Extrinsic Geometrical Methods for Neural Blind Deconvolution Simone Fiori
[email protected]
Dipartimento di Elettronica, Intelligenza Artificiale e Telecomunicazioni Universit`a Politecnica delle Marche Ancona (Italy, EU)
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.1/34
Talk overview Blind deconvolution (BD): Signal model, basic assumptions, Bayesian (‘Bussgang’-type) BD.
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.2/34
Talk overview Blind deconvolution (BD): Signal model, basic assumptions, Bayesian (‘Bussgang’-type) BD. Automatic gain control (AGC): Source of geometrical structure of the parameter space.
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.2/34
Talk overview Blind deconvolution (BD): Signal model, basic assumptions, Bayesian (‘Bussgang’-type) BD. Automatic gain control (AGC): Source of geometrical structure of the parameter space. Algorithms: Geodesic-based and projection-based.
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.2/34
Talk overview Blind deconvolution (BD): Signal model, basic assumptions, Bayesian (‘Bussgang’-type) BD. Automatic gain control (AGC): Source of geometrical structure of the parameter space. Algorithms: Geodesic-based and projection-based. Numerical experiments and comparison.
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.2/34
BD: Channel Model Channel output signal model: T
xn = h sn + νn , def
sn = [sn sn−1 sn−2 . . . sn−Lh +1 ]T is the system’s input vector-stream at time n = 1, . . . , N,
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.3/34
BD: Channel Model Channel output signal model: T
xn = h sn + νn , def
sn = [sn sn−1 sn−2 . . . sn−Lh +1 ]T is the system’s input vector-stream at time n = 1, . . . , N, sn denotes the sampled source signal,
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.3/34
BD: Channel Model Channel output signal model: T
xn = h sn + νn , def
sn = [sn sn−1 sn−2 . . . sn−Lh +1 ]T is the system’s input vector-stream at time n = 1, . . . , N, sn denotes the sampled source signal, νn represents a zero-mean white measurement disturbance independent of the source signal,
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.3/34
BD: Channel Model Channel output signal model: T
xn = h sn + νn , def
sn = [sn sn−1 sn−2 . . . sn−Lh +1 ]T is the system’s input vector-stream at time n = 1, . . . , N, sn denotes the sampled source signal, νn represents a zero-mean white measurement disturbance independent of the source signal, Lh denotes the length of system impulse response h.
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.3/34
BD: Filter model FIR filter output signal mode: zm,n = wTm xn , wm = [w0 w2 w3 . . . wLw −1 ]T denotes the filter’s impulse response at learning-iteration m = 1, . . . , M,
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.4/34
BD: Filter model FIR filter output signal mode: zm,n = wTm xn , wm = [w0 w2 w3 . . . wLw −1 ]T denotes the filter’s impulse response at learning-iteration m = 1, . . . , M, def
xn = [xn xn−1 xn−2 . . . xn−Lw +1 ]T denotes the filter input samples at time n = 1, . . . , N,
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.4/34
BD: Filter model FIR filter output signal mode: zm,n = wTm xn , wm = [w0 w2 w3 . . . wLw −1 ]T denotes the filter’s impulse response at learning-iteration m = 1, . . . , M, def
xn = [xn xn−1 xn−2 . . . xn−Lw +1 ]T denotes the filter input samples at time n = 1, . . . , N, Lw denotes the length of the filter impulse response.
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.4/34
BD: Channel-Filter Model Channel-filter cascade output model: zm,n = cm sn−δm + Nm,n , where: cm denotes instantaneous amplitude distortion,
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.5/34
BD: Channel-Filter Model Channel-filter cascade output model: zm,n = cm sn−δm + Nm,n , where: cm denotes instantaneous amplitude distortion, δm instantaneous group delay,
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.5/34
BD: Channel-Filter Model Channel-filter cascade output model: zm,n = cm sn−δm + Nm,n , where: cm denotes instantaneous amplitude distortion, δm instantaneous group delay, N m,n denotes so-termed deconvolution noise:
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.5/34
BD: Channel-Filter Model Channel-filter cascade output model: zm,n = cm sn−δm + Nm,n , where: cm denotes instantaneous amplitude distortion, δm instantaneous group delay, N m,n denotes so-termed deconvolution noise: zero-mean, white, Gaussian random process,
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.5/34
BD: Channel-Filter Model Channel-filter cascade output model: zm,n = cm sn−δm + Nm,n , where: cm denotes instantaneous amplitude distortion, δm instantaneous group delay, N m,n denotes so-termed deconvolution noise: zero-mean, white, Gaussian random process, incorrelated with the source signal. Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.5/34
BD: Basic Hypotheses Channel’s impulse response satisfies hT h = 1 and its inverse has finite energy.
