Project

Back to overview

Normative theory of synaptic plasticity across multiple time scales

Applicant Pfister Jean-Pascal
Number 137200
Funding scheme Ambizione
Research institution Institut für Physiologie Medizinische Fakultät Universität Bern
Institution of higher education University of Berne - BE
Main discipline Biophysics
Start/End 01.10.2011 - 31.08.2014
Approved amount 598'048.00
Show all

All Disciplines (2)

Discipline
Biophysics
Neurophysiology and Brain Research

Keywords (6)

synaptic plasticity; short-term plasticity; long-term plasticity; sequence learning; spiking neuron model; recurrent network

Lay Summary (English)

Lead
Lay summary
Learning and memory are the crucial processes by which our past experiences influence our present decisions. Those processes - which ultimately characterize who we are - are widely thought to be implemented at the level of the individual connections between neurons: the synapses. Despite their fundamental role, the dynamics of synapses remain poorly understood. Indeed, far from being static transmission relays transferring information from one neuron to the next, synapses are highly modifiable (plastic) elements and those changes act on a wide range of different time scales (nine orders of magnitude, i.e from milliseconds to weeks). For example, short-term plasticity  (which is induced by presynaptic activity) lasts for tens to hundreds of milliseconds  while long-term plasticity (which requires both pre- and the postsynaptic activity) lasts for hours to days.

In addition to the complexity induced by this wide range of time scales, the biophysical processes underlying synaptic dynamics are mediated by a large number of biochemical elements whose interactions are not yet well understood. So is there any hope to understand synaptic plasticity? Here we approach the question from a reverse-engineering perspective. This approach is well illustrated by the quote from Horace Barlow: "A wing would be a most mystifying structure if one did not know that birds flew". In the context of synaptic plasticity, this means that we will start by postulating potential functional roles to which synapses obey, then derive the plasticity rules and finally compare the results to biological data.

The grant proposal is divided into 2 parts. In the first project, the goal is to elaborate a unifying theoretical framework which embraces both short-term and long-term plasticity aspects. This study will extend our previous modeling work on short-term plasticity. In our earlier work on short-term plasticity, we showed that the biologically observed short-term plasticity dynamics can be matched to the dynamics of an optimal estimator of the presynaptic membrane potential based on the past observed spikes. We will extend this framework by assuming that the model parameters describing the presynaptic membrane potential dynamics are not known a priori by the synapse but have to be learned (at a slower time scale). Another generalization of this model will be to assume that the synapse estimates a potentially non-linear function of the presynaptic membrane potential where this non-linear function is parametrized by the long-term weight that can evolve on a slower time scale. Both extensions will provide testable predictions on how short-term and long-term plasticity should interact from a normative perspective.

The second project aims at developing a normative model for temporal sequence learning in spiking recurrent networks. Several brain systems need to be able to store and recall precise sequence of spikes. For example, the juvenile songbirds learn a song from their fathers and reproduce it for the rest of their lives. As a consequence, juvenile songbirds must learn a stereotypical spiking pattern that will eventually produce the father’s song. Here we will address this problem from a general perspective. Instead of learning a specific spiking temporal sequence, the recurrent network will learn a distribution of sequences. We will derive analytically the corresponding learning rule and compare it to biologically plausible learning rule.
Direct link to Lay Summary Last update: 21.02.2013

