BERT, which stands for Bidirectional Encoder Representations from Transformers, is a recently introduced language representation model based upon the transfer learning paradigm. We extend its fine-tuning procedure to address one of its major limitations - applicability to inputs longer than a few hundred words, such as transcripts of human call conversations. Our method is conceptually simple. We segment the input into smaller chunks and feed each of them into the base model. Then, we propagate each output through a single recurrent layer, or another transformer, followed by a softmax activation. We obtain the final classification decision after the last segment has been consumed. We show that both BERT extensions are quick to fine-tune and converge after as little as 1 epoch of training on a small, domain-specific data set. We successfully apply them in three different tasks involving customer call satisfaction prediction and topic classification, and obtain a significant improvement over the baseline models in two of them.

Noninvasive MR-guided focused ultrasound (MRgFUS) treatments are promising alternatives to the surgical removal of malignant tumors. A significant challenge is assessing the treated tissue immediately after MRgFUS procedures. Although current clinical assessment uses the immediate nonperfused volume (NPV) biomarker derived from contrast enhanced imaging, the use of contrast agent prevents continuing MRgFUS treatment if margins are not adequate. In addition, the NPV has been shown to provide variable accuracy for the true treatment outcome as evaluated by follow-up biomarkers. This work presents a novel, noncontrast, learned multiparametric MR biomarker that is conducive for intratreatment assessment. MRgFUS ablations were performed in a rabbit VX2 tumor model. Multiparametric MRI was obtained both during and immediately after the MRgFUS ablation, as well as during follow-up imaging. Segmentation of the NPV obtained during follow-up imaging was used to train a neural network on noncontrast multiparametric MR images. The NPV follow-up segmentation was registered to treatment-day images using a novel volume-conserving registration algorithm, allowing a voxel-wise correlation between imaging sessions. Contrasted with state-of-the-art registration algorithms that change the average volume by 16.8%, the presented volume-conserving registration algorithm changes the average volume by only 0.28%. After registration, the learned multiparametric MR biomarker predicted the follow-up NPV with an average DICE coefficient of 0.71, outperforming the DICE coefficient of 0.53 from the current standard of NPV obtained immediately after the ablation treatment. Noncontrast multiparametric MR imaging can provide a more accurate prediction of treated tissue immediately after treatment. Noncontrast assessment of MRgFUS procedures will potentially lead to more efficacious MRgFUS ablation treatments.

In this paper, we propose a novel Reinforcement Learning approach for solving the Active Information Acquisition problem, which requires an agent to choose a sequence of actions in order to acquire information about a process of interest using on-board sensors. The classic challenges in the information acquisition problem are the dependence of a planning algorithm on known models and the difficulty of computing information-theoretic cost functions over arbitrary distributions. In contrast, the proposed framework of reinforcement learning does not require any knowledge on models and alleviates the problems during an extended training stage. It results in policies that are efficient to execute online and applicable for real-time control of robotic systems. Furthermore, the state-of-the-art planning methods are typically restricted to short horizons, which may become problematic with local minima. Reinforcement learning naturally handles the issue of planning horizon in information problems as it maximizes a discounted sum of rewards over a long finite or infinite time horizon. We discuss the potential benefits of the proposed framework and compare the performance of the novel algorithm to an existing information acquisition method for multi-target tracking scenarios.

Often we wish to transfer representational knowledge from one neural network to another. Examples include distilling a large network into a smaller one, transferring knowledge from one sensory modality to a second, or ensembling a collection of models into a single estimator. Knowledge distillation, the standard approach to these problems, minimizes the KL divergence between the probabilistic outputs of a teacher and student network. We demonstrate that this objective ignores important structural knowledge of the teacher network. This motivates an alternative objective by which we train a student to capture significantly more information in the teacher's representation of the data. We formulate this objective as contrastive learning. Experiments demonstrate that our resulting new objective outperforms knowledge distillation and other cutting-edge distillers on a variety of knowledge transfer tasks, including single model compression, ensemble distillation, and cross-modal transfer. Our method sets a new state-of-the-art in many transfer tasks, and sometimes even outperforms the teacher network when combined with knowledge distillation. Code: http://github.com/HobbitLong/RepDistiller.

Adversarial attacks and defenses are currently active areas of research for the deep learning community. A recent review paper divided the defense approaches into three categories; gradient masking, robust optimization, and adversarial example detection. We divide gradient masking and robust optimization differently: (1) increasing intra-class compactness and inter-class separation of the feature vectors improves adversarial robustness, and (2) marginalization or removal of non-robust image features also improves adversarial robustness. By reframing these topics differently, we provide a fresh perspective that provides insight into the underlying factors that enable training more robust networks and can help inspire novel solutions. In addition, there are several papers in the literature of adversarial defenses that claim there is a cost for adversarial robustness, or a trade-off between robustness and accuracy but, under this proposed taxonomy, we hypothesis that this is not universal. We follow up on our taxonomy with several challenges to the deep learning research community that builds on the connections and insights in this paper.