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.6/34
BD: Basic Hypotheses Channel’s impulse response satisfies hT h = 1 and its inverse has finite energy. Channel is time-invariant or slowly time-varying.
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.6/34
BD: Basic Hypotheses Channel’s impulse response satisfies hT h = 1 and its inverse has finite energy. Channel is time-invariant or slowly time-varying. Source stream sn is a stationary, ergodic, independent identically distributed (IID) random process with mean IE s [sn ] = 0 and variance IE s [s2n ] = 1.
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.6/34
BD: Basic Hypotheses Channel’s impulse response satisfies hT h = 1 and its inverse has finite energy. Channel is time-invariant or slowly time-varying. Source stream sn is a stationary, ergodic, independent identically distributed (IID) random process with mean IE s [sn ] = 0 and variance IE s [s2n ] = 1. The probability density function p s(s) of the source signal is symmetric around zero and non-Gaussian.
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.6/34
BD: Applications Equalization of communication channels.
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.7/34
BD: Applications Equalization of communication channels. Optomagnetic memory-support storage and retrieval enhancement.
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.7/34
BD: Applications Equalization of communication channels. Optomagnetic memory-support storage and retrieval enhancement. Image deblurring.
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.7/34
BD: Applications Equalization of communication channels. Optomagnetic memory-support storage and retrieval enhancement. Image deblurring. Geophysical measurements analysis.
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.7/34
BD: Bussgang filtering The model reveals that the relationship between zm,n and cm sn−δm is deterministic but for the deconvolution noise.
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.8/34
BD: Bussgang filtering The model reveals that the relationship between zm,n and cm sn−δm is deterministic but for the deconvolution noise.
⇓
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.8/34
BD: Bussgang filtering The model reveals that the relationship between zm,n and cm sn−δm is deterministic but for the deconvolution noise.
⇓
An estimator of the source sequence having form B(zm,n ) can be designed according to Bayesian estimation theory.
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.8/34
BD: Filter Structure -
Delay
-
Delay
···
-
-
Delay
xn ··· w0,m
j z
? P 9
wLw −1,m
zm,n ?
ˆ B(z) cm sˆn−δm ?
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.9/34
A ‘Virtuous Cycle’ -
Inverse filter updating
Works better
Works ner
Source stream estimation
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.10/34
BD as Optimization Problem On the basis of the available Bayesian estimator, the error criterion may be minimized: def 1
h i 1 2 2 C(wm ) = IENm,n [Nm,n ] = IEzm,n zm,n − B(zm,n ) . 2 2
Thanks to ergodicity, the ensemble average IE[·] is estimated by: N X 1 Φ(zm,n ) , IEzm,n [Φ(zm,n )] ≈ N n=1
for any vector-valued function Φ : R → R p . Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.11/34
BD: Automatic Gain Control For practical reasons, it is customary to set the energy-contraint: w20 + w22 + w23 + · · · + w2Lw −1 = 1 . Namely, the filter’s impulse response should belong – at any time – to the unit hyper-sphere: S
p−1 def
= {v ∈ R p |vT v = 1} .
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.12/34
Geometry of S
p−1
At every point v ∈ S p−1 , the tangent space has structure: TvS
p−1 def
= {u ∈ R p |uT v = 0} .
If S p−1 ֒→ R p , which is equipped with the standard Euclidean metric, then the normal space has structure: p−1 def Nv S = {λv|λ ∈ R} .
