Ventral Visual Stream and Deep Networks

Transcription

1 Ma191b Winter 2017 Geometry of Neuroscience

2 References for this lecture: Tomaso A. Poggio and Fabio Anselmi, Visual Cortex and Deep Networks, MIT Press, 2016 F. Cucker, S. Smale, On the mathematical foundations of learning, Bulletin of the American Math. Society 39 (2001) N.1, 1 49.

3 Modeling Ventral Visual Stream via Deep Neural Networks Ventral Visual Stream considered responsible for object recognition abilities dorsal (green) and ventral (purple) visual streams responsible for first 100msc time of processing visual information from initial visual stimulus to activation of inferior temporal cortex neurons

4 mathematical model describing learning of invariant representations in the Ventral Visual Stream working hypothesis: main computational goal of the Ventral Visual Stream is compute neural representations of images that are invariant with respect to certain groups of transformations (mostly affine transformations: translations, rotations, scaling) model based on unsupervised learning far fewer examples are needed to train a classifier for recognition if using an invariant representation Gabor functions and frames optimal templates for simultaneously maximizing invariance with respect to translations and scaling architecture: hierarchy of Hubel Wiesel modules

5 a significant difference between (supervised) learning algorithms and functioning of the brain is that learning in the brain seems to require a very small number of labelled examples conjecture: key to reducing sample complexity of object recognition is invariance under transformations two aspects: recognition and categorization for recognition it is clear that complexity is greatly increased by transformations (same objects seen from different perspectives, in different light conditions, etc.) for categorizations also (distinguishing between different classes of objects: cats/dogs, etc.) transformations can hide intrinsic characteristics of an object empirical evidence: accuracy of a classifier per number of examples greatly improved in the presence of an oracle that factors out transformations (solid curve, rectified; dashed, non-rectified)

6

7 order of magnitude for number of different object categorizations (e.g. distinguishable different types of dogs) smaller than magnitude for different viewpoints generated by group actions reducing the variability by transformations makes greatly reduces the learning task complexity refer to sample complexity as number of examples needed for estimating a target function within an assigned error rate transform problem of distinguishing images into problem of distinguishing orbits under a given group action

8 Feedforward architecture in the ventral stream two main stages 1 retinotopic areas computing a representation that is invariant under affine transformations 2 approximate invariance to other object-specific dependencies, not described by group actions (parallel pathways) first stage realized through Gabor frames analysis overall model relies on a mathematical model of learning (Cucker-Smale)

9 architecture layers: red circle = vector computed by one of the modules, double arrow = its receptive field; image at level zero (bottom), vector computed at top layer consists of invariant features (fed as input to a supervised learning classifier)

10 biologically plausible algorithm (Hubel Wiesel modules) two types of neurons roles: simple cells: perform an operation of inner product with a template t H Hilbert space; a further non-linear operation (a threshold) is also applied complex cells: aggregate the outputs of several simple cells steps: (assume G finite subgroup of affine transformations) 1 unsupervised learning of group G by storing memory of orbit G t = {gt : g G} of a set of templates t H 2 computation of invariant representation: new image I H compute gt k, I for g G and t k, k = 1,..., K templates and µ k h (I) = 1 σ h ( gt k, I ) #G g G σ h a set of nonlinear functions (e.g. threshold functions)

11 computed µ k h (I) called signature of I signature µ k h (I) clearly G-invariant Selectivity Question: how well does µ k h (I) distinguish different objects? different meaning G I G I Main Selectivity Result (Poggio-Anselmi) want to be able to distinguish images within a given set of N images I, with an error of at most a given ɛ > 0 the signatures µ k h (I) can ɛ-approximate the distance between pairs among the N images with probability 1 δ provided that the number of templates used is at least K > c ɛ 2 log N δ more detailed discussion of this statement below; main point: need of the order of log(n) templates to distinguish N images

