ENSC Parcours IA

Data Challenge - Détection de clics d'odontocètes

✍️ Authors

• Tristan Gonçalves
• Thomas Chimbault

🐬 Context

As part of the Artificial Intelligence specialization at the ENSC, we participated in a data challenge provided by the University of Toulon in the ChallengeData website.

This challenge specifically aims to detect the presence of odontoceti clicks in underwater audio recordings in the Caribbean sea. The model will be evaluated on the ChallengeData website.

📊 Data Description

The dataset is composed of 23,168 audio files in WAV format, each of duration 200ms. The clicks are labeled with a binary variable: 1 if the file contains a click, 0 otherwise.

🎯 Challenge

The objective of the challenge is to create a model that predicts the presence of odontoceti clicks in the test set with the highest accuracy.

📏 Evaluation

The submissions are evaluated on the ROC AUC (area under the curve) metric.

The results must be submitted as a CSV file with 950 lines. Each line corresponds to a file of the test set and contains the prediction for this file. The prediction in percentage should be indicated and must not be rounded to binary labels.

🔎 Our approach

We first used classical machine learning model, such as Linear Regression or Random Forest.

Then, we used the ReservoirPy library, created and maintained by the Inria of Bordeaux. This uses the reservoir computing theory, part of the RNN's domain.

We also used a Convolutional Neural Network to classify the audio files. We used the Librosa library to extract the audio features.

Results

Classical approaches

Method	Result
Logistic Regression	0.5981
Decision Tree	0.6124
Bagged Tree	0.6351
Random Forest	0.6460
XGBoost	0.6301

Reservoir computing

We didn't get the result we expected by using reservoir computing. In fact, the issue we got resided in the format of the results. As a matter of fact, ReservoirPy's results were not probabilities and could be above 1 or below 0. Therefore, we had to put a threshold and round these extreme values to either 1 or 0. This distorted our results and we obtained a score of only 0.48. However, this method has been implemented very late during the project, so we maybe have wrongfully used this library.

Convolution Networks

Conv1D

Final Score: 0.9566

Conv2D

We tried to use 2D convolutions because it is well-known that audio files can be represented with spectrograms.
However, results were really not convincing, in regard of the energetic consumption of our model training. Indeed, our project being related to the study of underwater life, we thought that having a very consuming and heavy model was totally inappropriate.

For information, our best result there was 0.86.

💻 Installation

First of all, you may clone this repository on your computer.

git clone https://github.com/tristangclvs/spe_ia_clics_odontocetes.git

Then, download the .dataset archive here and extract it in the main root of the cloned folder.

Creating a virtual environment

You may want to create a virtual environment for python. \

python -m venv NameOfYourEnv

Then select your environment:

⊞ Windows:

NameOfYourEnv/Scripts/activate

🍏 Mac:

source NameOfYourEnv/bin/activate

To run the code in this repository, you will need to install the necessary dependencies:

pip install -r requirements.txt

Repository Structure

The repository is structured as follows:

• CNN_topic: contains the files used to train the convolutional neural networks (1D & 2D).
• images: contains the images used in the README file and a notebook for plotting some charts.
• notebooks: contains the notebooks used for the challenge.