The document provides an overview of Weave-D, which is described as a data accumulating, learning and fusing system inspired by the human brain. It supports incremental learning from heterogeneous and multimodal data like video. Weave-D is designed to handle data in chunks, apply previous knowledge to acquire new knowledge, generalize acquired knowledge, and prevent catastrophic forgetting. Potential applications are discussed in domains like medicine, finance, and forensics. The architecture of Weave-D and how it represents knowledge over time are briefly outlined.
3. Introduction
What is Weave-D?
Inspired by human brain
Data Accumulating, Learning and Fusing
System
Supports Multimodal data
Video
Incremental
learning
Inspiration
source
4. Why Weave-D?
Apply previous
knowledge to acquire
new knowledge
Heterogeneous
?
Handle
data
Come as chunks
Prevent
catastrophic
forgetting
Incremental
learning
?
Growth of information
Intuitive
Visualizing
information
Simple
Generalization of
acquired knowledge
?
Conceptualization
?
5. Business Value
Medical
What we can mine?
New patient has a cancer or not?
Effective medicine for certain diseases
Diseases distribution in the country
E.g. Anuradhapura – more kidney diseases
6. Business Value
Finance
Predict customers’ transactional behaviors, so
banks can plan their strategies ahead
Forensics or Police
Predict criminal behavior
Identify crimes with similar
evidence
And many more…
7. Similar Products
IBM Watson
Developed by IBM to compete in Jeopardy
A Question answering system
Consumes “millions” of Wikipedia pages and try
to find answers from the knowledge acquired
Finance and health care domains
12. C3
C1
Child (4-8 years old)
Child (8-12 years old)
Child (1-4 years old)
C4
Forest (Autumn)
Forest (Spring)
Forest (Winter)
C2
City (Day view)
C5 City (Night view)
(None)
Sunset view
Dataset 1
Dataset 2
Sunset view
Dataset 3
13. Demonstration - Scenario
Description
Sam is a sports enthusiast. He has a set of
images belonging to following sports; Croquet,
Polo, Rock-climbing, Sailing, Rowing,
Badminton. Also he has a small description of
the sport for each image. He needs to cluster
these images and text by the sports category.
Constraints
All the photos are not available to him at once.
He gets sets of images each day. (Incremental
learning)
14. User’s Point of View
Input
Query image
Expected outcomes
Set of related images and documents explaining the sport
Tasks
Setting up Weave-D
Training Weave-D
Querying from Weave-D
Sam doesn’t know what sport this is (Query image)
Meaningless file names!
Get documents explaining the sport denoted by image
20. Monetization Plans?
Promotions through Social Media
Facebook
Google+
Advertising on Data Mining websites
KDNuggets
Discussions
ICTA
Private Hospitals
Private Investigation Agencies
National
Hospital
Investments?
Project group
Sri Lanka
Police
21. Few years ahead in Money
Path
Sell 5 units
1 unit = 80K-100K
Part Time
Today
Initial
Investment
(Rs.100,000)
Full Time
January,
2014
1st Release
Advertising
campaign
(Rs. 15,000)
Sell 10 units
1 unit = 150K-200K
January,
2015
January,
2016
2nd Release
Labor cost (4
members)
(Rs. 60,000)
Break even
Other
(Rs. 25,000)
Profitable
22. Glimpse to the Future
Support mining information at different
granularities
Extend Weave-D Client-Server architecture
Support already existing standards (e.g.
PMML)
Data accmulation and fusion system. Seems like an already achieved thing and straightforwardLet me tell you how this is special from other tools out thereAppear Heterogenous (describe)Appear Incremental learning (describe)
Data is no longer homogeneous (it is a combination of images, text, audio)Weave-D supports heterogeneous dataData is no longer available at once, data arrives as streams, at different timesWeave-D can learn incrementallyNot all information is important, user should be able to select which features are importantWeave-D allows user to select important features of data
Would this be better if we present as a 2d flow chart?
Show config filesShow componentsShow as a diagram
Few points about what is the experiment what we’re trying to achieveExploratory mining techniqueVery difficult to measure the qualityBy InspectionCluster PurityShow horizontal not vertical
Sam has an image and he doesn’t know what sport this is. And the images and text files does not have very meaningful filenames. (otherwise he could have guessed the name and found the sport). What Sam can do is, he can query this image from Weave-D and find related images and text both. Then by reading the returned documents, he can figure out the sport.Rename data to have meaningless names!
Few examples explaning the same task Sam querying image and getting text in other domainsEx. Radiologist input and image of a cancer and get the full detailed reports relatedEx. Forensic investigators input an audio clip of a criminal and getting picture of a person as the resultFlexible architectureAllow user to form the architecture!Intuitive UI (Drag & Drop)