We present an empirical analysis of the state-of-the-art systems for referring expression recognition -- the task of identifying the object in an image referred to by a natural language expression -- with the goal of gaining insight into how these systems reason about language and vision. Surprisingly, we find strong evidence that even sophisticated and linguistically-motivated models for this task may ignore the linguistic structure, instead relying on shallow correlations introduced by unintended biases in the data selection and annotation process. For example, we show that a system trained and tested on the input image $\textit{without the input referring expression}$ can achieve a precision of 71.2% in top-2 predictions. Furthermore, a system that predicts only the object category given the input can achieve a precision of 84.2% in top-2 predictions. These surprisingly positive results for what should be deficient prediction scenarios suggest that careful analysis of what our models are learning -- and further, how our data is constructed -- is critical as we seek to make substantive progress on grounded language tasks.

Performance of neural networks can be significantly improved by encoding known invariance for particular tasks. Many image classification tasks, such as those related to cellular imaging, exhibit invariance to rotation. We present a novel scheme using the magnitude response of the 2D-discrete-Fourier transform (2D-DFT) to encode rotational invariance in neural networks, along with a new, efficient convolutional scheme for encoding rotational equivariance throughout convolutional layers. We implemented this scheme for several image classification tasks and demonstrated improved performance, in terms of classification accuracy, time required to train the model, and robustness to hyperparameter selection, over a standard CNN and another state-of-the-art method.

Deep neural networks require large amounts of resources which makes them hard to use on resource constrained devices such as Internet-of-things devices. Offloading the computations to the cloud can circumvent these constraints but introduces a privacy risk since the operator of the cloud is not necessarily trustworthy. We propose a technique that obfuscates the data before sending it to the remote computation node. The obfuscated data is unintelligible for a human eavesdropper but can still be classified with a high accuracy by a neural network trained on unobfuscated images.

Deep neural networks (DNNs) have become the state-of-the-art technique for machine learning tasks in various applications. However, due to their size and the computational complexity, large DNNs are not readily deployable on edge devices in real-time. To manage complexity and accelerate computation, network compression techniques based on pruning and quantization have been proposed and shown to be effective in reducing network size. However, such network compression can result in irregular matrix structures that are mismatched with modern hardware-accelerated platforms, such as graphics processing units (GPUs) designed to perform the DNN matrix multiplications in a structured (block-based) way. We propose MPDCompress, a DNN compression algorithm based on matrix permutation decomposition via random mask generation. In-training application of the masks molds the synaptic weight connection matrix to a sub-graph separation format. Aided by the random permutations, a hardware-desirable block matrix is generated, allowing for a more efficient implementation and compression of the network. To show versatility, we empirically verify MPDCompress on several network models, compression rates, and image datasets. On the LeNet 300-100 model (MNIST dataset), Deep MNIST, and CIFAR10, we achieve 10 X network compression with less than 1% accuracy loss compared to non-compressed accuracy performance. On AlexNet for the full ImageNet ILSVRC-2012 dataset, we achieve 8 X network compression with less than 1% accuracy loss, with top-5 and top-1 accuracies of 79.6% and 56.4%, respectively. Finally, we observe that the algorithm can offer inference speedups across various hardware platforms, with 4 X faster operation achieved on several mobile GPUs.

We propose a two-stage neural model to tackle question generation from documents. First, our model estimates the probability that word sequences in a document are ones that a human would pick when selecting candidate answers by training a neural key-phrase extractor on the answers in a question-answering corpus. Predicted key phrases then act as target answers and condition a sequence-to-sequence question-generation model with a copy mechanism. Empirically, our key-phrase extraction model significantly outperforms an entity-tagging baseline and existing rule-based approaches. We further demonstrate that our question generation system formulates fluent, answerable questions from key phrases. This two-stage system could be used to augment or generate reading comprehension datasets, which may be leveraged to improve machine reading systems or in educational settings.

Deep networks have gained immense popularity in Computer Vision and other fields in the past few years due to their remarkable performance on recognition/classification tasks surpassing the state-of-the art. One of the keys to their success lies in the richness of the automatically learned features. In order to get very good accuracy, one popular option is to increase the depth of the network. Training such a deep network is however infeasible or impractical with moderate computational resources and budget. The other alternative to increase the performance is to learn multiple weak classifiers and boost their performance using a boosting algorithm or a variant thereof. But, one of the problems with boosting algorithms is that they require a re-training of the networks based on the misclassified samples. Motivated by these problems, in this work we propose an aggregation technique which combines the output of multiple weak classifiers. We formulate the aggregation problem using a mixture model fitted to the trained classifier outputs. Our model does not require any re-training of the `weak' networks and is computationally very fast (takes $<30$ seconds to run in our experiments). Thus, using a less expensive training stage and without doing any re-training of networks, we experimentally demonstrate that it is possible to boost the performance by $12\%$. Furthermore, we present experiments using hand-crafted features and improved the classification performance using the proposed aggregation technique. One of the major advantages of our framework is that our framework allows one to combine features that are very likely to be of distinct dimensions since they are extracted using different networks/algorithms. Our experimental results demonstrate a significant performance gain from the use of our aggregation technique at a very small computational cost.