12 General problem: when two sets of random variables x, y are probabilistically related relation described by probability distribution P(x, y) some square loss problem (minimization problem) E(f ) = (y f (x)) 2 P(x, y) dx dy distribution itself unknown, but minimize empirical error E N (f ) = 1 N N (y i f (x i )) 2 i=1 over a set of random sampled data points {(x i, y i )} i=1,...,n if f N minimizes empirical error, want that the probability P( E(f N ) E N (f N ) > ɛ) is sufficiently small Problem depends on the function space where f N lives

13 General setting F. Cucker, S. Smale, On the mathematical foundations of learning, Bulletin of the American Math. Society 39 (2001) N.1, X compact manifold, Y = R k (for simplicity k = 1), Z = X Y with Borel measure ρ ξ random variable (real valued) on probability space (Z, ρ) expectation value and variance E(ξ) = ξ dρ, σ 2 (ξ) = E((ξ E(ξ)) 2 ) = E(ξ 2 ) E(ξ) 2 Z function f : X Y, least squares error of f E(f ) = (f (x) y) 2 dρ Z measures average error incurred in using f (x) as a model of the dependence between y and x Problem: how to minimize the error?

14 conditional probability ρ(y x) (probability measure on Y ) marginal probability ρ X (S) = ρ(π 1 (S)) on X, with projection π : Z = X Y X relation between these measures ( φ(x, y) dρ = Z X Y ) φ(x, y) dρ(y x) dρ X breaking of ρ(x, y) into ρ(y x) and ρ X (S) is breaking of Z into input X and output Y

15 regression function f ρ : X Y f ρ (x) = y dρ(y x) assumption: f ρ is bounded for fixed x X map Y to R via Y y y f ρ (x) expectation value is zero so variance σ 2 (x) = (y f ρ (x)) 2 dρ(y x) averaged variance σρ 2 = X Y σ 2 (x) dρ X = E(f ρ ) measures how well conditioned ρ is Note: in general ρ and f ρ not known but ρ X known

16 error, regression, and variance: E(f ) = ((f (x) f ρ (x)) 2 + σρ) 2 dρ X X What this says: σ 2 ρ is a lower bound for the error E(f ) for all f, and f = f ρ has the smallest possible error (which depends only on ρ) why identity holds:

17 Goal: learn (= find a good approximation for) f ρ given random samples of Z Z N z = ((x 1, y 1 ),..., (x N, y N )) sample set of points (x i, y i ) independently drawn with probability ρ empirical error E z (f ) = 1 N N (f (x i ) y i ) 2 i=1 for random variable ξ empirical mean E z (ξ) = 1 N N ξ(z i, y i ) i=1 given f : X Y take f Y : Z Y to be f Y : (x, y) f (x) y E(f ) = E(f 2 Y ), E z(f ) = E z (f 2 Y )

18 Facts of Probability Theory (quantitative versions of law of large numbers) ξ random variable on probability space Z with mean E(ξ) = µ and variance σ 2 (ξ) σ 2 Chebyshev: for all ɛ > 0 { P z Z m 1 : m } m ξ(z i ) µ ɛ i=1 σ2 mɛ 2 Bernstein: if ξ(z) E(ξ) M for almost all z Z then ɛ > 0 { } ( ) P z Z m 1 m : ξ(z i ) µ m ɛ mɛ 2 2 exp 2(σ Mɛ) i=1 Hoeffding: { P z Z m : 1 m } m ξ(z i ) µ ɛ i=1 ) 2 exp ( mɛ2 2M 2

19 Defect Function of f : X Y L z (f ) := E(f ) E z (f ) discrepancy between error and empirical error (only E z (f ) measured directly) estimate of defect if f (x) y M almost everywhere, then ɛ > 0, with σ 2 variance of fy 2 ( ) P{z Z m mɛ 2 : L z (f ) ɛ} 1 2ɛ exp 2(σ M2 ɛ) from previous Bernstein estimate taking ξ = f 2 Y when is f (x) y M a.e. satisfied? e.g. for M = M ρ + P M ρ = inf{ M : {(x, y) Z : y f ρ (x) M} measure zero } P sup f (x) f ρ (x) x X

