Project

Back to overview

Filtering with Spiking Neurons

English title Filtering with Spiking Neurons
Applicant Pfister Jean-Pascal
Number 179060
Funding scheme SNSF Professorships
Research institution Institut für Physiologie Medizinische Fakultät Universität Bern
Institution of higher education University of Berne - BE
Main discipline Biophysics
Start/End 01.09.2018 - 31.08.2020
Approved amount 655'881.00
Show all

All Disciplines (2)

Discipline
Biophysics
Neurophysiology and Brain Research

Keywords (12)

Point emission processes; normative approach; graphical model; spiking neuron model; Bayesian filtering; short-term plasticity; diffusion process; learning; synaptic plasticity; recurrent network; Decision making; Bayesian inference

Lay Summary (French)

Lead
Le nombre de neurones pouvant être enregistrés simultanément augmente d’année en année. Cependant les algorithmes actuels permettant de décoder un signal à partir d’enregistrements dynamiques et bruités ne fonctionnent pas bien lorsque le nombre de neurones enregistré est large. Ce projet a pour but de développer de nouveaux algorithmes de décodage plus adaptés au type de données expérimentales actuelles.
Lay summary
Les neurones communiquent entre eux en par le moyen de signaux digitaux (les potentiels d’actions) alors que beaucoup de variables qui doivent être encodées par le cerveau sont analogiques, comme la position, la vitesse ou la direction fixée par un animal. Comment donc décoder ces variables analogues extérieures à partir d’enregistrements de potentiels d’actions qui, de plus, ont été générés de façon partiellement aléatoire? Il existe une classe d’algorithmes - appelés filtres particulaires - qui permet d’extraire, de façon dynamique, un signal à partir d’observations bruitées. Cependant, ces filtres particulaires ne sont pas utilisables lorsque le nombre d’observables est large car il faudrait trop de particules pour correctement estimer l’incertitude sur le signal. Or, il se trouve que le nombre de neurones qui peuvent  être simultanément enregistrés augmente d’année en année. Le but du premier projet est donc de développer un nouveau type de filtre particulaire qui est utilisable lorsque le nombre d’observables (neurones simultanément enregistrés) est élevé. Le deuxième projet utilisera cette méthode à des enregistrements chez le singe (du laboratoire du Prof. Valerio Mante, UZH) afin d’obtenir une description statistique de la dynamique des neurones lors de la prise de décision.
Direct link to Lay Summary Last update: 21.12.2018

Responsible applicant and co-applicants

Employees

Associated projects

Number Title Start Funding scheme
150637 Inference and Learning with Spiking Neurons 01.09.2014 SNSF Professorships
175644 Bayesian synapses 01.10.2018 Project funding (Div. I-III)

Abstract

One of the most fascinating aspect of the brain is its ability to continuously extract relevant features from dynamic environment such as the position of a moving prey from ambiguous visual stimuli or the voice of a known person in a noisy crowd. This problem of extracting hidden and dynamic features from noisy observations can be addressed mathematically by Bayesian filtering theory. It remains however unknown how networks of neurons actually implement such a Bayesian filter.During the first 3 years of the SNF grant, we made significant progresses at the interface between mathematical filtering theory and its biological implementation. At the single synapse level, we proposed a new spiking neuron model for in-vivo dynamics (Surace and Pfister 2015) that is the key element for the validation of our synaptic filtering theory (Pfister et al. 2009; Pfister et al. 2010). At the neural network level, we proposed the first network dynamic model - the Neural Particle Filter (NPF) - that can perform nonlinear Bayesian filtering whilst implementable biologically (Kutschireiter et al. 2017). This NPF has several relevant properties. On the biological side, since this algorithm does not depend on importance weights, it can be easily implemented by a neural network. On the machine learning side, we could show that due to this weight-less particle filter approach, the NPF avoids the curse of dimensionality (Kutschireiter et al. 2017; Surace et al. 2017). Further, we could directly apply the NPF to an online-learning framework (Surace and Pfister 2016), this way deriving an online learning rule which even becomes local in the limit of small observation noise.The aim of the present proposal is to extend the original SNF prof grant proposal entitled inference and learning with spiking neurons in two directions. The first project will extend our recent work on the neural network implementation of nonlinear filtering. We proposed the first implementation of nonlinear filtering (the Neural Particle Filter) with a network of neurons, but those neurons were assumed, for simplicity reasons, to be analog neurons (Kutschireiter et al. 2017). Here we will consider a biologically more relevant scenario where neurons are spiking and potentially include refractory periods. We will also study the properties of this new Spiking Neural Particle Filter (SNPF), in particular its scalability in high dimensions.The second project will apply the SNPF to spiking data that have been recorded in Monkeys performing a decision making task (data from Prof. Valerio Mante Lab, INI, Zurich). The goal will be to bring a new view in conflicting literature about the nature of the statistical model that best captures recorded spiking trains. Some argued that the underlying firing rate slowly ramps up to a threshold - the ramping model (Mazurek 2003; Gold and Shadlen 2007) while others proposed a stepping model where the firing rate suddenly jumps to an elevated firing rate (Latimer et al. 2015). The idea here will be to propose a generalised model which includes both the ramping and the stepping model as specific case and determine to what extent the underlying rate is more consistent with a ramping or a stepping model.
-