SLIDE 1
Bootstrapping Labels for One-Hundred Million Images
Jimmy Whitaker
We are drowning in data
4/5/16 GTC 2016 2
Data Never Sleeps 2.0 - DOMO (2014)
Bootstrapping Labels for One-Hundred Million Images Jimmy Whitaker - - PowerPoint PPT Presentation
Bootstrapping Labels for One-Hundred Million Images Jimmy Whitaker - - PowerPoint PPT Presentation
Bootstrapping Labels for One-Hundred Million Images Jimmy Whitaker We are drowning in data Data Never Sleeps 2.0 - DOMO (2014) 4/5/16 GTC 2016 2 Ripe Opportunities Many problems to solve Limitless amounts of image data Deep
SLIDE 2
SLIDE 3
Ripe Opportunities
- Many problems to solve
- Limitless amounts of image data
- Deep Learning pushing State of the Art
everywhere
- GPUs making everything possible
4/5/16 GTC 2016 3
SLIDE 4
The Problem
- Deep Learning is data-driven
- ImageNet has 1.2 million training examples
- Few large, labeled image datasets exist
- It’s expensive to label data
- Our datasets are +100m images
- Few qualified to label
- Highly sensitive customer data
- Necessary subject matter expertise
4/5/16 GTC 2016 4
SLIDE 5
Ever labeled data?
- Not as easy as it seems:
- It’s repetitive
- Less accurate over time
- One day computers will do it all for you?
- But not yet
- Can some of this effort be automated?
4/5/16 GTC 2016 5
SLIDE 6
Many Approaches
- Mechanical Turk
- Costly
- Time Consuming
- Clustering
- Expensive
- How many clusters
- What features to
use?
4/5/16 GTC 2016 6
- Pre-trained classifiers
- What if pre-trained
classifiers don’t work well on data?
- Active learning
- Iterative labeling
- Open problem
Can we combine these into something useful?
SLIDE 7
The Goal
- Inspired by Image Similarity experience
and Jeremy Howard TED talk
- Use machines to filter the noise
- Reduce repetitive tasks
- Leverage human labeler
- Understand the data
- Label iteratively
- Allow exploration
4/5/16 GTC 2016 7
SLIDE 8
Our Approach
4/5/16 GTC 2016 8
SLIDE 9
Our Approach
4/5/16 GTC 2016 9
Compare Image Hashes to filter Duplicate Images
SLIDE 10
Our Approach
4/5/16 GTC 2016 10
SLIDE 11
Our Approach
4/5/16 GTC 2016 11
SLIDE 12
Our Approach
4/5/16 GTC 2016 12
SLIDE 13
Our Approach
4/5/16 GTC 2016 13
Prevents over- focusing on one portion of feature space
SLIDE 14
Our Approach
4/5/16 GTC 2016 14
SLIDE 15
Our Approach
4/5/16 GTC 2016 15
Label Images on the boundary of the class
SLIDE 16
Our Approach
4/5/16 GTC 2016 16
Improve CNN features for labeled classes
SLIDE 17
GUI
4/5/16 GTC 2016 17
SLIDE 18
Hardware
- Cirrascale GB5670
- 56 CPU Cores
- 8x NVIDIA K-80
- 512GB DDR4
- 1 TB SSD
4/5/16 GTC 2016 18
SLIDE 19
Benefits
- Create Large, Labeled Datasets
- High quality
- Allows data exploration
- Dramatic time reduction
- ~3-5x faster initially
- Multiplicative efficiency gains
- Flexible framework
- Perform data science with images
4/5/16 GTC 2016 19
SLIDE 20
CONFIDENTIAL 20