20 Hypothesis Space a learning process requires a datum of a class of functions (hypothesis space) within which the best approximation for f ρ C(X ) algebra of continuous functions on topological space X H C(X ) compact subset (not necessarily subalgebra) look for minimizer (not necessarily unique) f H = argmin f H (f (x) y) 2 because E(f ) = X (f f ρ) 2 + σρ 2 also minimizer f H = argmin f H (f f ρ ) 2 continuity: if for f H have f (x) y M a.e., bounds E(f 1 ) E(f 2 ) 2M f 1 f 2 and for E z also, so E and E z continuous compactness of H ensures existence of minimizer but not uniqueness (a uniqueness result when H convex) Z X

21 Empirical target function f H,z minimizer (non unique in general) f H,z = argmin f H 1 m m (f (x i ) y i ) 2 i=1 Normalized Error E H (f ) = E(f ) E(f H ) E H (f ) 0 vanishing at f H Sample Error E H (f H,z ) E(f H,z ) = E H (f H,z ) + E(f H ) = X (f H,z f ρ ) 2 + σ 2 ρ estimating E(f H,z ) by estimating sample and approximation errors, E H (f H,z ) and E(f H ) one on H the other independent of sample z

22 bias-variance trade-off bias = approximation error; variance = sample error fix H: sample error E H (f H,z ) decreases by increasing number m of samples fix m: approximation error E(f H ) decreases when enlarging H procedure: 1 estimate how close f H,z and f H depending on m 2 how to choose dim H when m is fixed first problem: how many examples need to draw to say with confidence 1 δ that X (f H,z f H ) 2 ɛ?

23 Uniformity Estimate (Vapnik s Statistical Learning Theory) covering number: S metric space, s > 0, number N (S, s) minimal l N so that disks in S radii s covering S; for S compact N (S, s) finite uniform estimate: H C(X ) compact, if for all f H have f (x) y M a.e., then ɛ > 0 ( ) P{z Z m ɛ mɛ 2 : sup L z (f ) ɛ} 1 N (H, )2 exp f H 8M 4(2σ M2 ɛ) with σ 2 = sup f H σ 2 (f 2 Y ) main idea: like previous estimate of defect but passing from a single function to a family of functions, using a uniformity based on covering number

24 Estimate of Sample Error H C(X ) compact, with f (x) y M a.e. for all f H, and σ 2 = sup f H σ 2 (fy 2 ), then ɛ > 0 ( ) P{z Z m ɛ mɛ 2 : E H (f z ) ɛ} 1 N (H, )2 exp 16M 8(4σ M2 ɛ) obtained from previous estimate using L z (f ) = E(f ) E z (f ) so answer to first question: to ensure probability above 1 δ need to take at least m 8(4σ M2 ɛ) ɛ 2 ( log(2n (H, ) ɛ 16M )) + log(1 δ ) obtained by setting ( ) ɛ mɛ 2 δ = N (H, )2 exp 16M 8(4σ M2 ɛ) need various techniques for estimating covering numbers N (H, s) depending on the choice of the compact set H

25 Second Question: Estimating the Approximation Error E(f H,z ) = E H (f H,z ) + E(f H ) focus on E(f H ), which depends on H and ρ (f H f ρ ) 2 + σρ 2 X second term independent of H so focus on first; f ρ bounded, but not in H nor necessarily in C(X ) Main idea: use finite dimensional hypothesis space H; estimate in terms of growth of eigenvalues of an operator Main technique: Fourier analysis; Hilbert spaces

