visual analytics from feature design to deep neural ...€¦ · visual analytics from feature...

64
Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova

Upload: others

Post on 27-Jun-2020

12 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Visual Analytics from Feature Design to Deep Neural Networks Understanding

Nicola Pezzotti and Anna Vilanova

Page 2: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Traditional Machine Learning

Feature SpaceHigh-Dimensional

Feature VectorsData

Feature

Extraction

It’s a dog

It’s a cat

Classifier (supervised)

Clustering(unsupervised)

Ite

ms(e

.g.,p

ictu

res)

Features (e.g., pixels)

Page 3: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Feature Space Design

Feature SpaceHigh-Dimensional

Feature VectorsData

Feature

Extraction

Ite

ms(e

.g.,p

ictu

res)

Features (e.g., pixels)

• Based on humane knowledge

and reasoning, assumptions

• Automatic Feature Extraction

Page 4: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Training/Learning

Feature

Space

Data

It’s a dog

It’s a cat

Classifier (supervised)

dogs

cats

Page 5: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Black Box

Use of Machine Learning – As a Black Box

Feature

Space

It’s a cat

Classifier (supervised)

It’s a dog

Page 6: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Black Box

Feature

Space

It’s a dog

Classifier (supervised)

It’s a cat

Use of Machine Learning – As a Black Box

Page 7: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Black Box

Feature

Space

It’s a cat

Classifier (supervised)

It’s a dog

Use of Machine Learning – As a Black Box

Page 8: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Black Box

Feature

Space

It’s a cat

Classifier (supervised)

It’s a dog

Use of Machine Learning – As a Black Box

Page 9: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Black Box

Black Box difficult acceptance

Feature

SpaceIt’s benign

Classifier (supervised)

It’s malign

Page 10: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Black Box

Feature

Space

Data

It’s a dog

It’s a cat

Classifier (supervised)

dogs

cats

User in the loop

Page 11: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Visual Analytics

Visualization

InteractionData Analysis

Machine learning

Page 12: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Visual Analytics – Opening the black-box

● Not enough getting a good or bad result you want to reason about the

process and the methods

● Why do I get a good or bad result?

Relation to feature space.

What makes a dog a dog ? what makes a cat a cat?

Important in medical applications:

What features makes a benign tumor benign or malign?

● Training Data

● Evaluate Design Decisions

Page 13: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Feature Space Design

Feature SpaceHigh-Dimensional

Feature VectorsData

Feature

Extraction

Ite

ms(e

.g.,p

ictu

res)

Features (e.g., pixels)

• Based on humane knowledge

and reasoning, assumptions

• Automatic Feature Extraction

Page 14: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Add reference

Why automation is not enough?

Page 15: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Numbers do not tell the whole story

http://blog.revolutionanalytics.com/2017/05/the-datasaurus-dozen.html

Page 16: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Feature Space Visualization/UnderstandingParallel Coordinates

SPLOM

High-Dimensional

Feature Vectors

Ite

ms(e

.g.,p

ictu

res)

Features (e.g., pixels)

Page 17: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Dimensionality Reduction

Dimensionality

Reduction

Embedding

Dim-1

Dim

-2

Assumption: redundancy in the data

Goal: preserve “structure” of the HD space

High-Dimensional

Feature Vectors

Ite

ms(e

.g.,p

ictu

res)

Features (e.g., pixels)

What does “structure” mean?

Page 18: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Dimensionality Reduction (DR)

18

● Build a lower dimension in which distances between points reflect similarities in the

HD data

● Minimize an objective function that measure the discrepancy between similarities in

the data and similarities in the map

Page 19: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Dimensionality Reduction (DR)

19

● Build a lower dimension in which distances between points reflect similarities in the

HD data

● Minimize an objective function that measure the discrepancy between similarities in

the data and similarities in the map

Page 20: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Types of Dimensionality Reduction

Transformation to a projection/subspace (for Vis mainly 2D):

○ Linear:Resulting attributes are linear combination of existing attributes

○ Principal Component Analysis (PCA)

○ Linear Discriminant Analysis (LDA)

○ …

○ Non Linear: Resulting attributes do not have straightforward relation to original attributes

■ Multi-Dimensional Scaling (MDS) - preserve distances

■ t-Distributed Stochastic Neighbor Embedding (t-SNE) – preserve neighbourhods

https://lvdmaaten.github.io/tsne/

■ …

Page 21: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

MNIST dataset – Handwritten numbers

21

Page 22: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

1

0

PCA on MNIST dataset

Page 23: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

tSNE on MNIST dataset

2

3

4

56

7

8

1

0

9

Page 24: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

PCA vs tSNE on MNIST dataset

PCA tSNE

2

3

4

56

7

8

1

0

9

Page 25: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Overview of systems that do feature analysis

Page 26: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

201214.1 M Cancer Cases 8.2 M Cancer Deaths [WHO Report, 2014]

Prostate Cancer1 out of 6 men

Example of Visual Analytics using tSNE

Page 27: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Dose A

Dose C

Dose B

3Renata Raidou

Page 28: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Dose A

Dose C

Dose B

Treatment Conventional vs. tailored to tumor characteristics

R. Raidou et al. EuroVis 2015

Page 29: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Tumor Tissue Characterization

Gleason Scores (GS)

Appearant Diffusion Coefficient (MRI)

