Supervised training of deep neural nets typically relies on minimizing cross-entropy. However, in many domains, we are interested in performing well on metrics specific to the application. In this paper we propose a direct loss minimization approach to train deep neural networks, which provably minimizes the application-specific loss function. This is often non-trivial, since these functions are neither smooth nor decomposable and thus are not amenable to optimization with standard gradient-based methods. We demonstrate the effectiveness of our approach in the context of maximizing average precision for ranking problems. Towards this goal, we develop a novel dynamic programming algorithm that can efficiently compute the weight updates. Our approach proves superior to a variety of baselines in the context of action classification and object detection, especially in the presence of label noise.
Black-box alpha (BB-$\alpha$) is a new approximate inference method based on the minimization of $\alpha$-divergences. BB-$\alpha$ scales to large datasets because it can be implemented using stochastic gradient descent. BB-$\alpha$ can be applied to complex probabilistic models with little effort since it only requires as input the likelihood function and its gradients. These gradients can be easily obtained using automatic differentiation. By changing the divergence parameter $\alpha$, the method is able to interpolate between variational Bayes (VB) ($\alpha \rightarrow 0$) and an algorithm similar to expectation propagation (EP) ($\alpha = 1$). Experiments on probit regression and neural network regression and classification problems show that BB-$\alpha$ with non-standard settings of $\alpha$, such as $\alpha = 0.5$, usually produces better predictions than with $\alpha \rightarrow 0$ (VB) or $\alpha = 1$ (EP).
We submitted two systems to the SemEval-2016 Task 12: Clinical TempEval challenge, participating in Phase 1, where we identified text spans of time and event expressions in clinical notes and Phase 2, where we predicted a relation between an event and its parent document creation time. For temporal entity extraction, we find that a joint inference-based approach using structured prediction outperforms a vanilla recurrent neural network that incorporates word embeddings trained on a variety of large clinical document sets. For document creation time relations, we find that a combination of date canonicalization and distant supervision rules for predicting relations on both events and time expressions improves classification, though gains are limited, likely due to the small scale of training data.
$k$ Nearest Neighbors ($k$NN) is one of the most widely used supervised learning algorithms to classify Gaussian distributed data, but it does not achieve good results when it is applied to nonlinear manifold distributed data, especially when a very limited amount of labeled samples are available. In this paper, we propose a new graph-based $k$NN algorithm which can effectively handle both Gaussian distributed data and nonlinear manifold distributed data. To achieve this goal, we first propose a constrained Tired Random Walk (TRW) by constructing an $R$-level nearest-neighbor strengthened tree over the graph, and then compute a TRW matrix for similarity measurement purposes. After this, the nearest neighbors are identified according to the TRW matrix and the class label of a query point is determined by the sum of all the TRW weights of its nearest neighbors. To deal with online situations, we also propose a new algorithm to handle sequential samples based a local neighborhood reconstruction. Comparison experiments are conducted on both synthetic data sets and real-world data sets to demonstrate the validity of the proposed new $k$NN algorithm and its improvements to other version of $k$NN algorithms. Given the widespread appearance of manifold structures in real-world problems and the popularity of the traditional $k$NN algorithm, the proposed manifold version $k$NN shows promising potential for classifying manifold-distributed data.
We tested 14 very different classification algorithms (random forest, gradient boosting machines, SVM - linear, polynomial, and RBF - 1-hidden-layer neural nets, extreme learning machines, k-nearest neighbors and a bagging of knn, naive Bayes, learning vector quantization, elastic net logistic regression, sparse linear discriminant analysis, and a boosting of linear classifiers) on 115 real life binary datasets. We followed the Demsar analysis and found that the three best classifiers (random forest, gbm and RBF SVM) are not significantly different from each other. We also discuss that a change of less then 0.0112 in the error rate should be considered as an irrelevant change, and used a Bayesian ANOVA analysis to conclude that with high probability the differences between these three classifiers is not of practical consequence. We also verified the execution time of "standard implementations" of these algorithms and concluded that RBF SVM is the fastest (significantly so) both in training time and in training plus testing time.
Document classification for text, images and other applicable entities has long been a focus of research in academia and also finds application in many industrial settings. Amidst a plethora of approaches to solve such problems, machine-learning techniques have found success in a variety of scenarios. In this paper we discuss the design of a machine learning-based semi-supervised job title classification system for the online job recruitment domain currently in production at CareerBuilder.com and propose enhancements to it. The system leverages a varied collection of classification as well clustering algorithms. These algorithms are encompassed in an architecture that facilitates leveraging existing off-the-shelf machine learning tools and techniques while keeping into consideration the challenges of constructing a scalable classification system for a large taxonomy of categories. As a continuously evolving system that is still under development we first discuss the existing semi-supervised classification system which is composed of both clustering and classification components in a proximity-based classifier setup and results of which are already used across numerous products at CareerBuilder. We then elucidate our long-term goals for job title classification and propose enhancements to the existing system in the form of a two-stage coarse and fine level classifier augmentation to construct a cascade of hierarchical vertical classifiers. Preliminary results are presented using experimental evaluation on real world industrial data.
