An information theoretical approach to prefrontalexecutive function

Etienne Koechlin et al

Stefan KlampflComputational Intelligence Seminar A, SS 2009

Jun 10, 2009

Introduction

I The prefrontal cortex (PFC) subserves executive control -the ability to select actions on the basis of internal plans orgoals

I We still know very little about its functional organization

I Current theories fail to describe functional specializationand functional integration simultaneously

I This approach by Koechlin et alI uses information theory to describe how executive function can

be divided into hierarchically ordered control processesI maps these processes onto a network of brain regions using

data from neuroimaging experimentsI reveals functional integration within this network by describing

connectivity analyses

Quantifying executive control

I The demand of executive control can be measured as theamount of information required for action selection

I Total amount of information required for selecting an actiona: H(a) = − log2 p(a)

I For the current stimulus s:

H(a) = I (s, a) + Q(a|s)

I Sensorimotor control I (s, a) = log2[p(s, a)/p(a)p(s)]: theinformation provided by stimulus s for selecting a

I Cognitive control Q(a|s) = − log2 p(a|s): informationrequired for selecting a which is unrelated to stimulus s

Quantifying executive control

I Cognitive control can be further subdivided into informationprovided by the immediate context and into informationprovided by past events:

Q(a|s) = I (c , a|s) + Q(a|s, c)

I Contextual control I (c , a|s): information conveyed bycontextual signals c independent of s

I Episodic control Q(a|s, c): remaining information conveyedby past events, independent of c

Quantifying executive control

I Episodic control can be further decomposed into control owingto signals preceding and following a past instruction cue u

Q(a|s, c) = I (u, a|s, c) + Q(a|s, c , u)

I I (u, a|s, c): information conveyed by the instruction cue u (orevents following it)

I Branching control Q(a|s, c , u): remaining informationconveyed by past events preceding u

I enables a task or behavioral episode to be interrupted andtemporally suspended into a pending state

Quantifying executive control

I Executive function can be subdivided into hierarchicallyordered control processes

I Each process is responsible for selecting an action based oninformation that is successively more remote in time

sensory control H(a) = I (s, a) + Q(a|s)contextual control Q(a|s) = I (c , a|s) + Q(a|s, c)

episodic control Q(a|s, c) = I (u, a|s, c) + Q(a|s, c , u)

branching control Q(a|s, c , u)

I Hypothesis: these processes are associated with distinctregions of the PFC

The cascade model of cognitive control

regions along the anterior-posterior (rostral-caudal) axis of lateral PFC

Experimental setup

I fMRI recordings of 12 human subjects

I 2 behavioral experiments to separately vary the amount ofsensory, contextual, and episodic control

I series of visual stimuli broken down into successive blocks(behavioral episodes) preceded by distinct instruction cues(episodic signals)

I motor experiment: respond with left or right hand (R1/R2)depending on the color of the stimulus

I task experiment: perform different letter discrimination tasks(T1/T2) depending on the color (vowel/consonant orupper/lower case)

Experimental setup

R1/R2: respond with left or righthand

T1/T2: perform eithervowel/consonant or upper/lowercase discrimination task

NG: no go stimulus

I (s, a): sensory controlI (c , a|s): contextual controlQ(a|s, c): episodic control

Reaction times to stimuli

motor experiment:open circles: I (s, a) = 0open squares: I (s, a) = 1

task experiment:solid circles: I (c , a|s) = 0solid squares: I (c , a|s) = 1

I reaction times significantly increased with sensory, contextualand episodic control

I episode effect was significantly steeper in the task than in themotor experiment

Topography of brain activations

fmRI data confirmed the implementation of the cascade model inthe frontal lobes

green: stimulus effect; yellow: context effect, but no stimuluseffect; red: episode effect, but no stimulus and context effect

Analysis of regional activations

I Rostral LFPC activationsonly depend on episodiccontrol

I Caudal LFPC regionsshowed effects of contextand episode, but not ofstimulus

I Episode effect wassteeper in the taskexperiment

I Premotor regions showedeffects of stimulus,context and episode

I All effects wereindependent ofhemisphere

Effective connectivity in the frontal lobes

I Model was reformulated as a model of structural linearequations with path coefficients

I Observed effects result from top-down control from rostral tocaudal LPFC and premotor regions

Effective connectivity in the frontal lobes

I (Koechlin et al., 1999) investigated the neural correlates ofbranching

I Subjects were required to temporally suspend an ongoing taskin favor of another

I Branching control is associated with yet more antererior PFCregions (frontopolar cortex)

Summary

I The human lateral PFC is functionally organized as ahierarchy of control processes (sensory, contextual, episodic,branching control) implemented along the anterior-posterioraxis (cascade model)

I This hierarchy is based on the temporal structure of executivecontrol

I Direct evidence for the cascade model was obtained by fMRIdata

I Information theory provides a quantitative approach tounderstanding the function of the PFC

I The cascade model is supported by a wide variety ofneuroimaging, single-unit, and lesion studies

I The cascade model offers a conceptual framework forunderstanding functional fractionation and functionalintegration within the lateral PFC

References

Koechlin, E., Basso, G., Pietrini, P., Panzer, S., and Grafman, J. (1999).

The role of the anterior prefrontal cortex in human cognition.

Nature, 399:148–151.

Koechlin, E., Ody, C., and Kouneiher, F. (2003).

The architecture of cognitive control in the human prefrontal cortex.

Science, 302(5648):1181–1185.

Koechlin, E. and Summerfield, C. (2007).

An information theoretical approach to prefrontal executive function.

Trends in Cognitive Sciences, 11(6):229–235.