Dynamic Contrast Enhanced (MRI)

(Pharmacokinetic modelling)

R. Raidou et al. EuroVis 2015

Page 30: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Tumor Tissue Characterization: Current Visual Analysis

Feng et al. ECR2015

R. Raidou et al. EuroVis 2015

Page 31: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Visual Tool

R. Raidou et al. EuroVis 2015

Page 32: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Identification and Exploration of Intra-tumor Regions

R. Raidou et al. EuroVis 2015

Page 33: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Visual Analysis Tool

R. Raidou et al. EuroVis 2015

Page 34: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Employing Visual Analytics to Aid the Design of White Matter Hyperintensity Classifiers

R.G. Raidou, H. Kuijf, N. Sepasian, N. Pezzotti,

(clinician from UMC), M. Breeuwer, A. Vilanova

Page 35: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Deep Learning

Neuronal NetworkData

It’s a dog

It’s a cat

Page 36: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional
Page 37: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional
Page 38: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

MOTIVATION

Res-Net and VGG

DenseNet

U-Net

Page 39: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Generative Semantic

Interpretable AI

Page 40: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Generative Semantic

Interpretable AI

Page 41: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

From Training to Input Generation

GENERATIVE

Gradient is back-propagated&

parameters are adjusted

Forward Pass

It’s a Cat!It’s a Dog!

Page 42: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

From Training to Input Generation

GENERATIVE

Gradient is back-propagated&

the Input is adjusted

Forward Pass

Show me

a Dog!

Show me

a Dog!

Page 43: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Feature Visualization

GENERATIVE

[Feature Visualization - Olah, Mordvinstev, Schubert - 2017]

Page 44: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

GoogLeNet Architecture

GENERATIVE

[Feature Visualization - Olah, Mordvinstev, Schubert - 2017]

Conv2d0 mixed3a mixed4a mixed4b mixed4e

Page 45: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Problems of Generative Methods

● Difficult to interpret

● Generated through an optimization

● Scalability issuehow many filters can we analyze at the same time?

GENERATIVE

Badly trained network?How informative?

Page 46: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Interpretable AI

Generative Semantic

Page 47: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

● Response of the Model against ALL the input○ Training set

○ Test set

○ Online examples

● Show the semantic relationships learned by the

network

● Scalability of computations is key○ Huge training sets

○ Cognitive overload of users

Visualization of Semantic Relationships

SEMANTIC

Page 48: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Latent Space Sampling

SEMANTIC Forward Pass

Trained Network

20739

[20, 7, 3, 9]

Page 49: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Latent Space Sampling

SEMANTIC Forward Pass

Trained Network

571117

[20, 7, 3, 9] [5, 7, 11, 17]

Page 50: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Good

Model

Bad

Model

SEMANTIC

Page 51: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Good

Model?

SEMANTIC

Page 52: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Hierarchical-SNE for Deep Network Analysis

92 s

ImageNet Test Set (100k)

SEMANTIC

[Hierarchical Stochastic Neighbor Embedding - Pezzotti et al. - 2016]

Page 53: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

SEMANTIC

[Hierarchical Stochastic Neighbor Embedding - Pezzotti et al. - 2016]

Page 54: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

SEMANTIC

[Hierarchical Stochastic Neighbor Embedding - Pezzotti et al. - 2016]

Page 55: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

DeepEyes

SEMANTIC

● A semantic tool to support the design of Deep Neural Networks

Deep Neural Network

TrainingDesign

Trial & Error

Page 56: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

DeepEyes

SEMANTIC

● A semantic tool to support the design of Deep Neural Networks

Deep Neural Network

TrainingDesign

DeepEyes

Informed Decisions

Page 57: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

SEMANTIC

DeepEyes

[DeepEyes: Progressive Visual Analytics for Designing Deep Neural Networks - Pezzotti et al. - 2018]

Page 58: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

SEMANTIC

DeepEyes

[DeepEyes: Progressive Visual Analytics for Designing Deep Neural Networks - Pezzotti et al. - 2018]

Page 59: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

SEMANTIC

DeepEyes

[DeepEyes: Progressive Visual Analytics for Designing Deep Neural Networks - Pezzotti et al. - 2018]

Page 60: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

SEMANTIC

DeepEyes: Detailed Analysis

[DeepEyes: Progressive Visual Analytics for Designing Deep Neural Networks - Pezzotti et al. - 2018]

Activation HeatmapInput Map Filter Map

Page 61: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

SEMANTIC

DeepEyes: Detailed Analysis

[DeepEyes: Progressive Visual Analytics for Designing Deep Neural Networks - Pezzotti et al. - 2018]

Activation HeatmapInput Map Filter Map

Dead Filters

Page 62: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Conclusions

Page 63: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

● Interpretability is important for ML models○ Validation

○ Improvement

● Scalability of analytics solutions is key○ Large datasets

○ Limited human cognitive power

● Interpretability will be even more relevant○ A shift towards unsupervised techniques

○ Active Learning solutions

○ Reinforcement Learning with Humans-in-the-loop

Conclusions

CONCLUSIONS

Page 64: Visual Analytics from Feature Design to Deep Neural ...€¦ · Visual Analytics from Feature Design to Deep Neural Networks Understanding Nicola Pezzotti and Anna Vilanova. Traditional

Thank you!Questions?