Passive visual systems typically fail to recognize objects in the amodal setting where they are heavily occluded. In contrast, humans and other embodied agents have the ability to move in the environment, and actively control the viewing angle to better understand object shapes and semantics. In this work, we introduce the task of Embodied Visual Recognition (EVR): An agent is instantiated in a 3D environment close to an occluded target object, and is free to move in the environment to perform object classification, amodal object localization, and amodal object segmentation. To address this, we develop a new model called Embodied Mask R-CNN, for agents to learn to move strategically to improve their visual recognition abilities. We conduct experiments using the House3D environment. Experimental results show that: 1) agents with embodiment (movement) achieve better visual recognition performance than passive ones; 2) in order to improve visual recognition abilities, agents can learn strategical moving paths that are different from shortest paths.

Several tasks in machine learning are evaluated using non-differentiable metrics such as mean average precision or Spearman correlation. However, their non-differentiability prevents from using them as objective functions in a learning framework. Surrogate and relaxation methods exist but tend to be specific to a given metric. In the present work, we introduce a new method to learn approximations of such non-differentiable objective functions. Our approach is based on a deep architecture that approximates the sorting of arbitrary sets of scores. It is trained virtually for free using synthetic data. This sorting deep (SoDeep) net can then be combined in a plug-and-play manner with existing deep architectures. We demonstrate the interest of our approach in three different tasks that require ranking: Cross-modal text-image retrieval, multi-label image classification and visual memorability ranking. Our approach yields very competitive results on these three tasks, which validates the merit and the flexibility of SoDeep as a proxy for sorting operation in ranking-based losses.

We present a method to extract a weighted finite automaton (WFA) from a recurrent neural network (RNN). Our algorithm is based on the WFA learning algorithm by Balle and Mohri, which is in turn an extension of Angluin's classic \lstar algorithm. Our technical novelty is in the use of \emph{regression} methods for the so-called equivalence queries, thus exploiting the internal state space of an RNN. This way we achieve a quantitative extension of the recent work by Weiss, Goldberg and Yahav that extracts DFAs. Experiments demonstrate that our algorithm's practicality.

Due to the lack of enough training data and high computational cost to train a deep neural network from scratch, transfer learning has been extensively used in many deep-neural-network-based applications, such as face recognition, image classification, speech recognition, etc. A commonly-used transfer learning approach involves taking a part of a pre-trained model, adding a few layers at the end, and re-training the new layers with a small dataset. This approach, while efficient and widely used, imposes a security vulnerability because the pre-trained models used in transfer learning are usually available publicly to everyone, including potential attackers. In this paper, we show that without any additional knowledge other than the pre-trained model, an attacker can launch an effective and efficient brute force attack that can craft instances of input to trigger each target class with high confidence. Note that we assume that the attacker does not have access to any target-specific information, including samples from target classes, re-trained model, and probabilities assigned by Softmax to each class, and thus called target-agnostic attack. These assumptions render all previous attacks impractical, to the best of our knowledge. To evaluate the proposed attack, we perform a set of experiments on face recognition and speech recognition tasks and show the effectiveness of the attack. Our work sheds light on a fundamental security challenge of transfer learning in deep neural networks.

With the popularity of social networks, and e-commerce websites, sentiment analysis has become a more active area of research in the past few years. On a high level, sentiment analysis tries to understand the public opinion about a specific product or topic, or trends from reviews or tweets. Sentiment analysis plays an important role in better understanding customer/user opinion, and also extracting social/political trends. There has been a lot of previous works for sentiment analysis, some based on hand-engineering relevant textual features, and others based on different neural network architectures. In this work, we present a model based on an ensemble of long-short-term-memory (LSTM), and convolutional neural network (CNN), one to capture the temporal information of the data, and the other one to extract the local structure thereof. Through experimental results, we show that using this ensemble model we can outperform both individual models. We are also able to achieve a very high accuracy rate compared to the previous works.

