Jean-Pascal Pfister: Why spikes? Information theory and control theory perspectives
| When |
Feb 05, 2025
from 12:15 PM to 01:00 PM |
|---|---|
| Where | Bernstein Center, Hansastr. 9a, Lecture Hall. |
| Contact Name | Gundel Jaeger |
| Add event to calendar |
 vCal iCal
|
Abstract
The discovery of action potentials by Lord Edgar Adrian has been made 100 years ago and yet, embarrassingly, it remains largely unknown why and when spike-based representation is preferable over analog-based representation. Several hypotheses on the benefit of spike-based representation have been suggested (such as communication over long distances, energy efficiency, noise robustness, cell-wide synchronisation, fast responses, etc…), however a systematic and fair comparison between a spiking-based and an analog-based representation of information is lacking for all the proposed hypotheses.
In this talk, we will address the why spike question from an information theory as well as from a control theory perspective. In the first part, we will address the noise robustness hypothesis from a channel capacity perspective and study the conditions under which quantization is beneficial for information transmission. In particular, by considering the additive uniform channel with power constraint, we can demonstrate analytically that there is a phase transition as a function of the exponent of the power cost. In case of a sublinear cost (i.e. exponent smaller than one), the capacity achieving input distribution is quantized (i.e the input can take only discrete values between 0 and 1 - and possibly only value of 0 and 1 when the noise is large) whereas when the cost is supralinear, the capacity achieving distribution has full support on the interval [0,1]).
In the second part of the talk, in order to take into account the dynamical aspect of brain computation, we will take a control theory perspective to the why spike question since the brain can be seen as an event-based controller. In particular, we will investigate the conditions under which event-based control is preferable over analog-based control. We consider a simple system made of two antagonist muscles whose task is to track on oscillatory signal where the response of the motor units within a muscle are given by some response kernel that is characterised by some delay and time constant. In this simple set-up, we study the performance of event-based and analog-based controllers that are either proportional or integral. We found a qualitative phase transition. For large delays, the event-based control is preferable over the analog-based control whereas it is the opposite for small delays.
Taken together, those results highlight the need to carefully study the conditions under which spikes are beneficial. This work has been done in collaboration with Jonas Stapmanns, Luke Eilers and Catarina Reis Dias.