Quantification is the machine learning task of estimating test-data class proportions that are not necessarily similar to those in training. Apart from its intrinsic value as an aggregate statistic, quantification output can also be used to optimize classifier probabilities, thereby increasing classification accuracy. We unify major quantification approaches under a constrained multi-variate regression framework, and use mathematical programming to estimate class proportions for different loss functions. With this modeling approach, we extend existing binary-only quantification approaches to multi-class settings as well. We empirically verify our unified framework by experimenting with several multi-class datasets including the Stanford Sentiment Treebank and CIFAR-10.
Previous approaches to model and analyze facial expression analysis use three different techniques: facial action units, geometric features and graph based modelling. However, previous approaches have treated these technique separately. There is an interrelationship between these techniques. The facial expression analysis is significantly improved by utilizing these mappings between major geometric features involved in facial expressions and the subset of facial action units whose presence or absence are unique to a facial expression. This paper combines dimension reduction techniques and image classification with search space pruning achieved by this unique subset of facial action units to significantly prune the search space. The performance results on the publicly facial expression database shows an improvement in performance by 70% over time while maintaining the emotion recognition correctness.
This paper describes a Hierarchical Composition Recurrent Network (HCRN) consisting of a 3-level hierarchy of compositional models: character, word and sentence. This model is designed to overcome two problems of representing a sentence on the basis of a constituent word sequence. The first is a data-sparsity problem in word embedding, and the other is a no usage of inter-sentence dependency. In the HCRN, word representations are built from characters, thus resolving the data-sparsity problem, and inter-sentence dependency is embedded into sentence representation at the level of sentence composition. We adopt a hierarchy-wise learning scheme in order to alleviate the optimization difficulties of learning deep hierarchical recurrent network in end-to-end fashion. The HCRN was quantitatively and qualitatively evaluated on a dialogue act classification task. Especially, sentence representations with an inter-sentence dependency are able to capture both implicit and explicit semantics of sentence, significantly improving performance. In the end, the HCRN achieved state-of-the-art performance with a test error rate of 22.7% for dialogue act classification on the SWBD-DAMSL database.
In this paper, we explore degrees of freedom in deep sigmoidal neural networks. We show that the degrees of freedom in these models is related to the expected optimism, which is the expected difference between test error and training error. We provide an efficient Monte-Carlo method to estimate the degrees of freedom for multi-class classification methods. We show degrees of freedom are lower than the parameter count in a simple XOR network. We extend these results to neural nets trained on synthetic and real data, and investigate impact of network's architecture and different regularization choices. The degrees of freedom in deep networks are dramatically smaller than the number of parameters, in some real datasets several orders of magnitude. Further, we observe that for fixed number of parameters, deeper networks have less degrees of freedom exhibiting a regularization-by-depth.
Deep Convolutional Neural Networks (CNN) enforces supervised information only at the output layer, and hidden layers are trained by back propagating the prediction error from the output layer without explicit supervision. We propose a supervised feature learning approach, Label Consistent Neural Network, which enforces direct supervision in late hidden layers. We associate each neuron in a hidden layer with a particular class label and encourage it to be activated for input signals from the same class. More specifically, we introduce a label consistency regularization called "discriminative representation error" loss for late hidden layers and combine it with classification error loss to build our overall objective function. This label consistency constraint alleviates the common problem of gradient vanishing and tends to faster convergence; it also makes the features derived from late hidden layers discriminative enough for classification even using a simple $k$-NN classifier, since input signals from the same class will have very similar representations. Experimental results demonstrate that our approach achieves state-of-the-art performances on several public benchmarks for action and object category recognition.
This paper connects a vector-based composition model to a formal semantics, the Dependency-based Compositional Semantics (DCS). We show theoretical evidence that the vector compositions in our model conform to the logic of DCS. Experimentally, we show that vector-based composition brings a strong ability to calculate similar phrases as similar vectors, achieving near state-of-the-art on a wide range of phrase similarity tasks and relation classification; meanwhile, DCS can guide building vectors for structured queries that can be directly executed. We evaluate this utility on sentence completion task and report a new state-of-the-art.
We derive a relationship between network representation in energy-efficient neuromorphic architectures and block Toplitz convolutional matrices. Inspired by this connection, we develop deep convolutional networks using a family of structured convolutional matrices and achieve state-of-the-art trade-off between energy efficiency and classification accuracy for well-known image recognition tasks. We also put forward a novel method to train binary convolutional networks by utilising an existing connection between noisy-rectified linear units and binary activations.