Performance monitoring, anomaly detection, and root-cause analysis in complex cyber-physical systems (CPSs) are often highly intractable due to widely diverse operational modes, disparate data types, and complex fault propagation mechanisms. This paper presents a new data-driven framework for root-cause analysis, based on a spatiotemporal graphical modeling approach built on the concept of symbolic dynamics for discovering and representing causal interactions among sub-systems of complex CPSs. We formulate the root-cause analysis problem as a minimization problem via the proposed inference based metric and present two approximate approaches for root-cause analysis, namely the sequential state switching ($S^3$, based on free energy concept of a restricted Boltzmann machine, RBM) and artificial anomaly association ($A^3$, a classification framework using deep neural networks, DNN). Synthetic data from cases with failed pattern(s) and anomalous node(s) are simulated to validate the proposed approaches. Real dataset based on Tennessee Eastman process (TEP) is also used for comparison with other approaches. The results show that: (1) $S^3$ and $A^3$ approaches can obtain high accuracy in root-cause analysis under both pattern-based and node-based fault scenarios, in addition to successfully handling multiple nominal operating modes, (2) the proposed tool-chain is shown to be scalable while maintaining high accuracy, and (3) the proposed framework is robust and adaptive in different fault conditions and performs better in comparison with the state-of-the-art methods.

Learning from 3D Data is a fascinating idea which is well explored and studied in computer vision. This allows one to learn from very sparse LiDAR data, point cloud data as well as 3D objects in terms of CAD models and surfaces etc. Most of the approaches to learn from such data are limited to uniform 3D volume occupancy grids or octree representations. A major challenge in learning from 3D data is that one needs to define a proper resolution to represent it in a voxel grid and this becomes a bottleneck for the learning algorithms. Specifically, while we focus on learning from 3D data, a fine resolution is very important to capture key features in the object and at the same time the data becomes sparser as the resolution becomes finer. There are numerous applications in computer vision where a multi-resolution representation is used instead of a uniform grid representation in order to make the applications memory efficient. Though such methods are difficult to learn from, they are much more efficient in representing 3D data. In this paper, we explore the challenges in learning from such data representation. In particular, we use a multi-level voxel representation where we define a coarse voxel grid that contains information of important voxels(boundary voxels) and multiple fine voxel grids corresponding to each significant voxel of the coarse grid. A multi-level voxel representation can capture important features in the 3D data in a memory efficient way in comparison to an octree representation. Consequently, learning from a 3D object with high resolution, which is paramount in feature recognition, is made efficient.

Making predictions of future frames is a critical challenge in autonomous driving research. Most of the existing methods for video prediction attempt to generate future frames in simple and fixed scenes. In this paper, we propose a novel and effective optical flow conditioned method for the task of video prediction with an application to complex urban scenes. In contrast with previous work, the prediction model only requires video sequences and optical flow sequences for training and testing. Our method uses the rich spatial-temporal features in video sequences. The method takes advantage of the motion information extracting from optical flow maps between neighbor images as well as previous images. Empirical evaluations on the KITTI dataset and the Cityscapes dataset demonstrate the effectiveness of our method.

The understanding of variations in genome sequences assists us in identifying people who are predisposed to common diseases, solving rare diseases, and finding the corresponding population group of the individuals from a larger population group. Although classical machine learning techniques allow researchers to identify groups (i.e. clusters) of related variables, the accuracy, and effectiveness of these methods diminish for large and high-dimensional datasets such as the whole human genome. On the other hand, deep neural network architectures (the core of deep learning) can better exploit large-scale datasets to build complex models. In this paper, we use the K-means clustering approach for scalable genomic data analysis aiming towards clustering genotypic variants at the population scale. Finally, we train a deep belief network (DBN) for predicting the geographic ethnicity. We used the genotype data from the 1000 Genomes Project, which covers the result of genome sequencing for 2504 individuals from 26 different ethnic origins and comprises 84 million variants. Our experimental results, with a focus on accuracy and scalability, show the effectiveness and superiority compared to the state-of-the-art.