26 Fourier Series: start with case of X = T n = (S 1 ) n torus Hilbert space L 2 (X ) Lebesgue measure with complete orthonormal system φ α (x) = (2π) n/2 exp(iα x), α = (α 1,..., α n ) Z n Fourier series expansion f = α Z n c α φ α finite dimensional subspaces H N L 2 (X ) spanned by φ α with α B, dimension N(B) number of lattice points in ball radius B in R n N(B) (2B) n/2 H hypothesis space: ball H N,R of radius R in norm in H N

27 Laplacian on torus X = T n Laplacian : C (X ) C (X ) (f ) = n 2 f x 2 i=1 i Fourier series basis φ α are eigenfunctions of with eigenvalue α 2 more general X : bounded domain X R n with smooth boundary X and a complete orthonormal system φ k of L 2 (X ) (Lebesgue measure) of eigenfunctions of Laplacian with (φ k ) = ζ k φ k, φ k X 0, k 1 0 < ζ 1 ζ 2 ζ k subspace H N of L 2 (X ) generated by {φ 1,..., φ N } hypothesis space H = H N,R ball of radius R for in H N

28 Construction of f H Lebesgue measure µ on X and measure ρ (marginal probability ρ X induced by ρ on Z = X Y ) consider regression function f ρ (x) = Y y dρ(y x) assumption f ρ bounded on X so in L 2 ρ(x ) and in L 2 µ(x ) choice of R: assume also that R f ρ, which implies R f ρ ρ then f H is orthogonal projection of f ρ onto H N using inner product in L 2 ρ(x ) goal: estimate approximation error E(f H ) for this f H

29 Distorsion factor: identity function on bounded functions extends to J : L 2 µ(x ) L 2 ρ(x ) distorsion of ρ with respect to µ D ρµ = J operator norm: how much ρ distorts the ambient measure µ reasonable assumption: distorsion is finite in general ρ not known, but ρ X is known, so D ρµ can be computed

30 Weyl Law Weyl law on rate of growth of eigenvalues of the Laplacian (acting on functions vanishing on boundary of domain X R n ) N(λ) lim λ λ n/2 = (2π) n B n Vol(X ) B n volume of unit ball in R n ; N(λ) number of eigenvalues (with multiplicity) up to λ Weyl law: Li Yau version ζ k n n + 2 4π2 ( k B n Vol(X ) ) 2/n P. Li and S.-T. Yau, On the parabolic kernel of the Schrödinger operator, Acta Math. 156 (1986), from this get a weaker estimate, using explicit volume B n ( ) k 2/n ζ k Vol(X )

31 Approximation Error and Weyl Law norm K : for f = k=1 c k φ k with φ k eigenfunctions of ( ) 1/2 f K := ck 2 ζ k k=1 like L 2 -norm but weighted by eigenvalues of Laplacian in l 2 measure of c = (c k ) Approximation Error Estimate: for H and f H as above ( ) k 2/n E(f H ) Dρµ 2 f ρ 2 K Vol(X ) + σ2 ρ proved using Weyl law and estimates d µ (f ρ, H N ) 2 = f ρ f H ρ = d ρ (f ρ, H N ) J d µ (f ρ, H N ) k=n+1 where f ρ = k c kφ k c k φ k 2 µ = k=n+1 c 2 k = k=n+1 c 2 k ζ k 1 ζ k 1 ζ N+1 f ρ 2 K

32 Solution of the bias-variance problem mimimize E(f H,z ) by minimizing both sample error and approximation error minimization as a function of N N (for the choice of hypothesis space H = H N,R ) select integer N N that minimizes A(N) + ɛ(n) where ɛ = ɛ(n) as in previous estimate of sample error and A(N) = D 2 ρµ ( ) 2/n k f ρ 2 K + σρ 2 Vol(X ) from previous relation between m, R = f ρ, δ and ɛ obtain ( ɛ 288M2 N log( 96RM ) log( 1 ) m ɛ δ ) 0 find N that minimizes ɛ with this constraint no explicit closed form solution for N minimizing A(N) + ɛ(n) but can be estimated numerically in specific cases