Responsible applicant and co-applicants

Employees

Publications

Publication
Nerve injury-induced neuropathic pain causes disinhibition of the anterior cingulate cortex
Blom Sigrid, Pfister Jean-Pascal, Senn Walter, Nevian Thomas (2014), Nerve injury-induced neuropathic pain causes disinhibition of the anterior cingulate cortex, in Journal of Neuroscience, 34(17), 5754-5764.
Adaptive Gaussian Poisson process: a model for in vivo neuronal dy- namics
Surace Simone Carlo, Pfister Jean-Pascal (2013), Adaptive Gaussian Poisson process: a model for in vivo neuronal dy- namics, in Cosyne Abstracts 2013, Salt Lake City USA. , 1.
Learning Activity Patterns in Recurrent Networks of Visible and Hidden Spiking Neurons
Brea Johanni, Senn Walter, Pfister Jean-Pascal (2013), Learning Activity Patterns in Recurrent Networks of Visible and Hidden Spiking Neurons, in Cosyne Abstracts 2013, Salt Lake City USA. , 1.
Matching recall and storage in sequence learning with spiking neural networks.
Brea Johanni, Senn Walter, Pfister Jean-Pascal (2013), Matching recall and storage in sequence learning with spiking neural networks., in The Journal of neuroscience : the official journal of the Society for Neuroscience, 33(23), 9565-75.
A triplet spike-timing-dependent plasticity model generalizes the Bienenstock-Cooper-Munro rule to higher-order spatiotemporal correlations.
Pfister Jean-Pascal (2011), A triplet spike-timing-dependent plasticity model generalizes the Bienenstock-Cooper-Munro rule to higher-order spatiotemporal correlations., in PNAS, 108(48), 19383-19388.
Sequence learning with hidden units in spiking neural networks
Brea Johanni, Senn Walter, Pfister Jean-Pascal (2011), Sequence learning with hidden units in spiking neural networks, in Advances in Neural Information Processing Systems, 24, 1422-1430.
Normative theory of STP predicts a link between short-term facilitation and refractoriness
Surace Simone, Pfister Jean-Pascal, Normative theory of STP predicts a link between short-term facilitation and refractoriness, in Cosyne Abstracts 2014, Salt Lake City USA. .
Reinforcement Learning in Cortical Networks
Senn Walter, Pfister Jean-Pascal, Reinforcement Learning in Cortical Networks, in Encyclopedia of Computational Neuroscience.

Collaboration

Group / person Country
Types of collaboration
Carandini lab, UCL Great Britain and Northern Ireland (Europe)
- in-depth/constructive exchanges on approaches, methods or results
Laboratory of Michael Long, NYU Medical School United States of America (North America)
- in-depth/constructive exchanges on approaches, methods or results
Engineering Department, University of Cambridge Great Britain and Northern Ireland (Europe)
- in-depth/constructive exchanges on approaches, methods or results
Birdsong Group, Institute of Neuroinformatics, University of Zurich / ETH Zurich Switzerland (Europe)
- in-depth/constructive exchanges on approaches, methods or results

Communication with the public

Communication Title Media Place Year
New media (web, blogs, podcasts, news feeds etc.) Wie das Gehirn verschiedene Melodien lernen kann German-speaking Switzerland 2013

Associated projects

Number Title Start Funding scheme
150637 Inference and Learning with Spiking Neurons 01.09.2014 SNSF Professorships

Abstract

Learning and memory are the crucial processes by which our past experiences influence our present decisions. Despite their fundamental importance, we do not yet fully understand the behavior of synapses (the connections between neurons) which are widely believed to underlie learning and memory. Indeed synapses are highly dynamical elements with time constants from the millisecond range up to the hour or days range. Another source of complexity is that the biophysical processes underlying synaptic dynamics are mediated by more than hundred different proteins. Instead of studying synaptic dynamics from a bottom-up (biophysical) perspective, we will follow a top-down (normative) approach which has encountered an increasing success during the last 50 years. Indeed, this approach seems better suited if we want to better understand the functions of synaptic transmission and plasticity and not get lost in a mass of irrelevant details. Here, we propose two different projects that strive to understand the computational principles that govern synaptic plasticity across multiple time scales.The first project (that will be carried out by the PhD student) aims to link short-term plasticity and long-term plasticity. We will start from a normative model of short-term plasticity that we recently developed. This model sees short-term plasticity as an optimal (Bayesian) estimator of the presynaptic membrane potential. We will first validate this model with in vivo data that will be collected in Hahnloser's lab in Zurich. Then we will extend it in two different directions that will both include long-term aspects. First, we will relax the assumption that the synapse knows a priori the statistics of the presynaptic membrane potential. As a consequence the synapse must learn, at a slower time scale, the parameters describing the presynaptic dynamics. The second extension will be to assume that the task of the synapse is to estimate a (potentially non-linear) function of both the presynaptic membrane potential and the long-term weight. As a consequence, in this framework, the induction of long-term potentiation or long-term depression influences the short-term properties of the synapse. This formalism leads to directly testable predictions on the interaction between short-term and long-term plasticity.The second project aims to develop a normative model for temporal sequence learning in spiking recurrent networks. Indeed, several brain systems need to be able to store and recall precise sequence of spikes. For example, the juvenile songbirds learn a song from their fathers and reproduce it for the rest of their lives. As a consequence, juvenile songbirds must learn a stereotypical spiking pattern that will eventually produce the father's song. Here we will address this problem from a general perspective. Instead of learning a specific spiking temporal sequence, the recurrent network will learn a distribution of sequences. We will also derive the conditions under which the recall dynamics is stable. Finally, we will calculate the storage capacity of such networks.
-