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.13/34
Riemannian Gradient on S
p−1
Riemannian gradient of a smooth function f : S p−1 → R S p−1 is a vector ∇v f that satisfies: Tangency:
S p−1 ∇v
f ∈ T v S p−1 ,
Compatibility: h∇Sv With the above setting: S p−1 ∇v
p−1
f, uiv =
∂ f T ∂v
u, ∀u ∈ T v S p−1 .
∂f f = (I p − vv ) . ∂v T
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.14/34
Geodesics on S
p−1
A geodesic v(t) = G(t, v0 , g) is a curve on which a particle, amanating from v0 with velocity g, slides with constant scalar speed k gk. p−1 ¨ v ∈ N S , v p−1 v(0) = v ∈ S , 0 v˙ (0) = g ∈ T v S p−1 . 0 The solution is:
g G(t, v0 , g) = cos(k gkt)v0 + sin(k gkt) . k gk Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.15/34
S p−1 ∇v -based
Optimization
As an optimization law for searching for the minimum (or local minima) of a regular function f : S p−1 → R over S p−1 , we may use the Riemannian-gradient based rule: ( dv S p−1 f, dt = −∇v v(0) = v0 ∈ S p−1 .
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.16/34
BD: Geodesic-based Rule The general-purpose differential equation may be customized as: dw T ∂C(w) = −(I p − ww ) , dt ∂w with p = Lw and: ∂C(w) ∂w = IE x [γ(z)x] , γ(z)def = (B(z) − z)(B′ (z) − 1) .
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.17/34
BD: Geodesic-based Algorithm It is suggested to approximate the exact flow of the differential equation on a manifold via piece-wise geodesic arcs: p−1 wm = G tm , wm−1 , −∇Swm−1 C(w) , m ∈ {1, . . . , M} . where: tm denotes an appropriate sequence of adaptation stepsizes,
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.18/34
BD: Geodesic-based Algorithm It is suggested to approximate the exact flow of the differential equation on a manifold via piece-wise geodesic arcs: p−1 wm = G tm , wm−1 , −∇Swm−1 C(w) , m ∈ {1, . . . , M} . where: tm denotes an appropriate sequence of adaptation stepsizes, w0 ∈ S p−1 is arbitrarily selected.
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.18/34
BD: Geodesic-based Algorithm It is suggested to approximate the exact flow of the differential equation on a manifold via piece-wise geodesic arcs: p−1 wm = G tm , wm−1 , −∇Swm−1 C(w) , m ∈ {1, . . . , M} . where: tm denotes an appropriate sequence of adaptation stepsizes, w0 ∈ S p−1 is arbitrarily selected.
Up to numerical error, wm ∈ S p−1 at every iteration. Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.18/34
BD: Projection-based algorithm By the embedding S p−1 ֒→ R p , updates along the Euclidean gradient direction: ! ∂C(w) def v . wm = Π wm−1 − tm , Π (v) = √ ∂w w=wm−1 vT v where: tm denotes an appropriate sequence of adaptation stepsizes,
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.19/34
BD: Projection-based algorithm By the embedding S p−1 ֒→ R p , updates along the Euclidean gradient direction: ! ∂C(w) def v . wm = Π wm−1 − tm , Π (v) = √ ∂w w=wm−1 vT v where: tm denotes an appropriate sequence of adaptation stepsizes, w0 ∈ S p−1 is arbitrarily selected,
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.19/34
BD: Projection-based algorithm By the embedding S p−1 ֒→ R p , updates along the Euclidean gradient direction: ! ∂C(w) def v . wm = Π wm−1 − tm , Π (v) = √ ∂w w=wm−1 vT v where: tm denotes an appropriate sequence of adaptation stepsizes, w0 ∈ S p−1 is arbitrarily selected,
Π : R p → S p−1 is the selected back-projector. Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.19/34
Short Geodesic Arcs If the time to within the geodesic is extended is short enough, the geodesic-based algorithm traces the Riemannian-gradient flow. In fact, for t small enough, the S p−1 -geodesic may be approximated as: ! 2 2 k gk t G(t, v0 , g) ≈ 1 − v0 + gt , 2 which gives rise to the expression: wm − wm−1 ≈− t
S p−1 k∇wm−1 C(w)k2 t
2
wm−1 −
S p−1 ∇wm−1 C(w)
.
