neural network architectures for large-scale 3d...

48
Neural Network Architectures for Large-scale 3d Reconstruction Viren Jain Research & Machine Intelligence, Google Inc.

Upload: others

Post on 30-May-2020

6 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

Neural Network Architectures for Large-scale 3d Reconstruction

Viren Jain Research & Machine Intelligence, Google Inc.

Page 2: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

• reconstructing neural wiring diagrams from 3d, nanoscale images of brain tissue

Page 3: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

“connectomics”

Page 4: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

“connectomics”

Page 5: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

Which neurons form synaptic connections with each other?

‣ Associate each dendrite and axon with a parent cell body

‣ Find synapses (connections) between particular cells

Page 6: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

• Global architecture: pathways, projections- fMRI, DTI, tracer injections

• Local circuitry: detailed wiring diagrams- electron microscopy- molecular genetic fluorescent labeling

Page 7: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

1970’s: c. elegans300 neurons; printed photos; MRC in the UK

2013: retina, drosophila medulla1000 neurons; O(teravoxel); Max Planck Institute, MIT, HHMI

In progress: mouse cortical column, whole drosophila brain, song-bird100,000 neurons; O(petavoxel); Harvard, HHMI, Allen Institute, MPI

Machines now being designed & built for whole mouse brain100,000,000 neurons; O(100-1,000 petavoxels)

Page 8: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction
Page 9: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

Synapse Identification

Page 10: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

• Conventional image processing techniques and even modern ‘off-the-shelf’ machine learning techniques make too many mistakes

Page 11: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

.

pres

ynap

tic

postsynaptic

Page 12: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

pres

ynap

tic

postsynaptic

Page 13: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

pres

ynap

tic

postsynaptic

X

Page 14: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

• How to increase accuracy?

Page 15: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

• Image may contain trillions of voxels

• Linear time methods

• Need to decompose the problem

Image Segmentation for Connectomics

Page 16: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

Global Region Formation

Local Boundary Prediction

Page 17: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

“Simple” Global Region Formation

“Sophisticated” Local Boundary Prediction

Page 18: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

input image local boundaries nonlocal segmentation

Page 19: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

Multi-layer neural network architecture that can be trained by gradient learning to interpret images in a desired way.

Convolutional Networks

Page 20: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

• No prior research on best feature representations for EM data

• Not afraid to gather lots of training data

Why Convolutional Networks?

Page 21: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

LeNet-5 Digit/Letter Recognition Network

Yann LeCun and others Bell Labs (1980’s)

Page 22: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

Jain et al, ICCV 2007Turaga et al, NC 2010

Jain et al, CVPR 2010Turaga et al, NIPS 2009

Page 23: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

Helmstaedter et al, Nature 2013

“Connectomic reconstruction of the IPL in mouse retina”

Page 24: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

CLEAN NOISY PSNR=14.96 CN2 PSNR=24.25

BLS-GSM PSNR=23.78 FoE PSNR=23.02

CLEAN CN2

FoEBLS-GSM

Jain & Seung, NIPS 2008

test set performance

Natural Image Denoising with Convolutional Networks

Page 25: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction
Page 26: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

convolutional nets are related to CRF’s

CRF model:

mean field:

inference:

CN2 (equivalent convolutional network):

image restoration

p(x|y) � e(�[ 12

Pi xi(w�x)i+

Pi xi(yi+b)])

µi = tanh (� [(w � µ)i + yi + b])

µi(t + 1) = tanh (� [(w � µ(t))i + yi + b])

Page 27: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction
Page 28: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

• Takes a long time to learn the parameters (even on a GPU)

• Computations involved in backpropagation make (further) parallelization difficult

Page 29: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

• Deep and Wide Multiscale Recursive Networks

Page 30: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

• Model Statistical Structure in Label-Space

Page 31: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

(3) Model Statistical Structure in Label-Space

Page 32: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

• Labels for neighboring image locations are correlated (i.e., structured prediction).

• In supervised image labeling approaches, predictions for neighboring locations typically only correlated through dependence on overlapping input image patches.

(3) Model Statistical Structure in Label-Space

Page 33: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

• (Bayesian) MRF approach: model p(Y) and P(X|Y)

• Computational expense of probabilistic learning & inference can limit model complexity of MRFs in important ways (Jain 2007, 2009).

