Date de publication : Mar 13, 2017 3:19:25 PM
Click here for more informations and give your application
Title of the proposal research subject
Modeling the neural network responsible for song learning in birds
SONNET : SOng Neuronal NETwork
Scientific priorities
Theme: Modeling and simulations for health
Building theoretical models for the neural dynamics underlying brain function, including learning, is essential to (1) get a better understanding of general brain function, and to (2) explore new paths that are not accessible using purely experimental (neurobiological) methods. Modeling the dynamics of brain circuits such as the basal ganglia, which are well conserved along evolution and display homologous circuits in birds and mammals, will also shed light on the pathological mechanisms of neurodegenerative diseases (such as Parkinson’s diseases).
Scientific research context
Learning complex sensorimotor skills (such as driving) involves an ensemble of sub-cortical nuclei called the basal ganglia (BG), which make a loop network with the cortex and thalamus. The BG receive projections from almost all cortical areas, allowing stimulus-response pairing (e.g. bell sound and food) and, more generally, complex sensorimotor transformations (e.g. song or speech learning). The BG-thalamo-cortical circuit is very complex in mammals as it underlies a wide number of functions.
Songbirds learn their complex vocalizations by imitation of a tutor. They devote a large part of their brain to the perception and production of song, making it an ideal animal model to study the representation of sensorimotor learning. More particularly, songbirds have a specific and segregated BG-thalamo-cortical circuit solely devoted to song learning (in juveniles) and plasticity (in adults). In this promising animal model, BG function in sensorimotor learning can be approached in a simplified (compared to mammalian circuits) yet realistic model (Doupe et al., 2005). Experiments have shown that this circuit drives song variability, allowing vocal exploration of the sensorimotor space, and guides song changes to minimize vocal error during juvenile song learning (Mooney 2009).The physiological mechanisms underlying this learning remain largely unknown, and theoretical models are still few and far from a dynamic representation of the physiological processes at play (Fiete et al. 2007, Hahnloser & Ganguli 2013, Hanuschkin et al. 2013). Nevertheless, modeling these physiological mechanisms is essential to test new hypothesis concerning the various phases during learning and their organization.
We propose to reevaluate the role of the BG-talamo-cortical circuit in sensorimotor learning in a theoretical model based on recent experimental advances. The PhD project consists in developing this new theoretical model of the neuronal circuit involved in song learning in successive stages. The model will reflect the various behavioral phases observed during juvenile song learning, and will be constrained by available anatomical and physiological data on related brain circuits. After establishing new equations for the dynamics of the system under study, analytical investigation of these equations (wherever possible) will be combined with numerical simulations of the model’s dynamics through artificial neural networks. We propose to explore how such a neural network model may give rise to the various behavioral phases observed in young birds. Our model will then guide the experimental investigation of BG function in song learning through its precise predictions. Moreover, it will bring forward new models of machine learning for sequence acquisition and imitative learning.
The originality of our proposal comes from the desire to create a model that is not purely theoretical, but (1) takes into account real physiological processes, (2) generalize from the machine learning viewpoint and will be transposable to other similar problems (e.g. sensorimotor learning in robots), and (3) produce possible comparison with speech acquisition in children during development (Moulin-Frier et al. 2014). Indeed, understanding song learning in birds will shed light on speech acquisition in humans.
