This presentation gave deep dive into various machine learning and deep learning algorithms followed by an overview of the hardware and software technologies for democratization of AI including OpenPOWER/POWER9 solutions.
2. Objectives
• Introduce the foundations of data science and artificial intelligence
• Give an overview of state-of-the-art machine learning technologies
• Demonstrate the benefits of leveraging H2O DriverlessAI and the IBM
hardware (Power processors coupled with NVLink) for AI projects in
enterprises
3. Agenda
• AI vs machine learning vs deep learning
• Supervised learning vs unsupervised learning
• Training vs inferencing
• Use cases
• Importance of data and data types (structured data, unstructured data)
• Data analysis
• Feature engineering
• Types of machine learning problems (regression, classification etc.)
• Machine learning algorithms
• AI technology landscape
• Role of GPUs and Power systems in AI development
• Best practices in AI development
• Automatic machine learning (AutoML)
5. Machine learning is good for…
1.Complex set of rules impossible to code
2.Long list of rules
3.Adapt to new data
4.Getting insights into large amounts of data
6. Training
• Data intensive:
historical data sets
• Compute intensive:
100% accelerated
• Develop a model for
use on the edge as
inference
Inference
• Enables the computer
to act in real time
• Low Power
• Out at the edge
7. AUTOMOTIVE
Auto sensors
reporting location,
problems
COMMUNICATIONS
Location-based
advertising
CONSUMER PACKAGED GOODS
Sentiment analysis of
what’s hot, problems
$
FINANCIAL SERVICES
Risk & portfolio analysis
New products
EDUCATION & RESEARCH
Experiment sensor analysis
HIGH TECHNOLOGY /
INDUSTRIAL MFG.
Mfg. quality
Warranty analysis
LIFE SCIENCES
Clinical trials
MEDIA/ENTERTAINMENT
Viewers / advertising
effectiveness
ON-LINE SERVICES /
SOCIAL MEDIA
People & career matching
HEALTH CARE
Patient sensors,
monitoring, EHRs
OIL & GAS
Drilling exploration
sensor analysis
RETAIL
Consumer sentiment
TRAVEL &
TRANSPORTATION
Sensor analysis for
optimal traffic flows
UTILITIES
Smart Meter analysis
for network capacity,
LAW ENFORCEMENT
& DEFENSE
Threat analysis - social
media monitoring, photo
analysis
AI Enterprise Use Cases
11. Exploratory Data Analysis
• Trivial but critical to data science process
• Plots
• Time-series data
• Histograms
• Pair-wise scatterplots
• Summary statistics
• Mean, Median, Mode, Maximum, Minimum, Upper and lower quartiles
• Outlier analysis
• Fill missing data
13. (Goodfellow 2016)
Learning Multiple Components
CHAPTER 1. INTRODUCTION
Input
Hand-
designed
program
Output
Input
Hand-
designed
features
Mapping from
features
Output
Input
Features
Mapping from
features
Output
Input
Simple
features
Mapping from
features
Output
Additional
layers of more
abstract
features
Rule-based
systems
Classic
machine
learning Representation
learning
Deep
learning
14. (Goodfellow 2016)
Depth: Repeated Composition
CHAPTER 1. INTRODUCTION
Visible layer
(input pixels)
1st hidden layer
(edges)
2nd hidden layer
(corners and
contours)
3rd hidden layer
(object parts)
CAR PERSON ANIMAL
Output
(object identity)
Figure 1.2: Illustration of a deep learning model. It is difficult for a computer to understand
16. Machine Learning Algorithms:
•Tree Based Methods (Decision Trees, Random Forests, and Gradient Boosting)
•Generalized Linear Models
•Linear Regression
•Logistic Regression
•Support Vector Machines
•Unsupervised Learning Techniques (Clustering, Principal Component Analysis)
•Neural Networks
•Neural Network Topologies
•Convolutional Neural Networks (R-CNN, F-CNN, U-Net for Medical Imaging)
•Sequence Models (RNN, LSTM)
•Autoencoders
•Generative Adversarial Networks
Learning Paradigms:
•Transfer Learning
17. AI Solutions Engineering
• Exploratory Data Analysis, Data Visualization
• Data Engineering
• Data Augmentation
• ETL
• Hyper parameter tuning and search algorithms (Random, Bayesian, and TPE Search)
• Data Leakage
• Bias Detection and Mitigation
• Interpretability (Explainability, Fairness, Accountability, Transparency, Ethics)
• LIME, Anchors, TreeInterpreter, Partial Dependency Plots, Deconvolution etc.
• Inference Optimization (NVidia TensorRT)
18. Elements of Enterprise AI
• Data Security
• Model Deployment and Operationalization
• Inferencing Scenarios (In database, cloud etc.)
• Interoperability issues
• Model Retraining
• Model Maintenance (Versioning, Documentation etc.)
• Regulatory Compliance (HIPAA, SEC, GDPR etc.)
• Model Security (Adversarial Attacks on AI Models)
• Resource Management (Scheduling, Multitenancy etc.)
• Collaboration Tools
19. Why now?
• Data explosion
• GPUs
• Some advancements in machine learning algorithms
20. Technologies for Democratization of Deep Learning
• HPC, Distributed Computing Clusters, Public and Private Clouds
• Multicore-processors (Power9)
• GPUs
• Storage Technologies
• Open Source Deep Learning Frameworks
• TensorFlow, Keras, PyTorch, FastAI, MXNet
• Traditional Machine Learning Frameworks
• Scikit-Learn, H2O, XGBoost, IBM SnapML
• Automatic Machine Learning Frameworks
• TPOT, auto-sklearn, auto-PyTorch, auto-keras
• H2O DriverlessAI, Data Robot, IBM AutoAI
• Data Processing Libraries
• Pandas, Cudf, Dask, Dask-cudf
• Open Source Databases
• MongoDB, Cassandra, EnterpriseDB, MariaDB, Redis, Neo4J
21. Feature
Engineering
HPC Cluster/Public Cloud/Private Cloud/Hybrid Cloud
Distributed Storage (Storage for AI)
Data
Analysis/Engineering/
Warehousing/Mining
Model Development,
Testing & Validation
Deployment &
Inferencing
Retraining, Online Training
& Model Versioning
HPC Schedulers, Cloud Middleware, Kubernetes, HELM, Containers, Virtualization
Databases, Big Data Tools, