Confidence Regularized Self-Training

Yang Zou Zhiding Yu Xiaofeng Liu Vijayakumar

Bhagavatula

Jinsong Wang

Source Domain (Labeled) Target Domain (Unlabeled)

Image

classification

Unsupervised Domain Adaptation (UDA)

Semantic

segmentation

Car

Adaptation

Adaptation

Target Images (Cityscapes)

Deep

CNN

Source Labels (GTA5)

Source Images (GTA5)

Predictions (Cityscapes)

Pseudo-labels (Cityscapes)

GTA5 → Cityscapes

Yang Zou, Zhiding Yu et al., Unsupervised Domain Adaptation for Semantic Segmentation via Class-Balanced Self-Training, ECCV18

UDA through Iterative Deep Self-Training

where: x: input sample p: class predication vector y: label vector !": pseudo-label vector

w: network parameters s: source sample index t: target sample index △$%&: probability simplex

Pseudo-label

generation

Network

retraining

Car

Person

Bus

Balanced softmax Pseudo-label !"∗ Network output

after self-training

Class-Balanced Self-Training (CBST)

Yang Zou, Zhiding Yu et al., Unsupervised Domain Adaptation for Semantic Segmentation via Class-Balanced Self-Training, ECCV18

Samples from VisDA-17 (With label “Car”)

Building

Person

Building

Sky

Building

Sample from BDD100K

Green: Correct Red: Misclassified

Preventing Overconfidence in Self-Training

Pseudo-label

Label

regularization

Model

regularization

Network output after

self-training

Pseudo-label

generation

Image label: “Car”

Network output before

self-training

Network

retraining

(b) Label regularized self-training

(c) Model regularized self-training

(a) Self-training without confidence regularization

Backbone

Network

Car

Person

Bus

Network

retraining

Network

retraining

Car

Person

BusPseudo-label

generation

Network

retraining

where: !: regularizer weight

Label Regularized Self-Training (LR)

Balanced softmax Pseudo-label "#∗ Network output

after self-training

Car

Person

BusPseudo-label

generation

Network

retraining

where: !: regularizer weight

Model Regularized Self-Training (MR)

Balanced softmax Pseudo-label "#∗ Network output

after self-training

LR-Entropy (LRENT) MR-KLDiv (MRKLD)

MR-Entropy (MRENT)

MR-L2 (MRL2)

Pseudo-label solver

Regularized retraining loss

v.s. probability

Pseudo-label generation loss

v.s. probability

Proposed Confidence Regularizers

0 0.25 0.5 0.75 1y

0

1

2

3

regula

rize

d lo

ss

LRNET

=0=0.5=1

y*

0 0.25 0.5 0.75 1p

-10

2

4

6

8

10

regula

rize

d lo

ss

MRENT

=0=0.5=1

p*

0 0.25 0.5 0.75 1p

-10

2

4

6

8

10

regula

rize

d lo

ss

MRKLD

=0=0.5=1

p*

0 0.25 0.5 0.75 1p

-10

2

4

6

8

10

regula

rize

d lo

ss

MRL2

=0=0.5=1

p*

Probabilistic Explanation

Proposition 1. CRST can be modeled as a regularized maximum likelihood for classification (RCML) problem

optimized via classification expectation maximization.

Theoretical Analysis

Proposition 2. Given pre-determined !"′s, CRST is convergent with gradient descent for network retraining

optimization.

Proposition 4. Self-training with MRKLD is equivalent to self-training with pseudo-label uniformly smoothed

by $ =(Kα−α)/(K+Kα), where α is the regularizer weight.

Proposition 3. If !" are equal for all k, the soft pseudo-label of LRENT given is exactly the same as softmax

with temperature

Convergence Analysis

Softlabel and Softmax with temperature

Label smoothing

Results on VisDA-17

Results on Office-31

Experiment: UDA for Image Classification

Results on SYNTHIA -> Cityscapes (mIoU* - 13 class)

Results on GTA5 -> Cityscapes

Experiment: UDA for Semantic Segmentation

Original Image Ground Truth Source Model CBST CRST (MRKLD)

Experiment: Qualitative Results (GTA5 -> Cityscapes)

Conclusions

§ Compared with supervised learning, self-training is an under-determined problem (EM with latent variables).

§ Our work shows the importance of confidence regularizations as inductive biases to help under-constrained

problems such as unsupervised domain adaptation and semi-supervised learning.

§ CRST is still aligned with entropy minimization. The proposed confidence regularization only serves as a safety

measure to prevent over self-training/entropy minimization.

§ MR-KLD is most recommended in practice for its efficiency and good performance.

Future Works

§ This work could potentially inspire many other meaningful regularizations/inductive biases for similar problems.

Conclusions and Future Works