33 back to the visual cortex modeling (Poggio-Anselmi) stored templates t k, k = 1,..., K and new images I in some finite dimensional approximation H N to a Hilbert space simple cells perform inner products gt k, I in H N estimate in terms of 1D-projections: I R d some in general large d; projections t k, I for a set of normalized vectors t k S d 1 (unit sphere) Z : S d 1 R +, Z(t) = µ t (I) µ t (I ) distance between images d(i, I ) think of as a distance between two probability distributions P I, P I on R d measure distance in terms of d(p I, P I ) Z(t) dvol(t) S d 1

34 model this in terms of ˆd(P I, P I ) := 1 K K Z(t k ) k=1 want to evaluate the error incurred in using ˆd(P I, P ci ) (1D projections and templates) to estimate d(p I, P I ) as in the Cucker-Smale setting, evaluate the error and the probability of error in terms of the Hoeffding estimate 1 K d(p I, P I ) ˆd(P I, P ci ) = Z(t k ) E(Z) K k=1 probability of error ( ) 1 K P Z(t k ) E(Z) K > ɛ k=1 if a.e. bound Z(t) E(Z) M 2e Kɛ2 2M 2

35 want this estimate to hold uniformly over a set of N images: want same bound to hold over each pair so error probability is at most ) ) N(N 1) exp ( Kɛ2 2Mmin 2 N 2 exp ( Kɛ2 2Mmin 2 δ 2 with M min the smallest M over all pairs This is at most a given δ 2 whenever K 4M2 min ɛ 2 log N δ

36 Group Actions and Orbits {t k } k=1,...,k given templates G finite subgroup of the affine group (translations, rotations, scaling) G acts on set of images I: orbit GI projection P : R d R K of images I onto span of templates t k Johnson Lindenstrauss lemma: low distorsion embeddings of sets of points from a high-dimensional to a low-dimensional Euclidean space (special case with map an orthogonal projection) given 0 < ɛ < 1; given finite set X of n points in R d take K > 8 log(n)/ɛ 2 then there is a linear map f given by a multiple of an orthogonal projection onto a (random) subspace of dimension K such that, for all u, v X (1 ɛ) u v 2 R d f (u) f (v) 2 R K (1 + ɛ) u v 2 R d result depends on concentration of measure phenomenon

37 up to a scaling, for a good choice of subspace spanned by templates, can take P to satisfy Johnson-Lindenstrauss lemma starting from finite set X = {u} of images, can generate another set by including all group translates X G = {g u : g G, u X } then for Johnson-Lindenstrauss lemma required accuracy for X G log(n #G) K > 8 ɛ 2 so can estimate sufficiently well the distance between images in R d using the distance between projections t k, gi of their group orbits onto the space of templates by t k, gi = g 1 t k, I for unitary representations it would seem one needs to increase by K #G K the number of templates to distinguish orbits, but in fact by argument above need an increase K K + 8 log(#g)/ɛ 2

38 given t k, gi = g 1 t k, I computed by the simple cells, pooling by complex cells by computing µ k h (I) = 1 #G σ h ( gt k, I ) g G σ h a set of nonlinear functions: examples µ k average(i) = 1 #G g G gt k, I µ k energy(i) = 1 #G g G gtk, I 2 µ k max(i) = max g G gt k, I other nonlinear functions: especially useful case, when σ h : R R + is injective Note: stored knowledge of gt k for g G allows the system to be automatically invariant wrt G action on images I

39 Localization and uncertainty principle would like templates t(x) to be localized in x: small outside of some interval x would also like ˆt to be localized in frequency: small outside an interval ω but... uncertainty principle: localized in x / delocalized in ω x ω 1

40 Optimal localization optimal possible localization when x ω = 1 realized by the Gabor functions t(x) = e iω 0x e x2 2σ 2

41 each cell applies a Gabor filter; plotted n y /n x anisotropy ratios