A novel approach for automatic acoustic novelty detection using a denoising autoencoder with bidirectional LSTM neural networks

Marchi, Erik; Vesperini, Fabio; Eyben, Florian; Squartini, Stefano; Schuller, Björn

doi:10.1109/icassp.2015.7178320

Cited by 184 publications

(138 citation statements)

References 22 publications

Supporting

Mentioning

138

Contrasting

Order By: Relevance

“…NOVELTY DETECTION To detect novelty, we adopt an autoencoder-based approach proposed by Japkowicz et al [15], and focus on camera images as the domain of interest, although this general approach is also applicable to a wide variety of complex domains, including acoustic signals [22], network server anomalies [33], data mining [14], document classification [21] and others. Classification of novel data amounts to a one-class classification problem since we assume that we only have access to the set of images that have been observed and collected by our system, and no examples of any other "unfamiliar" images.…”

Section: Prior Estimate Of Collision Probabilitymentioning

confidence: 99%

Safe Visual Navigation via Deep Learning and Novelty Detection

Richter

Roy

2017

Robotics: Science and Systems XIII

174

View full text Add to dashboard Cite

Abstract-Robots that use learned perceptual models in the real world must be able to safely handle cases where they are forced to make decisions in scenarios that are unlike any of their training examples. However, state-of-the-art deep learning methods are known to produce erratic or unsafe predictions when faced with novel inputs. Furthermore, recent ensemble, bootstrap and dropout methods for quantifying neural network uncertainty may not efficiently provide accurate uncertainty estimates when queried with inputs that are very different from their training data. Rather than unconditionally trusting the predictions of a neural network for unpredictable real-world data, we use an autoencoder to recognize when a query is novel, and revert to a safe prior behavior. With this capability, we can deploy an autonomous deep learning system in arbitrary environments, without concern for whether it has received the appropriate training. We demonstrate our method with a vision-guided robot that can leverage its deep neural network to navigate 50% faster than a safe baseline policy in familiar types of environments, while reverting to the prior behavior in novel environments so that it can safely collect additional training data and continually improve. A video illustrating our approach is available at: http://groups.csail.mit.edu/rrg/videos/safe visual navigation.

show abstract

Section: Prior Estimate Of Collision Probabilitymentioning

confidence: 99%

Safe Visual Navigation via Deep Learning and Novelty Detection

Richter

Roy

2017

Robotics: Science and Systems XIII

174

View full text Add to dashboard Cite

show abstract

“…The final decision of this multi-classifier ensemble learning is given by the maximum score based on Eqs. (15) and (16) …”

Section: Two-stage Ensemble Learning Frameworkmentioning

confidence: 99%

“…Notably, this DNN algorithm can effectively improve feature representation of original time-frequency features and then promote the classification performance [8]. Other DL algorithms, such as deep belief network (DBN) [13], convolutional neural networks (CNN) [14,15], and auto-encoder (AE) [16], have also been effectively used for SEC. However, it is still time-costing to train a deep network model by these DL algorithms for a large-scale sound dataset.…”

Section: Introductionmentioning

confidence: 99%

Robust sound event classification with bilinear multi-column ELM-AE and two-stage ensemble learning

Zhang

Yin

Zhang

et al. 2017

J AUDIO SPEECH MUSIC PROC.

View full text Add to dashboard Cite

The automatic sound event classification (SEC) has attracted a growing attention in recent years. Feature extraction is a critical factor in SEC system, and the deep neural network (DNN) algorithms have achieved the state-of-the-art performance for SEC. The extreme learning machine-based auto-encoder (ELM-AE) is a new deep learning algorithm, which has both an excellent representation performance and very fast training procedure. However, ELM-AE suffers from the problem of unstability. In this work, a bilinear multi-column ELM-AE (B-MC-ELM-AE) algorithm is proposed to improve the robustness, stability, and feature representation of the original ELM-AE, which is then applied to learn feature representation of sound signals. Moreover, a B-MC-ELM-AE and two-stage ensemble learning (TsEL)-based feature learning and classification framework is then developed to perform the robust and effective SEC. The experimental results on the Real World Computing Partnership Sound Scene Database show that the proposed SEC framework outperforms the state-of-the-art DNN algorithm.

show abstract

“…That is, we face this sort of problem when it is difficult to collect abnormal sounds to be detected. To detect abnormal sounds, some methods use a criterion for how different input sounds are from ''normal'' sounds [22][23][24].…”

Section: Abnormal Sound Detectionmentioning

confidence: 99%