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.20/34
Pre-whitening The source stream sn is IID. After passing through the channel, the samples gain second-order statistical correlation. Second-order correlation is easy to remove by data pre-whitening. Let us define: def
R xx = IE xn [xn xTn ] . Whitened filter-input vector-stream: − 12 xˆ n = R xx xn def
.
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.21/34
BD Algorithms at a Glance Collect the filter-input stream and build-up the multivariate stream xn .
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.22/34
BD Algorithms at a Glance Collect the filter-input stream and build-up the multivariate stream xn . Whiten the multivariate signal xn .
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.22/34
BD Algorithms at a Glance Collect the filter-input stream and build-up the multivariate stream xn . Whiten the multivariate signal xn . Choose a starting point for the inverse filter impulse response w0 and learning parameters.
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.22/34
BD Algorithms at a Glance Collect the filter-input stream and build-up the multivariate stream xn . Whiten the multivariate signal xn . Choose a starting point for the inverse filter impulse response w0 and learning parameters. Compute the final inverse filter impulse response w M by the geodesic-based algorithm or the projection-based algorithm applied to the whitened input stream.
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.22/34
Figures of Performance Residual inter-symbol interference (ISI): def
ISIm =
2 TmT Tm − T m,max 2 T m,max
,
def
where Tm = h ⊗ wm and T m,max denotes the component of Tm having maximal absolute value.
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.23/34
Figures of Performance Residual inter-symbol interference (ISI): def
ISIm =
2 TmT Tm − T m,max 2 T m,max
,
def
where Tm = h ⊗ wm and T m,max denotes the component of Tm having maximal absolute value. Elapsed run-time on a 1.86GHz – 512MB platform.
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.23/34
Figures of Performance Residual inter-symbol interference (ISI): def
ISIm =
2 TmT Tm − T m,max 2 T m,max
,
def
where Tm = h ⊗ wm and T m,max denotes the component of Tm having maximal absolute value. Elapsed run-time on a 1.86GHz – 512MB platform. c Flops (counted by Matlab 5.3).
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.23/34
Source and Bayesian Estimator It is assumed that sn is a white random signal, √ √ uniformly distributed within [− 3, + 3], counting N = 5, 000 samples. In this case, a suitable Bayesian estimator is ˆ = κ tanh(λz). B(z) Parameters κ and λ may be pre-learnt on the basis, e.g., of the procedure introduced in S. F IORI, Analysis of modified ‘Bussgang’ algorithms (MBA) for channel equalization, IEEE Trans. on Circuits and Systems - Part I, Vol. 51, No. 8, pp. 1552 – 1560, August 2004.
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.24/34
Experiments on a Toy Channel The channel’s impulse response is h = [1] (Lh = 1) and the base manifold is S 2 (Lw = 3). In this experiment, the channel-filter-cascade impulse response Tm = h ⊗ wm = wm . If we let the learning trajectories depart from randomly generated w0 ∈ S 2 , they should eventually converge to one of the six attractors [±1 0 0]T , [0 ± 1 0]T or [0 0 ± 1]T .
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.25/34
Toy Channel – Geodesic Numerical results on 100 independent trials, M = 100 learning iterations per trial, learning stepsize 0.5.
Neuron weight w
3,m
1
0.5
0
−0.5
−1 1 0.5
1 0.5
0 0
−0.5 Neuron weight w
2,m
−0.5 −1
−1 Neuron weight w
1,m
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.26/34
Toy Channel – Projection Numerical results on 100 independent trials, M = 100 learning iterations per trial, learning stepsize 0.9.
Neuron weight w
3,m
1
0.5
0
−0.5
−1 1 0.5
1 0.5
0 0
−0.5 Neuron weight w
2,m
−0.5 −1
−1 Neuron weight w
1,m
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.27/34
Experiments on Telephonic Channel Sampled telephonic channel having duration Lh = 14. Zeros of BGR channel
Channel amplitude frequency response 2
1.5 1
|H(e )| (dB)
0.5
0
jω
0 −0.5 −1
−1 −2
−1.5 −2
−1
0 Real Part
1
−3 −4
2
Approximate inverse BGR channel
−2
0 ω
2
4
Channel phase frequency response
1
10 0
arg[H(e )] (rad)
0.5
−10
jω
Imaginary Part
1
0
−20 −30
−0.5
0
10
20
30
−40 −4
−2
0 ω
2
4
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.28/34
Experiments on BGR: Data Filter of length Lw = 14.
