[PPT] - Towards biologically plausible regularization mechanisms T. Vi PowerPoint Presentation

SLIDE 1

Towards biologically plausible regularization mechanisms

T. Vi´

eville et cie November 25, 2003

SLIDE 2

Which part of the brain do we study ?

What and where visual streams:

from [2] 2

SLIDE 4

E.g. motion processing:

from [7] from [4] 3

SLIDE 5

Fast brain: how fast ? backward connection may be faster than forward

from [2] 4

SLIDE 6

Forward / backward connections in the cortex

Following [6], let us review that the visual cortices can be considered as a hierarchy of cortical levels with reciprocal extrinsic corticocortical connections among the constituent cortical areas [5]. The notion of a hierarchy depends upon a distinction between forward and backward extrinsic connections. This distinction rests upown different laminar specificity [9, 10]

Forwards connections Backwards connections Sparse axonal bifurcations Abundant axonal bifurcation Topographically organized Diffuse topography Originate in supragranular layers 2/3 Originate in bilaminar/infragranular layers 5/6 Terminate largely in layer 4 Terminate predominantly in layer 1, but all layers except 4 Postsynaptic effects through fast Modulatory afferents activate slow

AMPA (1.3-2.4 ms decay) and GABAA (6 ms decay) receptors (50 ms decay) voltage-sensitive NMDA receptors

Comment: postsynaptic/modulatory effect is a conjecture, so is assumptions on synapses mechanisms

5

SLIDE 7

Forward connections are concerned with the promulgation and segregation of sensory information, consistent with: (i) their sparse axonal bifurcation; (ii) patchy axonal terminations; and (iii) topographic projections. Backward connections are considered to have a role in mediating contextual effects and in the co-ordination of processing channels, consistent with: (i) their frequent bifurcation; (ii) diffuse axonal terminations; and (iii) non-topographic projections [10] (iv) slow time-constants.

forward connections are driving and backward connections are modulatory,

as suggested by reversible inactivation [11, 8] and functional neuroimaging [1]

backward connections are more numerous and transcend more levels,

e.g. the ratio of forward efferent connections to backward afferents in the lateral geniculate is about 1:10/20; there are backward connections from TE and TEO to V1 but no monosynaptic connections from V1 to TE or TEO [10].

backward connections are more divergent than forward connections [14], one point in a given cortical area will connect to a region

5-8mm in diameter in another; the divergence region of a point in V5 (i.e. the region receiving backward afferents from V5) may include thick and inter-stripes in V2, whereas its convergence region (i.e. the region providing forward afferents to V5) is limited to the thick stripes [14]. They are faster than direct lateral connections [?] 6

SLIDE 8

The role of feedbacks: a few assumptions

where to process simple cues allow to decide in which part of the retinal map visual

processing have to occur:

a rough but fast edge detector may be used to determine which contrasted areas have to analyzed
large scale (smoothed, eliminating noise) orientation detector may help to tune further process of visual cues (e.g. figure/background

segmentation),

low-level but efficient focus of attention towards close, mobile or textured objects is derived from fast rotationally stabilized motion

perception;

what to process very fast object recognition, as experimented and modelized by, e.g.

[12, 13], allows a “model-based” processing of the visual information with the possibility to choose

among memorized previous processing modes, configurations of parameters tuned with respect to this first recognition. 7

SLIDE 9

Holistic perception: naive introspection shows that we are able to almost instantaneously perform the grouping of local

and even random tokens or attributes (e.g. the dalmatian picture), yielding a global object perception. This includes “hallucinations”

f 2D or 3D shapes from partial information (e.g.

the Kanizsa triangle).

Holistic perception may be related to fast-brain object labelization: feedbacks from what has been detected by this “1st stage” may be the key feature of holistic perception.

Opportunism :

One principle which seems to control all visual perception is that “the end justifies the means”. This means that in order to elaborate our percepts, our brain has an extraordinary capacity to combine several attributes (color, texture, motion, stereo, etc..), but always choosing those well adapted to a given context or to a given task. This occurs dynamically and without any conscious effort. Feedbacks in the visual cortex seems to be used to select the relevant

attributes, given a task or context, as soon as this state has been either detected or input to the system by higher “layers” of the cortex or obtained from a-priori information.

8

SLIDE 10

Feedbacks: a view of the Friston framework

The brain is a .. machine to find “causes” ν from inputs u
The world is a deterministic dynamical system :

˙ x = f(x, ν) x(−∞) = 0 u = g(x) (1) The variable x is the “hidden” deterministic system state (“universal” but trivial (!) representation)

9

SLIDE 11

In (1), we can eliminate (in fact “integrate”) the influence of x: we directly relate the input u = P(ν, β) to the recent history of the causes ν:

u(t) =

t

κ1(τ) ν(t − τ) dτ

influence from previous causes

+

t
t

κ2(τ, τ ′) ν(t − τ) ν(t − τ ′) dτdτ ′

modulatory influence between causes

+.. (2)

i.e. we parameterize this causal relationship with parameter:

β = [κ1(τ) =

∂u(t) ∂ν(t−τ)

t=0 , .., κ2(τ, τ ′) =

∂u(t) ∂ν(t−τ) ∂ν(t−τ′)

t=0 , ..]

