Learning Deep Representation for Imbalanced Classification Chen - - PowerPoint PPT Presentation

▶

Mar 01, 2024 302 likes •392 views

Learning Deep Representation for Imbalanced Classification Chen Huang, Yining Li, Chen Change Loy, Xiaoou Tang The Chinese University of Hong Kong SenseTime Group Limited Motivation Data imbalance in vision classification Wearing Not hat

SLIDE 1

Learning Deep Representation for Imbalanced Classification

Chen Huang, Yining Li, Chen Change Loy, Xiaoou Tang The Chinese University of Hong Kong SenseTime Group Limited

SLIDE 2

Data imbalance in vision classification

Motivation

Wearing hat Not wearing hat … Minority class Majority class

SLIDE 3

Deep embedding: Class-level

cluster- & class-level constraint

Study traditional re-sampling [ICML’03] and cost-sensitive learning [ICDM’03] scheme

Motivation

Class 1 minority Class 2 majority Class 1 minority Class 2 majority … Cluster 1 Cluster j Cluster 1 Cluster 2

Quintuplet embedding Triplet embedding

𝐸 𝑔 𝑦𝑗 , 𝑔 𝑦𝑗

𝑞

< 𝐸(𝑔 𝑦𝑗 , 𝑔(𝑦𝑗

𝑞))

𝐸 𝑔 𝑦𝑗 , 𝑔 𝑦𝑗

𝑞+

< 𝐸(𝑔 𝑦𝑗 , 𝑔(𝑦𝑗

𝑞−)) < 𝐸(𝑔 𝑦𝑗 , 𝑔(𝑦𝑗 𝑞−−)) < 𝐸(𝑔 𝑦𝑗 , 𝑔(𝑦𝑗 𝑜))

SLIDE 4

Large Margin Local Embedding

CNN CNN CNN CNN CNN Triple-header hinge loss Mini- batches Training samples … Embedding Quintuplet Shared parameters

Network architecture
Equal class re-sampling & class costs assignment in batches

SLIDE 5

Large Margin Local Embedding

Training step
Cluster-wise kNN search
Clustering by k-means
Generate quintuplets from

cluster & class membership

Re-sample batches equally

from each class

Forward their quintuplets to

CNN to compute loss

Back-propagation

Feature-based clustering Feature learning/updating Every 5000 iterations

SLIDE 6

Results

Large-scale CelebA face attributes dataset
200K celebrity images, each with 40 attributes
Highly imbalanced: average positive class rate 23%
We adopt a balanced accuracy
𝑢𝑝𝑢𝑏𝑚 𝑏𝑑𝑑𝑣𝑠𝑏𝑑𝑧 =

𝑢𝑞 + 𝑢𝑜 𝑂𝑞 + 𝑂𝑜

𝑐𝑏𝑚𝑏𝑜𝑑𝑓𝑒 𝑏𝑑𝑑𝑣𝑠𝑏𝑑𝑧 =

1 2 𝑢𝑞 𝑂𝑞 + 𝑢𝑜 𝑂𝑜

Total acc. Balanced acc. Triplet-kNN* 83 72 Anet+ 87 80 LMLE-kNN 90 84 *[Schroff et al., CVPR15] +[Liu et al., ICCV15]

SLIDE 7

Results

Relative gains w.r.t. class imbalance

10 20 30 40 10 20 30 40 50

Relative accuracy gain (%) Class imbalance level (%)

Face attribute More imbalanced Over Anet [28] Over PANDA [46] Over Triplet-kNN [33]

SLIDE 8

Take-home message

Learning deep feature embedding for imbalanced data classification
Cluster- and class-level quintuplets can preserve both locality across clusters

and discrimination between classes, irrespective of class imbalance

Large margin classification via fast cluster-wise kNN search