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.29/34
Experiments on BGR: Data Filter of length Lw = 14. w0 = [0 0 0 0 0 0 1 0 0 0 0 0 0 0]T .
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.29/34
Experiments on BGR: Data Filter of length Lw = 14. w0 = [0 0 0 0 0 0 1 0 0 0 0 0 0 0]T . Noiseless channel (i.e., with νn ≡ 0 identically).
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.29/34
Experiments on BGR: Data Filter of length Lw = 14. w0 = [0 0 0 0 0 0 1 0 0 0 0 0 0 0]T . Noiseless channel (i.e., with νn ≡ 0 identically). Learning stepsize: 1 for the geodesic-based algorithm and 0.9 for the projection-based algorithm.
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.29/34
Experiments on BGR: Data Filter of length Lw = 14. w0 = [0 0 0 0 0 0 1 0 0 0 0 0 0 0]T . Noiseless channel (i.e., with νn ≡ 0 identically). Learning stepsize: 1 for the geodesic-based algorithm and 0.9 for the projection-based algorithm. Learning iterations: M = 80 for both algorithms.
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.29/34
Experiments on BGR: Results 0
−12
Cost function C(wm)
m
ISI (dB)
−5 −10 −15 −20
−12.5
−13
−13.5
−25 0
20
40 Iterations
60
−14
80
1
1.2
0.8
1
Convolution TM=h*wM
Filter impulse response wM
−30
0.6 0.4 0.2 0 −0.2 −0.4
0
20
40 Iterations
60
80
0.8 0.6 0.4 0.2 0
1
2
3
4 5 6 7 8 9 10 11 12 13 Discrete−time index n
−0.2
0
5
10 15 20 Discrete−time index n
25
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.30/34
Numerical Complexity Comparison Algorithms were run on the same batch of 5, 000 channel output samples.
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.31/34
Numerical Complexity Comparison Algorithms were run on the same batch of 5, 000 channel output samples. learning iterations: M = 50.
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.31/34
Numerical Complexity Comparison Algorithms were run on the same batch of 5, 000 channel output samples. learning iterations: M = 50. The flops count refers to the number of floating point operations required by the implemented code to run, averaged over the total number of samples passing by (5, 000 × 50).
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.31/34
Numerical Complexity Comparison Algorithms were run on the same batch of 5, 000 channel output samples. learning iterations: M = 50. The flops count refers to the number of floating point operations required by the implemented code to run, averaged over the total number of samples passing by (5, 000 × 50). The time count refers to the total time required by each algorithm to run on the specified platform.
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.31/34
Complexity Comparison: Results Results of computational-complexity comparison of the geodesic-based algorithm and the projection-based algorithm.
A LGORITHM ISI (dB) Flops Time (sec.s) Geodesic-based −25.057 80.594 0.328 Projection-based −25.056 81.582 0.313
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.32/34
Summary Both algorithms are well-behaving.
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.33/34
Summary Both algorithms are well-behaving. The deconvolution performances are comparable for the two algorithms.
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.33/34
Summary Both algorithms are well-behaving. The deconvolution performances are comparable for the two algorithms. The geodesic-based algorithm may exhibit steadier convergence.
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.33/34
Summary Both algorithms are well-behaving. The deconvolution performances are comparable for the two algorithms. The geodesic-based algorithm may exhibit steadier convergence. The projection-based algorithm may be slightly lighter from a computational point of view.
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.33/34
Many thanks to... The organizers and E.T. Jaynes Foundation!
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.34/34
Many thanks to... The organizers and E.T. Jaynes Foundation! Everybody for the kind attention!
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.34/34
Many thanks to... The organizers and E.T. Jaynes Foundation! Everybody for the kind attention! The Italian team for winning the World Cup!!!!
Extrinsic Geometrical Methods for Neural Blind Deconvolution – p.34/34