Abstract: Choosing good features is an essential part of machine learning. Recent techniques aim to automate this process. For instance, feature learning intends to learn the transformation of raw data into a useful representation to machine learning tasks. In automatic audio classification tasks, this is interesting since the audio, usually complex information, needs to be transformed into a computationally convenient input to process. Another technique tries to generate features by searching a feature space. Genetic algorithms, for instance, have being used to generate audio features by combining or modifying them. We find this approach particularly interesting and, despite the undeniable advances of feature learning approaches, we wanted to take a step forward in the use of genetic algorithms to find audio features, combining them with more conventional methods, like PCA, and inserting search control mechanisms, such as constraints over a confusion matrix. This work presents the results obtained on particular audio classification problems.
Abstract: Speaker recognition is performed in high Additive White Gaussian Noise (AWGN) environments using principals of Computational Auditory Scene Analysis (CASA). CASA methods often classify sounds from images in the time-frequency (T-F) plane using spectrograms or cochleargrams as the image. In this paper atomic decomposition implemented by matching pursuit performs a transform from time series speech signals to the T-F plane. The atomic decomposition creates a sparsely populated T-F vector in “weight space” where each populated T-F position contains an amplitude weight. The weight space vector along with the atomic dictionary represents a denoised, compressed version of the original signal. The arraignment or of the atomic indices in the T-F vector are used for classification. Unsupervised feature learning implemented by a sparse autoencoder learns a single dictionary of basis features from a collection of envelope samples from all speakers. The approach is demonstrated using pairs of speakers from the TIMIT data set. Pairs of speakers are selected randomly from a single district. Each speak has 10 sentences. Two are used for training and 8 for testing. Atomic index probabilities are created for each training sentence and also for each test sentence. Classification is performed by finding the lowest Euclidean distance between then probabilities from the training sentences and the test sentences. Training is done at a 30dB Signal-to-Noise Ratio (SNR). Testing is performed at SNR’s of 0 dB, 5 dB, 10 dB and 30dB. The algorithm has a baseline classification accuracy of ~93% averaged over 10 pairs of speakers from the TIMIT data set. The baseline accuracy is attributable to short sequences of training and test data as well as the overall simplicity of the classification algorithm. The accuracy is not affected by AWGN and produces ~93% accuracy at 0dB SNR.
Abstract: This paper describes a cycle accurate simulation results of weight values learned by an auto-encoder behavior model in terms of pre-route simulation. Given the results we visualized the first layer representations with natural images. Many common deep learning threads have focused on learning high-level abstraction of unlabeled raw data by unsupervised feature learning. However, in the process of handling such a huge amount of data, the learning method’s computation complexity and time limited advanced research. These limitations came from the fact these algorithms were computed by using only single core CPUs. For this reason, parallel-based hardware, FPGAs, was seen as a possible solution to overcome these limitations. We adopted and simulated the ready-made auto-encoder to design a behavior model in VerilogHDL before designing hardware. With the auto-encoder behavior model pre-route simulation, we obtained the cycle accurate results of the parameter of each hidden layer by using MODELSIM. The cycle accurate results are very important factor in designing a parallel-based digital hardware. Finally this paper shows an appropriate operation of behavior model based pre-route simulation. Moreover, we visualized learning latent representations of the first hidden layer with Kyoto natural image dataset.
Abstract: In this paper we propose a novel approach for ascertaining human identity based on fusion of profile face and gait biometric cues The identification approach based on feature learning in PCA-LDA subspace, and classification using multivariate Bayesian classifiers allows significant improvement in recognition accuracy for low resolution surveillance video scenarios. The experimental evaluation of the proposed identification scheme on a publicly available database [2] showed that the fusion of face and gait cues in joint PCA-LDA space turns out to be a powerful method for capturing the inherent multimodality in walking gait patterns, and at the same time discriminating the person identity..