paulecoffet / birdsonglearningmodel

Model of zebra finch song learning with different learning strategies between day and night

Zebra finch song learning model

The Zebra Finch Bird Song Learning Model project aims to model the loss in performance of zebra finches after sleep which correlates with an increase in song reproduction at the end of their development. We believe this is due because the zebra finch restructures its knowledge during its sleep.

This project is done during my final internship for the Cogmaster in the AMAC team at the Institut des Systèmes Intelligents et de Robotique in Paris. This project is part of the DREAM project.

Architecture

The architecture of this model is a two part algorithm: One is a day part, and the second one is a night part. These parts do different things. The day part is an optimisation algorithm, whereas the night part is a restructuration algorithm.

Several model of learning for day and for night have been implemented.

• The list of day learning models
• The list of night learning models

Notebook

You can follow my research notebook to see the progression of my work.

birdsonglearningmodel's Issues

Entropy is too low

In the current generated song, the Wiener Entropy is way too low. This seems to be due to too slow oscillation in the parameters. New priors exploration in the branch entropy_fit will investigate this issue.

The hypothesis:

• Allow faster oscillations and put a prior of one fast oscillation for amplitude will increase wiener entropy.

Expected results:

• More complex sounds with more entropy, without deterioring the amplitude measure.

Learning during day sometimes lead to less good performances

Even if optimisation during the day is suppose to only increase the song mimicry performance, it sometimes has a huge negative impact. Optimizing locally seems to sometimes yield overall decrease in performance.

Training a gesture in context of its surrounding gesture was suppose to minimise this effect. It worked but the issue is still present.

Pitch is too high

The pitch found by the current algorithm is way to high. The YIN pitch is also too unstable to be fit perfectly.

The branch pitch_prio will try to improve this pitch fit, by changing both the priority measures and the prior. This branch is only for a better fit and is not bound to any hypothesis.