Deep Learning Approaches for Understanding Simple Speech Commands
Automatic classification of sound commands is becoming increasingly important, especially for embedded and mobile devices. Many of these devices contain both microphones and cameras. The manufacturers that develop and produce them would like to use the same methodology for sound and image classification tasks. It’s possible to achieve by representing sound commands as images, and then use convolutional neural networks when classifying images as well as sounds. In this research, we tried several approaches to the problem of sound classification that we applied in TensorFlow Speech Recognition Challenge organized by Google Brain team on the Kaggle platform. Here we show different representations of sounds (Wave frames, Spectrograms, Mel-Spectrograms, MFCCs) and apply several 1D and 2D convolutional neural networks to get the best performance. As a novelty of our work, we developed and trained from scratch two 1d network architectures that are topologically similar to 2d VGG and ResNet network types. These networks show similar performance with 2d networks when sound signal is represented by using melgrams. Our experiments reveal that we found appropriate sound representation and corresponding convolutional neural networks. As a result, we achieved good classification accuracy (91.8%) that allowed us to finish the challenge on 8-th place among 1315 teams.
Brain-computer interfaces find application in a number of different areas and have the potential to be used for research as well as for practical purposes. The clinical use of BCI includes current studies on neurorehabilitation ([Frolov et al., 2013; Ang et al., 2010]), and there is the prospect of using BCI to restore movement and communication capabilities, providing alternative effective pathways to those that may be lost due to injury or illness. The processing of electrophysiological data requires analysis of high-dimensional, nonstationary, noisy signals reflecting complex underlying processes and structures. We have shown that for non-invasive neuroimaging methods such as EEG the potential improvement lies in the field of machine learning and involves designing data analysis algorithms that can model physiological and psychoemotional variability of the user. The development of such algorithms can be conducted in different ways, including the classical Bayesian paradigm as well as modern deep learning architectures. The interpretation of nonlinear decision rules implemented by multilayer structures would enable automatic and objective knowledge extraction from the neurocognitive experiments data. Despite the advantages of non-invasive neuroimaging methods, a radical increase in the bandwidth of the BCI communication channel and the use of this technology for the prosthesis control is possible only through invasive technologies. Electrocorticogram (ECoG) is the least invasive of such technologies, and in the final part of this work we demonstrate the possibility of using ECoG to decode the kinematic characteristics of the finger movement.
We propose a novel multi-texture synthesis model based on generative adversarial networks (GANs) with a user-controllable mechanism. The user control ability allows to explicitly specify the texture which should be generated by the model. This property follows from using an encoder part which learns a latent representation for each texture from the dataset. To ensure a dataset coverage, we use an adversarial loss function that penalizes for incorrect reproductions of a given texture. In experiments, we show that our model can learn descriptive texture manifolds for large datasets and from raw data such as a collection of high-resolution photos. We show our unsupervised learning pipeline may help segmentation models. Moreover, we apply our method to produce 3D textures and show that it outperforms existing baselines.
Autonomous taxies are in high demand for smart city scenario. Such taxies have a well specified path to travel. Therefore, these vehicles only required two important parameters. One is detection parameter and other is control parameter. Further, detection parameters require turn detection and obstacle detection. The control parameters contain steering control and speed control. In this paper a novel autonomous taxi model has been proposed for smart city scenario. Deep learning has been used to model the human driver capabilities for the autonomous taxi. A hierarchical Deep Neural Network (DNN) architecture has been utilized to train various driving aspects. In first level, the proposed DNN architecture classifies the straight and turning of road. A parallel DNN is used to detect obstacle at level one. In second level, the DNN discriminates the turning i.e. left or right for steering and speed controls. Two multi layered DNNs have been used on Nvidia Tesla K 40 GPU based system with Core i-7 processor. The mean squared error (MSE) for the detection parameters viz. speed and steering angle were 0.018 and 0.0248 percent, respectively, with 15 milli seconds of realtime response delay.
A new public dataset of traffic sign images is presented. The dataset is intended for training and testing the algorithms of traffic sign recognition. We describe the dataset structure and guidelines for working with the dataset, comparing it with the previously published traffic sign datasets. The evaluation of modern detection and classification algorithms conducted using the proposed dataset has shown that existing methods of recognition of a wide class of traffic signs do not achieve the accuracy and completeness required for a number of applications.
It has been shown that the activations invoked by an image within the top layers of a large convolutional neural network provide a high-level descriptor of the visual content of the image. In this paper, we investigate the use of such descriptors (neural codes) within the image retrieval application. In the experiments with several standard retrieval benchmarks, we establish that neural codes perform competitively even when the convolutional neural network has been trained for an unrelated classification task (e.g. Image-Net). We also evaluate the improvement in the retrieval performance of neural codes, when the network is retrained on a dataset of images that are similar to images encountered at test time. We further evaluate the performance of the compressed neural codes and show that a simple PCA compression provides very good short codes that give state-of-the-art accuracy on a number of datasets. In general, neural codes turn out to be much more resilient to such compression in comparison other state-of-the-art descriptors. Finally, we show that discriminative dimensionality reduction trained on a dataset of pairs of matched photographs improves the performance of PCA-compressed neural codes even further. Overall, our quantitative experiments demonstrate the promise of neural codes as visual descriptors for image retrieval.
The performance of machine learning methods is heavily dependent on the choice of data representation (or features) on which they are applied. The rapidly developing field of representation learning is concerned with questions surrounding how we can best learn meaningful and useful representations of data. We take a broad view of the field and include topics such as deep learning and feature learning, metric learning, compositional modeling, structured prediction, reinforcement learning, and issues regarding large-scale learning and non-convex optimization. The range of domains to which these techniques apply is also very broad, from vision to speech recognition, text understanding, gaming, music, etc.
Objective: Brain-computer interfaces (BCIs) decode information from neural activity and send it to external devices. The use of Deep Learning approaches for decoding allows for automatic feature engineering within the specific decoding task. Physiologically plausible interpretation of the network parameters ensures the robustness of the learned decision rules and opens the exciting opportunity for automatic knowledge discovery. Approach: We describe a compact convolutional network-based architecture for adaptive decoding of electrocorticographic (ECoG) data into finger kinematics. We also propose a novel theoretically justified approach to interpreting the spatial and temporal weights in the architectures that combine adaptation in both space and time. The obtained spatial and frequency patterns characterizing the neuronal populations pivotal to the specific decoding task can then be interpreted by fitting appropriate spatial and dynamical models. Main results: We first tested our solution using realistic Monte-Carlo simulations. Then, when applied to the ECoG data from Berlin BCI competition IV dataset, our architecture performed comparably to the competition winners without requiring explicit feature engineering. Using the proposed approach to the network weights interpretation we could unravel the spatial and the spectral patterns of the neuronal processes underlying the successful decoding of finger kinematics from an ECoG dataset. Finally we have also applied the entire pipeline to the analysis of a 32-channel EEG motor-imagery dataset and observed physiologically plausible patterns specific to the task. Significance: We described a compact and interpretable CNN architecture derived from the basic principles and encompassing the knowledge in the field of neural electrophysiology. For the first time in the context of such multibranch architectures with factorized spatial and temporal processing we presented theoretically justified weights interpretation rules. We verified our recipes using simulations and real data and demonstrated that the proposed solution offers a good decoder and a tool for investigating motor control neural mechanisms.