The forecast of tropical cyclone trajectories is crucial for the protection of people and property. Although forecast dynamical models can provide high-precision short-term forecasts, they are computationally demanding, and current statistical forecasting models have much room for improvement given that the database of past hurricanes is constantly growing. Machine learning methods, that can capture non-linearities and complex relations, have only been scarcely tested for this application. We propose a neural network model fusing past trajectory data and reanalysis atmospheric images (wind and pressure 3D fields). We use a moving frame of reference that follows the storm center for the 24h tracking forecast. The network is trained to estimate the longitude and latitude displacement of tropical cyclones and depressions from a large database from both hemispheres (more than 3000 storms since 1979, sampled at a 6 hour frequency). The advantage of the fused network is demonstrated and a comparison with current forecast models shows that deep learning methods could provide a valuable and complementary prediction. Moreover, our method can give a forecast for a new storm in a few seconds, which is an important asset for real-time forecasts compared to traditional forecasts.

Deep learning methods have received increasing interest by the remote sensing community for multi-temporal land cover classification in recent years. Convolutional Neural networks that elementwise compare a time series with learned kernels, and recurrent neural networks that sequentially process temporal data have dominated the state-of-the-art in the classification of vegetation from satellite time series. Self-attention allows a neural network to selectively extract features from specific times in the input sequence thus suppressing non-classification relevant information. Today, self-attention based neural networks dominate the state-of-the-art in natural language processing but are hardly explored and tested in the remote sensing context. In this work, we embed self-attention in the canon of deep learning mechanisms for satellite time series classification for vegetation modeling and crop type identification. We compare it quantitatively to convolution, and recurrence and test four models that each exclusively relies on one of these mechanisms. The models are trained to identify the type of vegetation on crop parcels using raw and preprocessed Sentinel 2 time series over one entire year. To obtain an objective measure we find the best possible performance for each of the models by a large-scale hyperparameter search with more than 2400 validation runs. Beyond the quantitative comparison, we qualitatively analyze the models by an easy-to-implement, but yet effective feature importance analysis based on gradient back-propagation that exploits the differentiable nature of deep learning models. Finally, we look into the self-attention transformer model and visualize attention scores as bipartite graphs in the context of the input time series and a low-dimensional representation of internal hidden states using t-distributed stochastic neighborhood embedding (t-SNE).

This paper introduces a divide-and-conquer inspired adversarial learning (DA-CAL) approach for photo enhancement. The key idea is to decompose the photo enhancement process into hierarchically multiple sub-problems, which can be better conquered from bottom to up. On the top level, we propose a perception-based division to learn additive and multiplicative components, required to translate a low-quality image or video into its high-quality counterpart. On the intermediate level, we use a frequency-based division with generative adversarial network (GAN) to weakly supervise the photo enhancement process. On the lower level, we design a dimension-based division that enables the GAN model to better approximates the distribution distance on multiple independent one-dimensional data to train the GAN model. While considering all three hierarchies, we develop multiscale and recurrent training approaches to optimize the image and video enhancement process in a weakly-supervised manner. Both quantitative and qualitative results clearly demonstrate that the proposed DACAL achieves the state-of- the-art performance for high-resolution image and video enhancement. Despite many mobile camera technological advances we have today, our captured images often still come with limited dynamic range, undesirable color rendition, and unsatisfactory texture sharpness. Among many possible causes, low-light environments and under/overexposed regions usually introduce severe lack of texture details and low-dynamic range coverage, respectively. Another critical issue is the amplification (during the enhancement process) of noise in the dark and/or texture-less regions, where the enhancement may not even be necessary.

Convolutional neural networks are state-of-the-art for various segmentation tasks. While for 2D images these networks are also computationally efficient, 3D convolutions have huge storage requirements and therefore, end-to-end training is limited by GPU memory and data size. To overcome this issue, we introduce a network structure for volumetric data without 3D convolution layers. The main idea is to include projections from different directions to transform the volumetric data to a sequence of images, where each image contains information of the full data. We then apply 2D convolutions to these projection images and lift them again to volumetric data using a trainable reconstruction algorithm. The proposed architecture can be applied end-to-end to very large data volumes without cropping or sliding-window techniques. For a tested sparse binary segmentation task, it outperforms already known standard approaches and is more resistant to generation of artefacts.

We propose a novel neural architecture search algorithm via reinforcement learning by decoupling structure and operation search processes. Our approach samples candidate models from the multinomial distribution on the policy vectors defined on the two search spaces independently. The proposed technique improves the efficiency of architecture search process significantly compared to the conventional methods based on reinforcement learning with the RNN controllers while achieving competitive accuracy and model size in target tasks. Our policy vectors are easily interpretable throughout the training procedure, which allows to analyze the search progress and the discovered architectures; the black-box characteristics of the RNN controllers hamper understanding training progress in terms of policy parameter updates. Our experiments demonstrate outstanding performance compared to the state-of-the-art methods with a fraction of search cost.