Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
BIG DATA
TALENT AWARDS
Jordi Trill
BigData & Core Tech Business
Development Manager
Oraclewww.oracle.com
BIG DATA
TALENT AWARDS
Marc Torrent
Director
Big Data CoEwww.bigdatabcn.com
@CoEBigData
GANADOR
Categoría I: proyecto final de
máster o postgrado
Proyecto: Describing Socializing, Eating and
Sedentary Lifestyle Patterns from Egocentric
Images
Autor: Pedro Herruzo, Laura Portell y Alberto Soto
Universidad : UB
Advisor: Dr. Petia I. Radeva
Describing Socializing, Eating and Sedentary lifestyle Patterns
from Egocentric Images
Pedro Herruzo - Laura Portell - Alberto SotoMaster in Foundations in Data Science - UB
This is me today
This is me today
This is me today
This is me today
!
Here I was writing the thesis…
This is me tomorrow
This is me tomorrow
I should socialise more!
This is me tomorrow
!
The LAP-dataset• Own dataset.
• 45,297 egocentric images ~50GB
• 55 days tracked: 8-15h/day
• 4 volunteers’ daily activityNarrative Clip
120 photos/hour
Annotation tool• Labeling task (~75h). It’s easy and we created it!
Deep Learning E - S - T
Two lines:
1) Deep Learning (DL)
2) DL + Classical Machine Learning (ML)
E - S - T
How we predict lifestyle patterns?
Deep Learning
*Pre-trained over ImageNet ~14M images
*
Models: Multitask
0 0.8
1 0.9
0 0.3
yy Prediction
0
1
2} arg max
3
5
L = �X
N
7X
i=1
(yilog(yi) + (1� yi)log(1� yi))
Binary cross-entropy:
L = �0log(0.8)� 1log(0.2)� 1log(0.9)� 0log(0.1)� 0log(0.3)� 1log(0.7)
L = �0log(0.8)� 1log(0.2)� 1log(0.9)� 0log(0.1)� 0log(0.3)� 1log(0.7)
High LowImportance:
Visual Explanations for the model decisions: GRAD-CAM
State of the Art: Selvaraju et al. (2016)
Prototype of a Final Product
Prototype of a Final Product
Temporal costs
10 %11 %
37 % 11 %
7 %
5 %5 %
15 %
Labeling Data Github Installations MeetingsReading papers Coding Writing Presentation
total: 804h
Describing Socializing, Eating and Sedentary lifestyle Patterns
from Egocentric Images
Pedro [email protected]
Thanks for listening. Questions?
Master in Foundations in Data Science - UB
Laura [email protected]
Alberto [email protected]
GANADOR
Categoría II: tesis doctoral
Proyecto: Temporal Graph Mining and
Distributed Processing
Autor: Rohit Kumar
Universidad: UPC, ULB
Advisor: Alberto Abelló (UPC) & Toon Calders (ULB)
Temporal Graph Mining and Distributed Processing
Dr. Rohit Kumar
Advisors: Alberto Abelló (UPC) and Toon Calders (ULB)
Motivation
11
www.alstntec.com
www. farnetworks.com
World is getting more and more connected
Human – Human interactions- Social Media- Email- Phone- Chat- Blogs
Device – Device InteractionsHuman – Device Interactions
Graph is a natural model to represent this connectedness
Most common approach
12
www.alstntec.com
www. farnetworks.com
Static network
Problem
Scaling to big graphs!
13
Billions of interactions
Loss of Information
9 AM
8 AM
10 AM10 AM
8 AM
9 AM
Contribution
14
Cycle EnumerationNeighborhood Profile
Efficient Algorithms and Data Structure
New models for using temporal nature of network
Influence Propagation
Cost model for Partitioning of Graph in Spark GraphX
Resulted
• 5 Conference Paper (2 in A*, 2 in A and 1 in B Core Ranking)
• 1 Journal paper in KAIS
• 2 Workshop papers
• 1 Demo Paper in CIKM
• A open source system (iMaxer) for influence spread prediction.
15
16
All papers/code and summary of work @ http://rohit13k.github.io/
BIG DATA
TALENT AWARDS