Unsupervised Deep Learning by Neighbourhood Discovery ICML-2019 - - PowerPoint PPT Presentation

▶
unsupervised deep learning by neighbourhood discovery
SMART_READER_LITE
LIVE PREVIEW

Unsupervised Deep Learning by Neighbourhood Discovery ICML-2019 - - PowerPoint PPT Presentation

Aug 18, 2022 314 likes •540 views

Unsupervised Deep Learning by Neighbourhood Discovery ICML-2019 Jiabo Huang 1 Qi Dong 1 Shaogang Gong 1 Xiatian Zhu 2 1 Queen Mary University of London 2 Vision Semantic Ltd. Related Works & Motivation Related Works Clustering

slide-1
SLIDE 1

Unsupervised Deep Learning by Neighbourhood Discovery

1Queen Mary University of London 2Vision Semantic Ltd.

Jiabo Huang1 Qi Dong1 Shaogang Gong1 Xiatian Zhu2

ICML-2019

slide-2
SLIDE 2

Related Works & Motivation

Ø Motivation Ø Related Works

  • Clustering Analysis: Caron et al., ECCV, 2018
  • Sample (Instance) Specificity Learning: Wu et al., CVPR, 2018
  • Self-supervised Learning: Zhang et al., CVPR, 2017
  • Generative Model: Donahue et al., ICLR, 2016

(a) (a) Clustering analysis:

1/6

slide-3
SLIDE 3

Related Works & Motivation

Ø Motivation Ø Related Works

  • Clustering Analysis: Caron et al., ECCV, 2018
  • Sample (Instance) Specificity Learning: Wu et al., CVPR, 2018
  • Self-supervised Learning: Zhang et al., CVPR, 2017
  • Generative Model: Donahue et al., ICLR, 2016

(a) (a) Clustering analysis: class-consistent boundaries?

1/6

slide-4
SLIDE 4

Related Works & Motivation

Ø Motivation Ø Related Works

  • Clustering Analysis: Caron et al., ECCV, 2018
  • Sample (Instance) Specificity Learning: Wu et al., CVPR, 2018
  • Self-supervised Learning: Zhang et al., CVPR, 2017
  • Generative Model: Donahue et al., ICLR, 2016

(a) (b) (a) Clustering analysis: class-consistent boundaries? (b) Sample specificity learning:

1/6

slide-5
SLIDE 5

Related Works & Motivation

Ø Motivation Ø Related Works

  • Clustering Analysis: Caron et al., ECCV, 2018
  • Sample (Instance) Specificity Learning: Wu et al., CVPR, 2018
  • Self-supervised Learning: Zhang et al., CVPR, 2017
  • Generative Model: Donahue et al., ICLR, 2016

(a) (b) (a) Clustering analysis: class-consistent boundaries? (b) Sample specificity learning: correlation between samples?

1/6

slide-6
SLIDE 6

Related Works & Motivation

Ø Motivation Ø Related Works

  • Clustering Analysis: Caron et al., ECCV, 2018
  • Sample (Instance) Specificity Learning: Wu et al., CVPR, 2018
  • Self-supervised Learning: Zhang et al., CVPR, 2017
  • Generative Model: Donahue et al., ICLR, 2016

(a) (b) (c) (a) Clustering analysis: class-consistent boundaries? (b) Sample specificity learning: correlation between samples? (c) Ours: Anchor Neighbourhood Discovery

1/6

slide-7
SLIDE 7

Related Works & Motivation

Ø Motivation Ø Related Works

  • Clustering Analysis: Caron et al., ECCV, 2018
  • Sample (Instance) Specificity Learning: Wu et al., CVPR, 2018
  • Self-supervised Learning: Zhang et al., CVPR, 2017
  • Generative Model: Donahue et al., ICLR, 2016

(a) (b) (c) (a) Clustering analysis: class-consistent boundaries? (b) Sample specificity learning: correlation between samples? (c) Ours: Anchor Neighbourhood Discovery Training with neighbourhoods of high-confidence only

1/6

slide-8
SLIDE 8

Neighbourhood Discovery & Selection

Without ground-truth labels !-Neareset neighbourhood structure Consistent?

2/6

Consistent?

slide-9
SLIDE 9

Neighbourhood Discovery & Selection

Ø Observation: Consistency v.s. Similarity Distribution Entropy Sample Index High Entropy Similarity Sample Index Low Entropy Similarity Consistency Entropy

3/6

slide-10
SLIDE 10

Neighbourhood Discovery & Selection

Sample Index

Low Entropy

Similarity

!-neareset neighbours

4/6

slide-11
SLIDE 11

Neighbourhood Discovery & Selection

Selection

Class-consistent Neighbourhoods !-neareset neighbours

Sample Index

Low Entropy

Similarity

4/6

slide-12
SLIDE 12

Neighbourhood Discovery & Selection

Selection

Class-consistent Neighbourhoods !-neareset neighbours

Sample Index

Low Entropy

Similarity

4/6

slide-13
SLIDE 13

Neighbourhood Discovery & Selection

Selection

Class-consistent Neighbourhoods !-neareset neighbours

Sample Index

High Entropy

Similarity Sample Index

Low Entropy

Similarity

4/6

slide-14
SLIDE 14

Training Objectives & Strategy

Ø Neighbourhood Supervision Ø Curriculum Learning

5/6

1st Round

slide-15
SLIDE 15

Training Objectives & Strategy

Ø Neighbourhood Supervision

5/6

Ø Curriculum Learning

1st Round 2nd Round

slide-16
SLIDE 16

Training Objectives & Strategy

Ø Neighbourhood Supervision Ø Curriculum Learning

1st Round 2nd Round Last Round

5/6

slide-17
SLIDE 17

Experiments

Ø Small scale Image Classification (!NN) Ø Small scale Image Classification (LC)

CIFAR10 CIFAR100 SVHN CIFAR10 CIFAR100 SVHN

+12.5% +8.8% +6.0% +6.0% +1.7%

DeepCluster ECCV’18 Instance CVPR’18 AND (Ours)

6/6 Accuracy

  • 0.3%
slide-18
SLIDE 18

Experiments

Ø Small scale Image Classification (!NN) Ø Small scale Image Classification (LC)

CIFAR10 CIFAR100 SVHN CIFAR10 CIFAR100 SVHN

Ø Large scale Image Classification

CONV1 CONV2 CONV3 CONV4 CONV5 FC

+5.6%

DeepCluster ECCV’18 Instance CVPR’18 AND (Ours)

6/6

+12.5% +8.8% +6.0% +6.0% +1.7%

ILSVRC2012 Accuracy Accuracy

  • 0.3%
slide-19
SLIDE 19

Experiments

Ø Small scale Image Classification (!NN) Ø Small scale Image Classification (LC)

CIFAR10 CIFAR100 SVHN CIFAR10 CIFAR100 SVHN

Ø Large scale Image Classification Ø Fine-grained Image Classification (!NN)

CUB200 DOGS

CONV1 CONV2 CONV3 CONV4 CONV5 FC

+2.8% +5.3%

DeepCluster ECCV’18 Instance CVPR’18 AND (Ours)

6/6

+5.6% +12.5% +8.8% +6.0% +6.0% +1.7%

ILSVRC2012 Accuracy Accuracy Accuracy

  • 0.3%
slide-20
SLIDE 20

Poster#115 Unsupervised Deep Learning by Neighbourhood Discovery

Thank You!

Code: https://github.com/Raymond-sci/AND