Using artificial bat sonar neural networks for complex pattern recognition: Recognizing faces and the speed of a moving target

Abstract

Two sets of studies examined the viability of using bat-like sonar input for artificial neural networks in complex pattern recognition tasks. In the first set of studies, a sonar neural network was required to perform two face recognition tasks. In the first task, the network was trained to recognize different faces regardless of facial expressions. Following training, the network was tested on its ability to generalize and correctly recognize faces using echoes of novel facial expressions that were not included in the training set. The neural network was able to recognize novel echoes of faces almost perfectly (above 96% accuracy) when it was required to recognize up to five faces. In the second face recognition task, a sonar neural network was trained to recognize the sex of 16 faces (eight males and eight females). After training, the network was able to correctly recognize novel echoes of those faces as 'male' or as 'female' faces with accuracy levels of 88%. However, the network was not able to recognize novel faces as 'male' or 'female' faces. In the second set of studies, a sonar neural network was required to learn to recognize the speed of a target that was moving towards the viewer. During training, the target was presented in a variety of orientations, and the network's performance was evaluated when the target was presented in novel orientations that were not included in the training set. The different orientations dramatically affected the amplitude and the frequency composition of the echoes. The neural network was able to learn and recognize the speed of a moving target, and to generalize to new orientations of the target. However, the network was not able to generalize to new speeds that were not included in the training set. The potential and limitations of using bat-like sonar as input for artifical neural networks are discussed.

Computational analyses in cognitive neuroscience: In defense of biological implausibility

“…This computational analysis assessed the availability of shape information in sonar sounds and has been successfully applied to a variety ofpattern recognition tasks (see Dror, Florer, Rios, & Zagaeski, 1996). The fact that such information is explicitly and easily available in one dimension of the sonar sounds and not in others does not prove that the biological sonar system utilizes the computationally efficient dimension.…”

Gait-based person and gender recognition using micro-doppler signatures

“…The most common applications are: detecting presence, counting, tracking and identifying individuals [1]- [4]. In addition a few studies on face recognition [5]- [7], one handed gesture recognition [8], fall detections [9], [10] or mode of transport classification [11] have also been reported. A number of features set the different systems apart, such as the transmission mode of the wave, the type of wave used for microdoppler (acoustic or electromagnetic) and finally the data analysis methods.…”

“…There are three modes of transmission used, with the two most common being the continuous wave (CW) sonar at 40kHz [4], [6], [8], [11] or CW RF at 10.5GHz [2], [3], and the Pulse Doppler RF [9], [10] with a carrier at 5.8GHz and pulse modulated at a lower frequency. The third mode uses chirps, sweeping frequencies from 25kHz to 100kHz [5], [7]. The most common data analysis methods start with a Fast Fourier Transform (FFT) or a Short-Time Fourier Transform (STFT) to process the data and then a classifier is used for pattern recognition.…”

“…The most common data analysis methods start with a Fast Fourier Transform (FFT) or a Short-Time Fourier Transform (STFT) to process the data and then a classifier is used for pattern recognition. A broad range of classifiers are used: linear [3], neural network [5], standard deviation+linear [11], Bayesian [4], [6], Gaussian Mixture Model (GMM, [8]), Maximum Correlation Coefficient (MCC, [7]), Support Vector Machine (SVM, [10]), and k-Nearest Neighbour (kNN, [10]). The proposed solution uses a CW sonar at 40kHz with bandpass sampling and a novel data analysis method which will be described below.…”

Classifying still faces with ultrasonic sensing

“…Dror et al [3] trained a neural network to recognize 5 faces using the time-frequency spectrogram of the echo as input. The network recognized the 5 faces with an accuracy of 96%.…”