Deep learning approach to classification of acoustic signals using information features

The paper considers the problem of binary classification of acoustic signals of biological origin recorded real environment. Information characteristics such as entropy and statistical complexity are chosen as the characteristic description of objects. The solution methods are based on three neural network architectures modified by the authors (on the Inception core, on the Inception core with the Residual technology, on the Self-Attention structure with LSTM blocks). A dataset from the Kaggle competition for detecting acoustic signatures of whales was used, and a comparison was made between the models in terms of the quality of solving the problem under consideration on a standard set of metrics. The AUC ROC value of more than 90% was obtained, which indicates the successful solution of the problem of detecting a useful signal and indicates the possible applicability of information characteristics to similar tasks.