Академический Документы
Профессиональный Документы
Культура Документы
System G Team
Feature Extraction
Machine Learning
Classifiers
Learning Models
Cognitive
Reasoning
Risk Prediction
Big Graph
(e.g. Transaction
Graph in the
Financial Domain)
3
Judgement
Perception
Reasoning
Strategy
Observation
Memory
8
4
Observation
Memory
Graph
Database
Graph
Analytics
Graphical
Models
http://systemg.research.ibm.com
Memory
System G Team
Relationship,
Perception &
Contextual Analysis
Machine Learning:
Machine Reasoning:
Graph Middleware:
Deep Learning Tools
Bayesian Networks
Parallel Prog. Lib.
Visual and Text Sentiment Tools
Game Theory Tools
Power Optimization
Anomaly Detection Tools
Multimodal Analysis Platform
GPU Optimization
Mobile Cognition:
Graph Analytics:
iOS Cognition Tools
Topological Analysis
Robot Cognition Tools 3
Matching and Search
Path and Flow
Machine Learning
Technologies
Spatiotemporal Analytics:
Spatiotemporal Mining
Spatiotemporal Indexing
2
Network Analytics
Technologies
Graph Database:
Native Store
GBase
Graph Visualization:
Multivariate Graph
13
Dynamic Graph
Big Graph
10
Sensing &
Observation
4
Machine Reasoning
Technologies
Judgment
Perception &
Representation
Reasoning &
Strategy
Memory
1
Graph Database
Technologies
11
System G Team
Challenges
13
14
: observations
: hidden states
Transmitted images,
speech content, video
content
2013 IBM Corporation
16
17
Single CPU core (2.40GHz) of an Intel Xeon E5-2630 server with 64GB of RAM. An NVIDIA Titan GPU was
used for CNN computations.
18
19
iPad Pro
iPhone 6s
System G Team
~7 frames/sec
21
22
24
Training Videos
Feature Extractions
Low-level feature
SIFT (Visual)
Attempting board trick
STIP (Motion)
Feeding an animal
25
Deep
Learning
Decision
Tree
Fusion
Late
Fusion
Early
Fusion
MFCC (Audio)
SVM
Mid-level Concept
Landing a fish
Classifiers
Output
running
jumping
person
board
park
street
Feature
Extracting
PeopleSplitUp
Classifier
Key frames
PeopleMeet Classifier
PersonRuns Classifier
Detected
Embrace
Event
Merging
Detected
PeopleSplitUp
Postprocessing
Detected PersonRuns
Preliminary Events
27
Detected PeopleMeet
Event Identifying
Backwards Search
Forwards Search
Reasoned
Event
Concepts
28
29
30
31
Detection Results
32
33
2010
2009
2014
34
2013
System G Team
35
-- Forecasting what
will happen based on
our and others
potential actions.
36
UNCLASSIFIED
Acceleration
of Neural
network on
Mobile
Devices
Project Name
38
Outline
Background
Full Network Acceleration and Compression
Kernel Importance Measurement
Algorithmic Performance Evaluation
Computation Complexity Assessment
39
Acceleration CNN
on Local Device
40
Memory
Speed
Model Size
Weights
Mult.s
AlexNet
243MB
61M
725M
VGG-S
393MB
103M
2640M
VGG-16
552MB
138M
15484M
GoogLeNet
51MB
6.9M
1566M
41
iPhone 6S
(Plus)
iPad Air 2
iPhone 7 (Plus)
SoC
A9
A8X
A9X
A10 Fusion
CPU
2x Twister @
1.85 GHz
3x Typhone @
1.5 GHz
2x Twister @ 2.26
GHz
4-core
GPU
PVR GT7600
(6 cluster)
PVR GXA6850
(8 cluster)
RAM (shared
memory)
2GB LDDR4
2GB LDDR3
4GB LDDR4
3GB on Plus?
64-bit
128-bit
128-bit
Max # of threads
per group
512
512
512
42
Compression
(pruning) of CNN
Speeding up CNN
43
Think Different
redundant CNN, but do we really need such a CNN as the starting point?
!
g
in
Slim
N
N
C
44
Good to be Slim
45
Be Slim is Hard
46
features
47
A pre-trained
CNN
Extract CNN
Responses
Measure the
Importance of
Feature
Extractors
Prune Model
Fine-tuning
48
Intractable
tractable
Inconsistent
consistent
A pre-trained
CNN
Extract
Responses of
a High-level
Layer
Measure the
Importance of
Feature
Extractors
Backpropagate the
Importance &
Prune Model
Fine-tuning
Forward Propagation
Input
layers
Response
Response
Response
FC layers
49
Evaluation
Metal
iOS GPU Language
Project Name
53
Multi-threading in encoding
command is supported
*Put identical type of commands together as possible.
54
55
56
57
Computation-intensive layer.
Explore a way to reuse data in
cache to improve the speed.
Lots of implementation trick to
reduce overhead in GPU
computations.
E.g., perform compilation
optimization in your codes.
58
2.0x
2.4x
1.6x
1.6x
59