Quadruplet Selection Methods For Deep Embedding Learning

Recognition of objects with subtle differences has been used in many practical applications, such as car model recognition and maritime vessel identification. For discrimination of the objects in fine-grained detail, we focus on deep embedding learning by using a multi-task learning framework, in which the hierarchical labels (coarse and fine labels) of the samples are utilized both for classification and a quadruplet-based loss function. In order to improve the recognition strength of the learned features, we present a novel feature selection method specifically designed for four training samples of a quadruplet. By experiments, it is observed that the selection of very hard negative samples with relatively easy positive ones from the same coarse and fine classes significantly increases some performance metrics in a fine-grained dataset when compared to selecting the quadruplet samples randomly. The feature embedding learned by the proposed method achieves favorable performance against its state-of-the-art counterparts.

Towards End-to-End Speech Recognition

Dimitri Palaz

Standard automatic speech recognition (ASR) systems follow a divide and conquer approach to convert speech into text. Alternately, the end goal is achieved by a combination of sub-tasks, namely, feature extraction, acoustic modeling and sequence decoding, which are optimized in an independent manner. More recently, in the machine learning community deep learning approaches have emerged which allow training of systems in an end-to-end manner. Such approaches have found success in the area of natural language processing and computer vision community, and have consequently peaked interest in the speech community. The present thesis builds on these recent advances to investigate approaches to develop speech recognition systems in end-to-end manner. In that respect, the thesis follows two main axes of research. The first axis of research focuses on joint learning of features and classifiers for acoustic modeling. The second axis of research focuses on joint modeling of the acoustic model and the decoder. Along the first axis of research, in the framework of hybrid hidden Markov model/artificial neural networks (HMM/ANN) based ASR, we develop a convolution neural networks (CNNs) based acoustic modeling approach that takes raw speech signal as input and estimates phone class conditional probabilities. Specifically, the CNN has several convolution layers (feature stage) followed by multilayer perceptron (classifier stage), which are jointly optimized during the training. Through ASR studies on multiple languages and extensive analysis of the approach, we show that the proposed approach, with minimal prior knowledge, is able to learn automatically the relevant features from the raw speech signal. This approach yields systems that have less number of parameters and achieves better performance, when compared to the conventional approach of cepstral feature extraction followed by classifier training. As the features are automatically learned from the signal, a natural question that arises is: are such systems robust to noise? Towards that we propose a robust CNN approach referred to as normalized CNN approach, which yields systems that are as robust as or better than the conventional ASR systems using cepstral features (with feature level normalizations). The second axis of research focuses on end-to-end sequence-to-sequence conversion. We first propose an end-to-end phoneme recognition system. In this system the relevant features, classifier and the decoder (based on conditional random fields) are jointly modeled during training. We demonstrate the viability of the approach on TIMIT phoneme recognition task. Building on top of that, we investigate a ``weakly supervised'' training that alleviates the necessity for frame level alignments. Finally, we extend the weakly supervised approach to propose a novel keyword spotting technique. In this technique, a CNN first process the input observation sequence to output word level scores, which are subsequently aggregated to detect or spot words. We demonstrate the potential of the approach through a comparative study on LibriSpeech with the standard approach of keyword word spotting based on lattice indexing using ASR system.

EPFL2016

Word Sequence Modeling using Deep Learning

Joël Yvon Roland Legrand

For a long time, natural language processing (NLP) has relied on generative models with task specific and manually engineered features. Recently, there has been a resurgence of interest for neural networks in the machine learning community, obtaining state-of-the-art results in various fields such as computer vision, speech processing and natural language processing. The central idea behind these approaches is to learn features and models simultaneously, in an end-to-end manner, and making as few assumptions as possible. In NLP, word embeddings, mapping words in a dictionary on a continuous low-dimensional vector space, have proven to be very efficient for a large variety of tasks while requiring almost no a-priori linguistic assumptions. In this thesis, we investigate continuous representations of segments in a sentence for the purpose of solving NLP tasks that involve complex sentence-level relationships. Our sequence modelling approach is based on neural networks and takes advantage of word embeddings. A first approach models words in context in the form of continuous vector representations which are used to solve the task of interest. With the use of a compositional procedure, allowing arbitrarily-sized segments to be compressed onto continuous vectors, the model is able to consider long-range word dependencies as well. We first validate our approach on the task of bilingual word alignment, consisting in finding word correspondences between a sentence in two different languages. Source and target words in context are modeled using convolutional neural networks, obtaining representations that are later used to compute alignment scores. An aggregation operation enables unsupervised training for this task. We show that our model outperforms a standard generative model. The model above is extended to tackle phrase prediction tasks where phrases rather than single words are to be tagged. These tasks have been typically cast as classic word tagging problems using special tagging schemes to identify the segments boundaries. The proposed neural model focuses on learning fixed-size representations of arbitrarily-sized chunks of words that are used to solve the tagging task. A compositional operation is introduced in this work for the purpose of computing these representations. We demonstrate the viability of the proposed representations by evaluating the approach on the multiwork expression tagging task. The remainder of this thesis addresses the task of syntactic constituency parsing which, as opposed to the above tasks, aims at producing a structured output, in the form of a tree, of an input sentence. Syntactic parsing is cast as multiple phrase prediction problems that are solved recursively in a greedy manner. An extension using recursive compositional vector representations, allowing for lexical infor- mation to be propagated from early stages, is explored as well. This approach is evaluated on a standard corpus obtaining performance comparable to generative models with much shorter computation time. Finally, morphological tags are included as additional features, using a similar composition procedure, to improve the parsing performance for morphologically rich languages. State-of-the-art results were obtained for these task and languages.

EPFL2016

Towards End-to-End Speech Recognition

Dimitri Palaz

EPFL2016

Word Sequence Modeling using Deep Learning

Joël Yvon Roland Legrand

EPFL2016