Fliess fundamental formula and Volterra kernels

10

SLIDE 12

v u v u= P( , )

input prediction causes representation

v

forward connection backward connection

= R(u , )

Infered ‘‘causes’’ Inputs

adapted from [6]

The inference is coherent iff : u = P(R(u, Φ), β)

Expectation: is parameterized by forward connections Φ

expectation “infers” the causes from the given inputs

estiMation: is parameterized by backward connections β

estimation “predicts” the input from “a-priori” causes

11

SLIDE 13

Let us find the “maximally probable” estimation of the causes ν, knowing u: maxν log(p(ν|u)) = maxν  log(p(u|ν)) + log(p(ν)) −log(p(u))

don’t care

  maximal likehood maxν log(p(u|ν)) + log(p(ν)) Conditional information A priori information β tuning : u = P(ν, β) Φ tuning : ν = R(u, Φ) maxν log(p(P(ν, β)|ν)) + log(p(R(u, Φ)) Estimation Expectation

12

SLIDE 14

from [6] 13

SLIDE 15

Which probability ? additive error u = P(ν, β) + ǫ Gibb’s potential p(ǫ) = e−V (ǫ,θ)

V p p = e − V

e.g. exponential : V (ǫ, θ) = log(c(θ) + b(θ)T a(ǫ)) . . Gaussian, Poisson, binomial, uniform, ..

minβV (u − P(ν, β), θ) minΦ θV (u − P(R(u, Φ), β), θ)

in the formalism θ is “expectated” with Φ
the probability distribution must be “chosen”
convergence is “guaranteed” .. towards a sub-optimal (local minimum) solution
the minimization process corresponds to our “pde” approach

14

SLIDE 16

A variational view of a cortical map computation

computation

m

map

h(m) h

map input map output transformation

ptional input

, L

map parameters

min L , L = 1

2

||h − ¯

h||2

Λ + ||∇h||2 L

Λ:

measurement information metric:

precision of the input function, partial

bservations, missing data → linear relations between some measures
L: diffusion tensor:

(a) When the problem is ill-posed, we choose the solution which variations are minimized. (b) When partially or approximately defined, the value is defined using information “around’.

15

SLIDE 17

Implementation using an integral approximation of the differential operator:

∂hk

t

∂t (pu) = −∇Lt(pu) = l

Λk

l

hl

t − ¯

hl

t

(pu) −

v Luv kl hl t(pv)

(1st order scheme)

computes a weighted mean value around x minus the balanced value at x

from [3] .

extra cortical input or intra-cortical input from previous layers

(layers IV of the cortex) correspond to the input variable ¯ h,

extra cortical or backward intra-cortical output (layers V of the

cortex) correspond to the output the computed variable h,

local intra cortical connections (layers II/III of the cortex)

correspond to inter-parameters interaction, i.e. forward inputs to define Λ and L, thus σ,

backward intra cortical connections (layers I of the cortex)

correspond to inter-parameter interactions, i.e. backward inputs to define Λ and L, thus σ,

internal connections correspond to the local diffusion mechanism,

i.e. correspond to the integration over the operator variable. 16

SLIDE 18

Implementing feedbacks between cortical maps

minLi with Li = 1

2

||hi − ¯

hi||2

Λi(S∗h•i) + ||∇hi||2 Li(S∗h•i)

Forward/backward connections:

(i) backward cortical value are smoothed in space (spatial divergence), (ii) a cortical map value does not modify its own paramaters, (iii) forward connections forms a lattice (no loops in forward connections) .

Key fact: minimizing each criterion ∂hi

∂t = −∇Li is equivalent to minimize:

L• =

i(||∇iLi||)α Li,

α > 2

we have a view of the common action of these maps in interaction
we have a proof of stability / convergence

17

SLIDE 19

References

[1]

C. B¨

uchel and K.L. Friston. Modulation of connectivity in visual pathways by attention: cortical interactions evaluated with structural equation modelling and fmri. Cereb. Cortex, 7:768–778, 1997. [2]

J. Bullier. Integrated model of visual processing. Brain Res. Reviews, 36:96–107, 2001.

[3]

Y. Burnod. An adaptive neural network: the cerebral cortex. Masson, Paris, 1993. 2nd edition.

[4]

J. Chey, S. Grossberg, and E. Mingolla. Neural dynamics of motion processing and speed discrimination. Vision Res., 38:2769–2786,

1997. [5] D.J. Felleman and D.C. Van Essen. Distributed hierarchical processing in the primate cerebral cortex. Cereb Cortex, 1:1–47, 1991. [6] Karl Friston. Functional integration and inference in the brain. Prog Neurobiol, 68:113–143, 2002. [7] M.A. Giese and T. Poggio. Neural mechanisms for the recognition of biological movements and actions. Nature Neuroscience, 2003. in press. [8]

P. Girard and J. Bullier. Visual activity in area v2 during reversible inactivation of area 17 in the macaque monkey. J. Neurophysiol.,

62(6):1287–1301, 1989. [9] K.S. Rockland and D.N. Pandya. Laminar origins and terminations of cortical connections in the occipital lobe in the rhesus monkey. Brain Res, 179:3–20, 1979. [10] P.A. Salin and J. Bullier. Corticocortical connections in the visual system: structure and function. Psychol. Bull., 75:107–154, 1995. [11] J.H. Sandell and P.H. Schiller. Effect of cooling area 18 on striate cortex cells in the squirrel monkey. J. Neurophysiol., 48:38–48, 1982. [12]

S. Thorpe, D. Fize, and C. Marlot. Speed of processing in the human visual system. Nature, 381:520–522, 1996.

[13] S.J. Thorpe and M. Fabre-Thorpe. Seeking categories in the brain. Science, 291:260–263, 2001. [14]

S. Zeki and S. Shipp. The functional logic of cortical connections. Nature, 335:311–316, 1988.

18