Project
First of all, several minimal circuit models implementing inverse learning will be suggested. Composed of two neuronal populations, one auditory and one motor, they will reproduce syllabic sounds one by one. In a second time, the model will be modified to learn the syllabic sequence(s) of the tutor, i.e. the temporal organization of syllables in song. By adding a reward signal denoting song quality in a third time, the model will include a reinforcement learning mechanism. Finally, the circuit will be included in a BG-thalamo-cortical loop network and the representation underlying its inputs (auditory) and outputs (motor, with an artificial syrinx) will be developed. We will evaluate the learning capacity of such a network model given the physiological constraints at play (non-specific dopaminergic signal, high dimensionality of input and output signals, non-linear transformations in input and output streams…). All along the model development, various learning rules based synaptic plasticity between neurons will be evaluated. These rules will be compared based on several criteria: biological plausibility, performance in a machine-learning sense (speed of convergence, minimal final error…), and the quality of emerging neuronal representations (optimality from a coding perspective and biological plausibility). The exploration of these models will include the comparison with existing models developed in the team of Xavier Hinaut: layer networks, in particular for inverse learning, or recurrent neural networks such as Reservoir Computing models (Jaeger 2004, 2014) and LSTM (Gers et al., 2000) for the temporal learning aspects. Finally, these explorations will be completed by the analysis of electrophysiological data obtained by the team of Arthur Leblois.
Required Knowledge and background
- Good background in maths and/or physics ;
- A strong interest for neuroscience and the physiological processes underlying learning;
- Python programming with experience in scientific libraries Numpy/Scipy (or similar coding language: matlab, etc.);
- Experience in machine learning or data mining is a preferred;
- Independence and ability to manage a project;
- Good English reading/speaking skills.
Advisors
This work will rely on the past work of both supervisors: A. Leblois (physicist by training) concerning songbird neurophysiology, and X. Hinaut concerning neuronal models of sequence learning applied to language and complex action sequences.
Xavier HINAUT (CR2 Inria)
Arthur LEBLOIS (HDR / CR2 CNRS, neurophysiology).
The PhD will be co-supervised by AL and XH at the Institute for Neurodegenerative diseases (Institut des Maladies Neurodégénératives – IMN –, Pellegrin Hospital Campus, Bordeaux).
References
Doupe AJ, Perkel DJ, Reiner A, Stern EA (2005) Birdbrains could teach basal ganglia research a new song Trends Neurosci. 28:353-63
Fiete, I. R., Fee, M. S., & Seung, H. S. (2007). Model of birdsong learning based on gradient estimation by dynamic perturbation of neural conductances. Journal of neurophysiology, 98(4), 2038-2057.
Mooney, R. (2009). Neural mechanisms for learned birdsong. Learning & Memory, 16(11), 655-669. doi:10.1101/lm.1065209
Hahnloser, R. H. R., & Ganguli, S. (2013) Vocal Learning with Inverse Models. In Principles of neural coding. Quiroga, Rodrigo Quian, and Stefano Panzeri, eds. CRC Press.
Hanuschkin, A., Ganguli, S., & Hahnloser, R. (2013). A Hebbian learning rule gives rise to mirror neurons and links them to control theoretic inverse models. Frontiers in Neural Circuits, 7, 106.
Jaeger, H., & Haas, H. (2004). Harnessing nonlinearity: Predicting chaotic systems and saving energy in wireless communication. Science, 304(5667), 78-80.
Jaeger, H. (2014) Controlling recurrent neural networks by conceptors. arXiv:1403.3369
Moulin-Frier, C., Nguyen, S. M., & Oudeyer, P. Y. (2014). Self-organization of early vocal development in infants and machines: the role of intrinsic motivation. Frontiers in Psychology, 4, 1006.
Gers, F. A., Schmidhuber, J., & Cummins, F. (2000). Learning to forget: Continual prediction with LSTM. Neural computation, 12(10), 2451-2471.
Keywords
Sensorimotor learning, modeling, machine learning, inverse model, basal ganglia, recurrent neural networks, sequence learning.
Duration
3 years
Probable date of beginning
October 1st, 2017
Salary
Monthly gross salary : 1982,00 € the first and second year and 2085,00 € the third year
Monthly net salary (after medical insurance) : 1593,50 € the first and second year and 1676,31 € the third year
Application
Click here for more informations and give your application
More info
For more information on the post or transmit its application, contact: xavier <dot> hinaut (home) inria -dot- fr
For information of administrative order or in case of difficulty to transmit its application, contact: laure.pottier_schupp (home) inria.fr