(3) Model Statistical Structure in Label-Space

Page 34: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

input

output

x1

y1

z1

x2

y2

z2

x3

y3

z3

xN

yN

zN

x

y

z

x

y

z

x

y

z

x

y

zN

x

y

z

x

y

z

x

y

z

x

y

z

mlp_1

mlp_2

unsupervised

supervised

node = feature map arrow = filter

Page 35: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

input

output

x

y

z

x

y

z

x

y

z

xN

yN

zN

x

y

z

x

y

z

x

y

z

x

y

z

x

y

z

x

y

z

x

y

z

x

y

z

mlp_1

mlp_2

unsupervised

supervised

Page 36: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

DAWMRIteration 1

DAWMRIteration 3

Original Scale Downsampled FurtherDownsampled In

put C

hann

els

(3d

imag

e +

4d a

ffini

ty g

raph

)

. . .

1 2 1000. . .

. . .

1 2 1000. . .

. . .

1 2 1000. . .

. . .

Unsu

perv

ised

Feat

ure

Extra

ctio

n (fi

lterin

g,

enco

ding

, poo

ling,

su

bsam

plin

g). . . . . .

1 2 200. . .

x y z

Supe

rvise

dCl

assifi

catio

n(M

LP o

r SVM

)

Input Image Affinity Graph Affinity Graph Affinity Graph

Architecture of Single Iteration

DAWMRIteration 2

DAWMR Architecture

Huang & Jain, ICLR 2014

Page 37: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

Unsupervised Feature Learning Feature Extraction Supervised Learning

Trai

ning

Pipe

line

Impl

emen

tatio

nAr

chite

ctur

e

. . .

. . .

. .

.

Storage( HDD or

flash memory)

. .

.

Multicore CPU System

Thousands of CPU cores

GPU

• distributed CPU feature extraction takes advantage of parallel file system for accessing raw image data of training examples spread across large image space

• distributed CPU feature extraction stage allows expensive operations (sparse coding, multiscale downsampling, etc) without having to optimize implementation for GPUs

• GPU stage focused on simple and relatively “stable” multilayer perceptron computations

Page 38: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

image affinity graphsegmented

affinity graph

DAWMR iteration 1

DAWMR iteration 2

DAWMR iteration 3

recursive processing

Page 39: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

R(S, T ) =1�n2

�X

i,j 6=i

�(Si = Sj,Ti = Tj) + �(Si 6= Sj , Ti 6= Tj)Rand Index:

Huang & Jain, In Review

Page 40: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

Would be useful to have a model that could provide some measure of p(segmentation | image).

Could be used to identify spurious locations in the reconstruction or evaluate specific manipulations (e.g., agglomeration).

Shape Modeling with Neural Networks

negative positive

Page 41: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

Energy-based model for segmentation

Page 42: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

Combinatorial Energy Learning for Segmentation

binary shape descriptors for efficient representation of 3-D shape configurations

deep neural network architecture combines convolutional image network with fully-connected network that scores compatibility between a segmentation and 3-D image

a training procedure uses pairwise object relations within a segmentation to learn the energy-based model

connectivity region data structure for efficiently computing the energy of configurations of 3-D objects

Page 43: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

CELIS: Local energy model

Page 44: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

Combinatorial Agglomeration

Page 45: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

CELIS: Energy-based model

E(S, I) = compatibility between raw data and segmentation (learned)

minimize E by searching over the space of possible segmentations

Combinatorial Energy Learning for Image Segmentation, Submitted

Page 46: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

DAWMRIteration 1

DAWMRIteration 3

Original Scale Downsampled FurtherDownsampled In

put C

hann

els

(3d

imag

e +

4d a

ffini

ty g

raph

)

. . .

1 2 1000. . .

. . .

1 2 1000. . .

. . .

1 2 1000. . .

. . .

Unsu

perv

ised

Feat

ure

Extra

ctio

n (fi

lterin

g,

enco

ding

, poo

ling,

su

bsam

plin

g). . . . . .

1 2 200. . .

x y z

Supe

rvise

dCl

assifi

catio

n(M

LP o

r SVM

)

Input Image Affinity Graph Affinity Graph Affinity Graph

Architecture of Single Iteration

DAWMRIteration 2

DAWMR Architecture

Huang & Jain, ICLR 2014

Page 47: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

End-to-End Learning

input imagenonlocal

segmentation

Page 48: Neural Network Architectures for Large-scale 3d Reconstructionpocv16.eecs.berkeley.edu/talks/jain.pdf · 2016-07-06 · Neural Network Architectures for Large-scale 3d Reconstruction

Google: Samantha Ainsley Tim Blakely Tom Dean Michal Januszewski Peter Li Larry Lindsey Jeremy Maitin-Shepard Art Pope Mike Tyka

MPI: Winfried Denk Jorgen Kornfeld Fabian Svara

Janelia Bock Lab: Davi Bock Eric Perlman

Harvard/AIBS: Nuno da Costa Wei-Chung Lee Clay Reid

Janelia FlyEM: Bill Katz Stephen Plaza Pat Rivlin Proofreading Team