AI, bat recognition and beyond: Bioacoustic classification and detection using a biologically inspired mathematical model – a Hopfield network

Given the consequences of biodiversity loss, vast amounts of data collected by acoustic recorders, limitations of commercial classifiers and distrust of ‘black box’ methods, there is an urgent need for new accurate, efficient and transparent automated analysis tools. Much of the current research utilising artificial intelligence focuses on convolutional neural networks (CNNs) and image classification of spectrograms. While results are promising, they vary widely, are computationally expensive, and too battery-heavy to be incorporated into edge processing devices analysing live data. We present a novel application of a Hopfield Neural Network3 (HNN) - a biologically inspired associative memory model. Our network effectively uses one-shot learning - via the Fast Fourier Transform of the raw signal - avoiding the costly CNN conversion to images and backpropagation process. The software developed, stores sample sound files, initially single echolocation pulses from cryptic bat species (Pipistrellus pipistrellus, Pipistrellus pygmaeus). Then, for test sound files, the HNN converges and classifies each as one of the two stored memories, silence, or an unrecognised sound - a novel use of HNN spurious minima. A large public split dataset4 inspired the approach and was used in testing. In 7 seconds, the sample sounds were stored and a folder containing 10,384 sound files processed. Files were automatically moved to new folders, labelled as one of the four categories. Initial accuracy of the whole model was 82%, while further analysis discovered inaccuracies in the initial dataset labelling, which if corrected would boost accuracy further. Additionally, the software was successful in identifying the two species in raw data collected on an AudioMoth and Song Meter Mini. The work proved successful as a highly efficient proof of concept, able to process and recognise sounds recorded on a variety of devices and is now being developed and broadened in scope in collaboration with research and non-profit partners.

Citation

Lomas, W. and Gascoyne, A. (2024) AI, bat recognition and beyond: Bioacoustic classification and detection using a biologically inspired mathematical model – a Hopfield network. Poster presented at the 5th World Ecoacoustics Congress, Universidad Autónoma de Madrid.

Type

Other

Language

en

Description

Poster presented at the 5th World Ecoacoustics Congress, 8th-12th July 2024, Universidad Autónoma de Madrid.

Collections

Faculty of Science and Engineering