Multi-Attentive Detection of the Spider Monkey Whinny in the (Actual) Wild

Abstract

We study deep bioacoustic event detection through multi-head attention based pooling, exemplified by wildlife monitoring. In the multiple instance learning framework, a core deep neural network learns a projection of the input acoustic signal into a sequence of embeddings, each representing a segment of the input. Sequence pooling is then required to aggregate the information present in the sequence such that we have a single clip-wise representation. We propose an improvement based on Squeeze-and-Excitation mechanisms upon a recently proposed audio tagging ResNet, and show that it performs significantly better than the baseline, as well as a collection of other recent audio models. We then further enhance our model, by performing an extensive comparative study of recent sequence pooling mechanisms, and achieve our best result using multi-head selfattention followed by concatenation of the head-specific pooled embeddings -better than prediction pooling methods, as well as compared to other recent sequence pooling tricks. We perform these experiments on a novel dataset of spider monkey whinny calls we introduce here, recorded in a rainforest in the South-Pacific coast of Costa Rica, with a promising outlook pertaining to minimally invasive wildlife monitoring.

Automated acoustic detection of Geoffroy's spider monkey highlights tipping points of human disturbance

“…(figure 2) for example spectrograms). Secondly, compared to the gibbon-focused studies in [33] and [22] that in such cases we should typically use smaller and shallower model architectures in order to avoid overfitting on the limited training set, Rizos et al [34] instead showed that it was the deeper and more complex models that achieved the highest performance (compared to models like the ones used in [22,33]), including the one we use in this study.…”

“…The classifier used in this study for whinny detection was first proposed in [34] as an improvement upon a deep, convolution-based, neural network architecture for acoustic event detection [61]. This improvement was achieved via the addition of attention-like mechanisms [62] that learn how to apply importance weights to the also learnt features as described in detail in [34]. Specifically, the model uses a squeeze-and-excitation mechanism [62] after each convolutional layer to reweigh the outputs of the convolutional filters, as well as a multiple-head attention mechanism [63] for pooling the sequential audio representation into a single, fixed vector representation.…”

“…Secondly, compared to the gibbon-focused studies in [33] and [22], we have fewer positive examples for training, which impacts the performance of DL techniques [31]. That being said, even though studies like Dufourq et al [22] mention that in such cases we should typically use smaller and shallower model architectures in order to avoid overfitting on the limited training set, Rizos et al [34] instead showed that it was the deeper and more complex models that achieved the highest performance (compared to models like the ones used in [22,33]), including the one we use in this study.

Figure 2Land use model results.

…”

“…Deep learning (DL) is a machine learning (ML) discipline that has been revolutionary in its success in modelling high-complexity data, including in the domain of computational bioacoustics [31], in which there has been an increase in predictive performance compared to more traditional ML approaches. We opted to use DL for its capability in modelling PAM data, inspired by previous studies, such as the discrimination among calls of different primates recorded in a wildlife sanctuary [32], and the detection of Bornean [33] and Hainan [22] gibbon calls, as well as calls by Geoffroy's spider monkey ( Ateles geoffroyi ) [34] using PAM data. Here, we use the deep convolutional model proposed in [34] for acoustic call detection.…”

“…We opted to use DL for its capability in modelling PAM data, inspired by previous studies, such as the discrimination among calls of different primates recorded in a wildlife sanctuary [32], and the detection of Bornean [33] and Hainan [22] gibbon calls, as well as calls by Geoffroy's spider monkey ( Ateles geoffroyi ) [34] using PAM data. Here, we use the deep convolutional model proposed in [34] for acoustic call detection.…”