DEEPLY ACTIVE LEARNING:
Approximating Human Learning with Smaller Datasets Combined with Human Assistance
TO P 1 0 W O R L D’ S M O S T I N N OVAT I V E
C O M PA N I E S I N DATA S C I E N C E
S E P 2 6 , 2 0 1 6 @ A R I M O I N C2
1. Motivation
2. Deep Active Learning Approach
3. Experiment Results
4. Lessons Learned
Agenda
S E P 2 6 , 2 0 1 6 @ A R I M O I N C3
0%
5%
10%
15%
20%
25%
30%
2010 2011 2012 2013 2014 2015 TodayAlexNet ZF VGG ResNet GoogLeNet-v4
Classification Error
DEEP LEARNING REQUIRES
HUGE Amount of Labeled Data + Computations
H U M A N P E R F O R M A N C E
G P U - B A S E D D N N S
50K people worked on Amazon Mechanical Turk 2007–2009
S E P 2 6 , 2 0 1 6 @ A R I M O I N C4
Active LearningU N L A B E L E D DATA L A B E L A S U B S E T
T R A I N A L E A R N E R O N L A B E L E D DATA P I C K T H E B E S T N E X T P O I N T S TO L A B E L
Most uncertain
R E -T R A I N T H E L E A R N E R O N L A B E L E D DATA
S E P 2 6 , 2 0 1 6 @ A R I M O I N C5
Deep Active Learning framework
Learned Embeddings
C l u s t e r - b a s e d Q u e r y o f N ex t S a m p l e s t o L a b e l s
Silhouette(xi)=1-a/b
Samples w/ Lowest Silhouette Scores
L a b e l i n g
xi avg
avg
b=min
a=avg
L a b e l e d D a t a
S E P 2 6 , 2 0 1 6 @ A R I M O I N C7
Deep Active Learning framework
Learned Embeddings
C l u s t e r - b a s e d Q u e r y o f N ex t S a m p l e s t o L a b e l s
Silhouette(xi)=1-a/b
Samples w/ Lowest Silhouette Scores
L a b e l i n g
xi avg
avg
b=min
a=avg
L a b e l e d D a t a
S E P 2 6 , 2 0 1 6 @ A R I M O I N C8
LSTM-based Caption Generator
S O F T M A X S O F T M A X S O F T M A X
y1 y2 yn
LST
M
LST
M
LST
M
LST
MW O R D 2 V E C
x0 x1 xn-1v1
R N N / L S T M D e c o d e r
C N N E n c o d e r
Impress guests with lemon pattern fresh look stoneware ensures lasting use dishwasher safe
E m b e d d i n g Ve c t o r
EXPERIMENT
S E P 2 6 , 2 0 1 6 @ A R I M O I N C1 0
The Setup
Train Data Test Data
2000 Product Images 100 Product Images-Caption Pairs
PRODUCT TITLE GENERATOR
Deep Active Learning Random Labeling Fully Supervised
Train on 500 Samples Labeled Actively
Train on 500 Samples Labeled Randomly
Train on 2000 Labeled Samples
PRODUCT TITLE GENERATOR
S E P 2 6 , 2 0 1 6 @ A R I M O I N C
Product Image Reconstructed Image 100 Samples 300 Samples 500 Samples
Hand cream stirs senses refreshing seashore breeze scent moisturize soften skin
Hand cream stirs senses refreshing eucalyptus and mint scent moisturize and soften your skin
Hand cream stirs your senses with a refreshing seashore breeze scent moisturize and soften your skin
You’ll always wake beautiful day cute dishwasher microwave safe mug
Impress guests with lemon pattern fresh look stoneware ensures lasting use dishwasher safe
Impress guests with lemon pattern fresh look stoneware ensures lasting use dishwasher safe
Little girls dress featuring trendy drop waist silhouette crewneck short tulip fully lined knit bodice lends
Soft sophisticated essential bath towel quick dry holds lovely look rich color use benzoyl peroxide
Soft sophisticated hand towel quick dry highly absorbent benzoyl peroxide friendly
1 3
Generated Captions Examples
S E P 2 6 , 2 0 1 6 @ A R I M O I N C1 4
1. Successfully applied active learning in our DL
framework
2. Demonstrated an improvement of 3x in
perplexity score btw DAL vs. DRL
3. It turns out DAL works for text generation &
not just classification!
• It’s effective when your data has
clustering structure
4. Future work: Work on larger scale dataset
Summary & Future Work
DEEPLY ACTIVE LEARNING:
Approximating Human Learning with Smaller Datasets Combined with Human Assistance
TO P 1 0 W O R L D’ S M O S T I N N OVAT I V E
C O M PA N I E S I N DATA S C I E N C E