We propose using five data-driven community detection approaches from social networks to partition the label space for the task of multi-label classification as an alternative to random partitioning into equal subsets as performed by RAkELd: modularity-maximizing fastgreedy and leading eigenvector, infomap, walktrap and label propagation algorithms. We construct a label co-occurence graph (both weighted an unweighted versions) based on training data and perform community detection to partition the label set. We include Binary Relevance and Label Powerset classification methods for comparison. We use gini-index based Decision Trees as the base classifier. We compare educated approaches to label space divisions against random baselines on 12 benchmark data sets over five evaluation measures. We show that in almost all cases seven educated guess approaches are more likely to outperform RAkELd than otherwise in all measures, but Hamming Loss. We show that fastgreedy and walktrap community detection methods on weighted label co-occurence graphs are 85-92% more likely to yield better F1 scores than random partitioning. Infomap on the unweighted label co-occurence graphs is on average 90% of the times better than random paritioning in terms of Subset Accuracy and 89% when it comes to Jaccard similarity. Weighted fastgreedy is better on average than RAkELd when it comes to Hamming Loss.
The massive availability of digital repositories of human thought opens radical novel way of studying the human mind. Natural language processing tools and computational models have evolved such that many mental conditions are predicted by analysing speech. Transcription of interviews and discourses are analyzed using syntactic, grammatical or sentiment analysis to infer the mental state. Here we set to investigate if classification of Bipolar and control subjects is possible. We develop the Emotion Intensity Index based on the Dictionary of Affect, and find that subjects categories are distinguishable. Using classical classification techniques we get more than 75\% of labeling performance. These results sumed to previous studies show that current automated speech analysis is capable of identifying altered mental states towards a quantitative psychiatry.
The classification procedure of streaming data usually requires various ad hoc methods or particular heuristic models. We explore a novel non-parametric and systematic approach to analysis of heterogeneous sequential data. We demonstrate an application of this method to classification of the delays in responding to the prompts, from subjects with bipolar disorder collected during a clinical trial, using both synthetic and real examples. We show how this method can provide a natural and systematic way to extract characteristic features from sequential data.
Inferring implicit discourse relations in natural language text is the most difficult subtask in discourse parsing. Surface features achieve good performance, but they are not readily applicable to other languages without semantic lexicons. Previous neural models require parses, surface features, or a small label set to work well. Here, we propose neural network models that are based on feedforward and long-short term memory architecture without any surface features. To our surprise, our best configured feedforward architecture outperforms LSTM-based model in most cases despite thorough tuning. Under various fine-grained label sets and a cross-linguistic setting, our feedforward models perform consistently better or at least just as well as systems that require hand-crafted surface features. Our models present the first neural Chinese discourse parser in the style of Chinese Discourse Treebank, showing that our results hold cross-linguistically.
Object classification according to their shape and size is of key importance in many scientific fields. This work focuses on the case where the size and shape of an object is characterized by a current}. A current is a mathematical object which has been proved relevant to the modeling of geometrical data, like submanifolds, through integration of vector fields along them. As a consequence of the choice of a vector-valued Reproducing Kernel Hilbert Space (RKHS) as a test space for integrating manifolds, it is possible to consider that shapes are embedded in this Hilbert Space. A vector-valued RKHS is a Hilbert space of vector fields; therefore, it is possible to compute a mean of shapes, or to calculate a distance between two manifolds. This embedding enables us to consider size-and-shape classification algorithms. These algorithms are applied to a 3D database obtained from an anthropometric survey of the Spanish child population with a potential application to online sales of children's wear.
We propose a novel classification model for weak signal data, building upon a recent model for Bayesian multi-view learning, Group Factor Analysis (GFA). Instead of assuming all data to come from a single GFA model, we allow latent clusters, each having a different GFA model and producing a different class distribution. We show that sharing information across the clusters, by sharing factors, increases the classification accuracy considerably; the shared factors essentially form a flexible noise model that explains away the part of data not related to classification. Motivation for the setting comes from single-trial functional brain imaging data, having a very low signal-to-noise ratio and a natural multi-view setting, with the different sensors, measurement modalities (EEG, MEG, fMRI) and possible auxiliary information as views. We demonstrate our model on a MEG dataset.
A complex nature of big data resources demands new methods for structuring especially for textual content. WordNet is a good knowledge source for comprehensive abstraction of natural language as its good implementations exist for many languages. Since WordNet embeds natural language in the form of a complex network, a transformation mechanism WordNet2Vec is proposed in the paper. It creates vectors for each word from WordNet. These vectors encapsulate general position - role of a given word towards all other words in the natural language. Any list or set of such vectors contains knowledge about the context of its component within the whole language. Such word representation can be easily applied to many analytic tasks like classification or clustering. The usefulness of the WordNet2Vec method was demonstrated in sentiment analysis, i.e. classification with transfer learning for the real Amazon opinion textual dataset.