Deep learning has proven itself as a successful set of models for learning useful semantic representations of data. These, however, are mostly implicitly learned as part of a classification task. In this paper we propose the triplet network model, which aims to learn useful representations by distance comparisons. A similar model was defined by Wang et al. (2014), tailor made for learning a ranking for image information retrieval. Here we demonstrate using various datasets that our model learns a better representation than that of its immediate competitor, the Siamese network. We also discuss future possible usage as a framework for unsupervised learning.

Several machine learning models, including neural networks, consistently misclassify adversarial examples---inputs formed by applying small but intentionally worst-case perturbations to examples from the dataset, such that the perturbed input results in the model outputting an incorrect answer with high confidence. Early attempts at explaining this phenomenon focused on nonlinearity and overfitting. We argue instead that the primary cause of neural networks' vulnerability to adversarial perturbation is their linear nature. This explanation is supported by new quantitative results while giving the first explanation of the most intriguing fact about them: their generalization across architectures and training sets. Moreover, this view yields a simple and fast method of generating adversarial examples. Using this approach to provide examples for adversarial training, we reduce the test set error of a maxout network on the MNIST dataset.

Deep learning can achieve outstanding results in various fields. However, it requires so significant computational power that graphics processing units (GPUs) and/or numerous computers are often required for the practical application. We have developed a new distributed calculation framework called "Sashimi" that allows any computer to be used as a distribution node only by accessing a website. We have also developed a new JavaScript neural network framework called "Sukiyaki" that uses general purpose GPUs with web browsers. Sukiyaki performs 30 times faster than a conventional JavaScript library for deep convolutional neural networks (deep CNNs) learning. The combination of Sashimi and Sukiyaki, as well as new distribution algorithms, demonstrates the distributed deep learning of deep CNNs only with web browsers on various devices. The libraries that comprise the proposed methods are available under MIT license at http://mil-tokyo.github.io/.

This paper presents a novel approach to speaker subspace modelling based on Gaussian-Binary Restricted Boltzmann Machines (GRBM). The proposed model is based on the idea of shared factors as in the Probabilistic Linear Discriminant Analysis (PLDA). GRBM hidden layer is divided into speaker and channel factors, herein the speaker factor is shared over all vectors of the speaker. Then Maximum Likelihood Parameter Estimation (MLE) for proposed model is introduced. Various new scoring techniques for speaker verification using GRBM are proposed. The results for NIST i-vector Challenge 2014 dataset are presented.

Conditional restricted Boltzmann machines are undirected stochastic neural networks with a layer of input and output units connected bipartitely to a layer of hidden units. These networks define models of conditional probability distributions on the states of the output units given the states of the input units, parametrized by interaction weights and biases. We address the representational power of these models, proving results their ability to represent conditional Markov random fields and conditional distributions with restricted supports, the minimal size of universal approximators, the maximal model approximation errors, and on the dimension of the set of representable conditional distributions. We contribute new tools for investigating conditional probability models, which allow us to improve the results that can be derived from existing work on restricted Boltzmann machine probability models.

We recently have witnessed many ground-breaking results in machine learning and computer vision, generated by using deep convolutional neural networks (CNN). While the success mainly stems from the large volume of training data and the deep network architectures, the vector processing hardware (e.g. GPU) undisputedly plays a vital role in modern CNN implementations to support massive computation. Though much attention was paid in the extent literature to understand the algorithmic side of deep CNN, little research was dedicated to the vectorization for scaling up CNNs. In this paper, we studied the vectorization process of key building blocks in deep CNNs, in order to better understand and facilitate parallel implementation. Key steps in training and testing deep CNNs are abstracted as matrix and vector operators, upon which parallelism can be easily achieved. We developed and compared six implementations with various degrees of vectorization with which we illustrated the impact of vectorization on the speed of model training and testing. Besides, a unified CNN framework for both high-level and low-level vision tasks is provided, along with a vectorized Matlab implementation with state-of-the-art speed performance.

Tag recommendation has become one of the most important ways of organizing and indexing online resources like articles, movies, and music. Since tagging information is usually very sparse, effective learning of the content representation for these resources is crucial to accurate tag recommendation. Recently, models proposed for tag recommendation, such as collaborative topic regression and its variants, have demonstrated promising accuracy. However, a limitation of these models is that, by using topic models like latent Dirichlet allocation as the key component, the learned representation may not be compact and effective enough. Moreover, since relational data exist as an auxiliary data source in many applications, it is desirable to incorporate such data into tag recommendation models. In this paper, we start with a deep learning model called stacked denoising autoencoder (SDAE) in an attempt to learn more effective content representation. We propose a probabilistic formulation for SDAE and then extend it to a relational SDAE (RSDAE) model. RSDAE jointly performs deep representation learning and relational learning in a principled way under a probabilistic framework. Experiments conducted on three real datasets show that both learning more effective representation and learning from relational data are beneficial steps to take to advance the state of the art.

Social media platforms are often used by people to express their needs and desires. Such data offer great opportunities to identify users’ consumption intention from user-generated contents, so that better tailored products or services can be recommended. However, there have been few efforts on mining commercial intents from social media contents. In this paper, we investigate the use of social media data to identify consumption intentions for individuals. We develop a Consumption Intention Mining Model (CIMM) based on convolutional neural network (CNN), for identifying whether the user has a consumption intention. The task is domain-dependent, and learning CNN requires a large number of annotated instances, which can be available only in some domains. Hence, we investigate the possibility of transferring the CNN mid-level sentence representation learned from one domain to another by adding an adaptation layer. To demonstrate the effectiveness of CIMM, we conduct experiments on two domains. Our results show that CIMM offers a powerful paradigm for effectively identifying users’ consumption intention based on their social media data. Moreover, our results also confirm that the CNN learned in one domain can be effectively transferred to another domain. This suggests that a great potential for our model to significantly increase effectiveness of product recommendations and targeted advertising.

Learning mutually-grounded vision-language knowledge is a foundational task for cognitive systems and human-level artificial intelligence. Most of knowledge-learning techniques are focused on single modal representations in a static environment with a fixed set of data. Here, we explore an ecologically more-plausible setting by using a stream of cartoon videos to build vision-language concept hierarchies continuously. This approach is motivated by the literature on cognitive development in early childhood. We present the model of deep concept hierarchy (DCH) that enables the progressive abstraction of concept knowledge in multiple levels. We develop a stochastic method for graph construction, i.e. a graph Monte Carlo algorithm, to search efficiently the huge compositional space of the vision-language concepts. The concept hierarchies are built incrementally and can handle concept drift, allowing for being deployed in lifelong learning environments. Using a series of approximately 200 episodes of educational cartoon videos we demonstrate the emergence and evolution of the concept hierarchies as the video stories unfold. We also present the application of the deep concept hierarchies for context-dependent translation between vision and language, i.e. the transcription of a visual scene into text and the generation of visual imagery from text.

Error backpropagation is an extremely effective algorithm for assigning credit in artificial neural networks. However, weight updates under Backprop depend on lengthy recursive computations and require separate output and error messages — features not shared by biological neurons, that are perhaps unnecessary. In this paper, we revisit Backprop and the credit assignment problem. We first decompose Backprop into a collection of interacting learning algorithms; provide regret bounds on the performance of these sub-algorithms; and factorize Backprop's error signals. Using these results, we derive a new credit assignment algorithm for nonparametric regression, Kickback, that is significantly simpler than Backprop. Finally, we provide a sufficient condition for Kickback to follow error gradients, and show that Kickback matches Backprop's performance on real-world regression benchmarks.