For the tensor PCA (principal component analysis) problem, we propose a new hierarchy of algorithms that are increasingly powerful yet require increasing runtime. Our hierarchy is analogous to the sum-of-squares (SOS) hierarchy but is instead inspired by statistical physics and related algorithms such as belief propagation and AMP (approximate message passing). Our level-$\ell$ algorithm can be thought of as a (linearized) message-passing algorithm that keeps track of $\ell$-wise dependencies among the hidden variables. Specifically, our algorithms are spectral methods based on the Kikuchi Hessian matrix, which generalizes the well-studied Bethe Hessian matrix to the higher-order Kikuchi free energies. It is known that AMP, the flagship algorithm of statistical physics, has substantially worse performance than SOS for tensor PCA. In this work we `redeem' the statistical physics approach by showing that our hierarchy gives a polynomial-time algorithm matching the performance of SOS. Our hierarchy also yields a continuum of subexponential-time algorithms, and we prove that these achieve the same (conjecturally optimal) tradeoff between runtime and statistical power as SOS. Our results hold for even-order tensors, and we conjecture that they also hold for odd-order tensors. Our methods suggest a new avenue for systematically obtaining optimal algorithms for Bayesian inference problems, and our results constitute a step toward unifying the statistical physics and sum-of-squares approaches to algorithm design.

Spherical data is found in many applications. By modeling the discretized sphere as a graph, we can accommodate non-uniformly distributed, partial, and changing samplings. Moreover, graph convolutions are computationally more efficient than spherical convolutions. As equivariance is desired to exploit rotational symmetries, we discuss how to approach rotation equivariance using the graph neural network introduced in Defferrard et al. (2016). Experiments show good performance on rotation-invariant learning problems. Code and examples are available at https://github.com/SwissDataScienceCenter/DeepSphere

We consider network sparsification as an $L_0$-norm regularized binary optimization problem, where each unit of a neural network (e.g., weight, neuron, or channel, etc.) is attached with a stochastic binary gate, whose parameters are jointly optimized with original network parameters. The Augment-Reinforce-Merge (ARM), a recently proposed unbiased gradient estimator, is investigated for this binary optimization problem. Compared to the hard concrete gradient estimator from Louizos et al., ARM demonstrates superior performance of pruning network architectures while retaining almost the same accuracies of baseline methods. Similar to the hard concrete estimator, ARM also enables conditional computation during model training but with improved effectiveness due to the exact binary stochasticity. Thanks to the flexibility of ARM, many smooth or non-smooth parametric functions, such as scaled sigmoid or hard sigmoid, can be used to parameterize this binary optimization problem and the unbiasness of the ARM estimator is retained, while the hard concrete estimator has to rely on the hard sigmoid function to achieve conditional computation and thus accelerated training. Extensive experiments on multiple public datasets demonstrate state-of-the-art pruning rates with almost the same accuracies of baseline methods. The resulting algorithm $L_0$-ARM sparsifies the Wide-ResNet models on CIFAR-10 and CIFAR-100 while the hard concrete estimator cannot. We plan to release our code to facilitate the research in this area.

Structured latent attribute models (SLAMs) are a special family of discrete latent variable models widely used in social and biological sciences. This paper considers the problem of learning significant attribute patterns from a SLAM with potentially high-dimensional configurations of the latent attributes. We address the theoretical identifiability issue, propose a penalized likelihood method for the selection of the attribute patterns, and further establish the selection consistency in such an overfitted SLAM with diverging number of latent patterns. The good performance of the proposed methodology is illustrated by simulation studies and two real datasets in educational assessment.

Speech-related Brain Computer Interfaces (BCI) aim primarily at finding an alternative vocal communication pathway for people with speaking disabilities. As a step towards full decoding of imagined speech from active thoughts, we present a BCI system for subject-independent classification of phonological categories exploiting a novel deep learning based hierarchical feature extraction scheme. To better capture the complex representation of high-dimensional electroencephalography (EEG) data, we compute the joint variability of EEG electrodes into a channel cross-covariance matrix. We then extract the spatio-temporal information encoded within the matrix using a mixed deep neural network strategy. Our model framework is composed of a convolutional neural network (CNN), a long-short term network (LSTM), and a deep autoencoder. We train the individual networks hierarchically, feeding their combined outputs in a final gradient boosting classification step. Our best models achieve an average accuracy of 77.9% across five different binary classification tasks, providing a significant 22.5% improvement over previous methods. As we also show visually, our work demonstrates that the speech imagery EEG possesses significant discriminative information about the intended articulatory movements responsible for natural speech synthesis.