Leonardo Nogueira Matos

Combining global and local classifiers with Bayesian network

This paper introduces a classification method based on feature space segmentation. Since the classification task is equivalent to a probability distribution estimation, a Bayesian network is used as an inference mechanism for dealing with the underling probability distribution function that, presumably, is complex and factored. The article presents a method for splitting the feature space into regions that are associated to local classifiers. After that, a Bayesian network is used for combining their outputs. Experimental results reveal that this is a suitable approach for speeding up the training phase for large databases as well as to ensure good recognition rates

Multi-kernel approach to Parallelization of EM Algorithm for GMM Training

Most machine learning algorithms need to handle large datasets. This feature often leads to limitations on processing time and memory. The Expectation-Maximization (EM) is one of such algorithms, which is used to train one of the most commonly used parametric statistical models, the Gaussian Mixture Models (GMM). All steps of the algorithm are potentially parallelizable once they iterate over the entire data set. In this work, we propose a parallel implementation of EM for training GMM using CUDA cores. Experimentation scenario consists of five different datasets and four metrics. Results show a speedup of 12.7 if compared to sequential version. With coalesced access to CUDA global memory and shared memory usage, we have achieved up to 99.4% of actual occupancy, regardless the number of Gaussians considered.

On the infinite clipping of handwritten signatures

The coarsest quantization of velocity signature signals carries most information.In average, velocity signal keeps less than 1% of the total signature energy.Infinite clipping of signatures leads to high performance without disclosing it. The infinite clipping-a two-level quantization of a given signal based on its polarity-is applied to velocity signals from online handwritten signatures (i.e., pen movement recorded as a time series). In terms of signature authenticity verification, a counter-intuitive result is observed, similar to those reported by Licklider and Pollack, in the 1940s, concerning speech intelligibility. As in their work, the strong signal distortions caused by the infinite clipping have little effect on the conveyed information. Motivated by these observed results, we further investigate a related point of view, according to which the infinite clipping is just the two-level quantization of the first detail output from the Haar wavelet transform. As a consequence, we study the relevance of each subsequent wavelet analysis output, as compared to the energy concentration inside each corresponding signal subspace. Experimentally, we conclude that velocity subspace retains less than 1% of the total signal energy, whereas it carries most of the information necessary to attain authenticity verification performance similar to those of state-of-the-art systems.

Evaluating a zoning mechanism and class-modular architecture for handwritten characters recognition

In this article we propose a feature extraction procedure based on directional histograms and investigate the application of a nonconventional neural network architecture, applied to the problem of handwritten character recognition. This approach is inspired on some characteristics of the human visual system, as it focus attention on high spatial frequencies and on the recognition of local features. Two architectures were tested and evaluated: a conventional MLP (Multiple Layer Perceptron) and a class-modular MLP. Experiments developed with the Letter database produced a recognition rate of 93.67% for the class-modular MLP. Other set of experiments utilized the IRONOFF database resulting in recognition rates of 89.21% and 80.75% for uppercase and lowercase characters respectively, also with the class-modular MLP.

A Traffic Light Recognition Device

Traffic lights detection and recognition research has grown every year. Time is coming when autonomous vehicle can navigate in urban roads and streets and intelligent systems aboard those cars would have to recognize traffic lights in real time. This article proposes a traffic light recognition (TLR) device prototype using a smartphone as camera and processing unit that can be used as a driver assistance. A TLR device has to be able to visualize the traffic scene from inside of a vehicle, generate stable images, and be protected from adverse conditions. To validate this layout prototype, a dataset was built and used to test an algorithm that uses an adaptive background suppression filter (AdaBSF) and Support Vector Machines (SVMs) to detect traffic lights. The application of AdaBSF and subsequent classification with SVM to the dataset achieved 100% precision rate and recall of 65%. Road testing shows that the TLR device prototype meets the requirements to be used as a driver assistance device.

A Systematic Method for Detecting Parallelized Software Bottlenecks and Suggesting Modifications: The Case of the Expectation Maximization Algorithm

Parallelized algorithms can distribute the workload on the available multi-core processors. Graphical Processing Units (GPU) began to be used in general purpose computing thanks to its ability to simultaneously perform thousands of operations in their parallel coprocessors. Unfortunately, providing parallelized versions of typical sequential routines is not a trivial task. Even with the advent of CUDA, the NVIDIA’s more intuitive solution for GPU programming, developers need to acquire a deep knowledge of GPU architecture and the rationale of the target algorithms to optimize resources usage and reduce processing time. This paper proposes a systematic method for analyzing parallelized algorithms and propose guidelines for CUDA code refactoring in such a way faster and more efficient software, regarding hardware resources consumption, could be constructed. One of such kind of software is Automatic Speech Recognition (ASR) systems. Mainstream approaches for ASR use the Expectation Maximization (EM) algorithm to train Gaussian Mixture Models (GMM) to provide an Acoustic Model for ASR. These training phase is usually extensive time-consuming and so it’s well suited for a parallelized solution approach. We show the feasibility of our method identifying important issues in a literatures parallelized implementation of EM and further refactoring suggestion to enhance memory occupancy and decrease processing time. The results show a processing speedup of the EM algorithm around 40x (minimum) and 61x (maximum) when compared to the control version. The method was also effective in the improvement of the values for all the concerned performance metrics for GPU-based solutions.

Speech Recognition in Noisy Environments with Convolutional Neural Networks

One of the biggest challenges in speech recognition today is its use on a daily basis, in which distortion and noise in the environment are present and hinder the recognition task. In the last thirty years, hundreds of methods for noise-robust recognition were proposed, each with its own advantages and disadvantages. In this paper, the use of convolutional neural networks (CNN) as acoustic models in automatic speech recognition systems (ASR) is proposed as an alternative to the classical recognition methods based on HMM without any noise-robust method applied. The experiment showed that the presented method reduces the equal error rate in word recognition tasks with additive noise.

Brazilian Portuguese speech-driven answering system

Call centers are being increasingly incorporated into companies and institutions. There are usually two different types of call centers: customer services with human attendants and numeric keypad-driven automatic service. Human attendants have high costs. The use of numeric keypads are not intuitive and increases systems rejection rate. In order to contribute to the reduction of these limitations, this paper proposes an automated answering system with speech recognition for Brazilian Portuguese. As a case study, the system handles phone calls of the Computer Science Department secretary at a federal university. Current version is able to provide automatic call transfer to departments lectures, voice messages recording and automatic e-mailing. Recognition evaluation has been done by means of four different metrics. The metric WIP pointed a speech recognition rate of ~91% for limited vocabulary.

Acoustic models comparison using HTK and the Spoltech corpus to Brazilian Portuguese

This paper shows a comparison between Hidden Markov Models (HMM) trained with 12 mel-cepstral coefficients plus extra(s) parameter(s) and two different HMM initialization ways. Thus, it compares the models, in order to detect the more robust parameter added to the mel-cepstral vector in an Automatic Speech Recognizer (ASR) system for the Brazilian Portuguese. To perform such experiments, it uses the HTK to train the HMMs. All the HMMs models used the same speech training base, which is the Spoltech corpus.