Automatic Acoustic Mosquito Tagging with Bayesian Neural Networks

Abstract

Deep learning models are now widely used in decision-making applications. These models must be robust to noise and carefully map to the underlying uncertainty in the data. Standard deterministic neural networks are well known to be poor at providing reliable estimates of uncertainty and often lack the robustness that is required for real-world deployment. In this paper, we work with an application that requires accurate uncertainty estimates in addition to good predictive performance. In particular, we consider the task of detecting a mosquito from its acoustic signature. We use Bayesian neural networks (BNNs) to infer predictive distributions over outputs and incorporate this uncertainty as part of an automatic labelling process. We demonstrate the utility of BNNs by performing the first fully automated data collection procedure to identify acoustic mosquito data on over 1,500 hours of unlabelled field data collected with low-cost smartphones in Tanzania. We use uncertainty metrics such as predictive entropy and mutual information to help with the labelling process. We show how to bridge the gap between theory and practice by describing our pipeline from data preprocessing to model output visualisation. Additionally, we supply all of our data and code. The successful autonomous detection of mosquitoes allows us to perform analysis which is critical to the project goals of tackling mosquito-borne diseases such as malaria and dengue fever.

Propagating variational model uncertainty for bioacoustic call label smoothing

“… 7 For a recent example, residual networks (hence ResNets, 8 i.e., deep convolutional networks with residual connections every few layers for facilitating backwards propagation of the error signal for training) were shown to outperform the competition in a study on AED. 9 A ResNet similar to the winning method from the aforementioned study was also shown to be the best performer specifically for bioacoustic call detection in an extensive comparative study 10 against a non-residual deep convolutional network, 9 shallower networks of around two or three (1D or 2D) convolutional layers commonly used for AED, 11 , 12 , 13 , 14 as well as a combination of convolutional and recurrent (i.e., designed for sequential data) layers previously used for the bioacoustic detection of Bornean gibbon calls. 15 The success of the winning model of Rizos et al.…”

“…MC-based approaches comprise Bayes by backprop 45 and MC dropout, 46 and have been applied on a wide range of data domains, including audio. 13 , 47 …”

From Theoretical Models to Practical Deployment: A Perspective and Case Study of Opportunities and Challenges in AI-driven Cardiac Auscultation Research for Low-Income Settings

“…It has been shown that BNNs can reduce overfitting, which is especially beneficial for small datasets like the one examined in this Challenge [70]. However, since this approach does not consistently outperform the deterministic counterparts, as discussed by [71], our aim was to scrutinise the specific impact of this approach in our study. Constructing a complete BNN requires modelling the prior distribution over all model parameters, a task that can be computationally demanding [72].…”

“…However, it has been shown that including them during training and inference is a viable approximation to a complete BNN [70,72]. Inspired by [72], we added dropout layers to various segments of the ResNet50 architecture, particularly to the BasicBlock() and Bottleneck() modules, as per the ResNet implementation from [71]. This can be interpreted as a Monte Carlo approximation to BNNs.…”

The ACM Multimedia 2022 Computational Paralinguistics Challenge

“…B.4]. Its structure is based on prior models that have been successful in assisting domain experts for mosquito tagging [14]. As features, the baseline uses 128 log-mel spectrogram coefficients with a time window of 30 feature frames and a stride of 5 frames for training.…”