic reionization. At same time these kind will present new challenges processing extensive amount data generated, calling for development automated methods capable precisely estimating physical parameters their uncertainties. In this...

We are interested in learning data-driven representations that can generalize well, even when trained on inherently biased data. In particular, we face the case where some attributes (bias) of data, if learned by model, severely compromise its generalization properties. tackle this problem through lens information theory, leveraging recent findings for a differentiable estimation mutual information. propose novel end-to-end optimization strategy, which simultaneously estimates and minimizes...

Abstract This paper aims at investigating the action prediction problem from a pure kinematic perspective. Specifically, we address of recognizing future actions, indeed human intentions, underlying same initial (and apparently unrelated) motor act. study is inspired by neuroscientific findings asserting that acts very onset are embedding information about intention with which performed, even when different intentions originate class movements. To demonstrate this claim in computational and...

In this article, we analyze the electric field dependence of hole mobility in disordered poly(p-phenylene vinylene). The charge carrier is obtained from Monte Carlo simulations. Depending on strength three regions can be identified: percolation region, correlation and inverted region. Each region characterized by a different conduction mechanism thus functional field. Earlier studies have highlighted that Poole-Frenkel law, which appears based type caused randomly distributed dipoles. This...

We have performed a multiscale approach to study the influence of peripheral substitution in semiconducting properties discotic liquid-crystalline triindoles. Charge carrier mobility as high 1.4 cm2 V-1 s-1 was experimentally reported for triindoles substituted with alkynyl chains on periphery (Gómez-Lor et al. Angew. Chem., Int. Ed., 2011, 50, 7399-7402). In this work, our goal is get deeper understanding both molecular electronic structure and microscopic factors affecting charge transport...

Adversarial Examples represent a serious problem affecting the security of machine learning systems. In this paper we focus on defense mechanism based reconstructing images before classification using an autoencoder. We experiment several types autoencoders and evaluate impact strategies such as injecting noise in input during training latent space at inference time.We tested models adversarial examples generated with Carlini-Wagner attack, white-box scenario stacked system composed by...