OVERCOMING OVERFITTING IN MONKEY VOCALIZATION CLASSIFICATION: USING LSTM AND LOGISTIC REGRESSION

Abstrak

The problem of overfitting in a classification task involving animal vocalizations, namely squirrel monkeys, golden lion tamarins, and tailed macaques, is handled in this project. Acoustic features extracted for the audio data used in this research are MFCCs. The classification of subjects was done using the LSTM model. However, several architectures with LSTM also presented the problem of overfitting. To overcome this, a logistic regression model was used, which had a classification accuracy of 100%. These results indicate that for such a classification problem, logistic regression may be more appropriate than the complex architecture of LSTMs. Several LSTM architectures have been presented in this study to give an overall review of the observed challenges. Although the capability of LSTM in handling sequential data is very promising, sometimes simpler models might be preferred, as indicated by the results. This is a single-dataset work, and the findings may not generalize well to other domains. The work contributes much-needed insight into the choice of models for audio classification tasks and identifies the trade-off between model complexity and performance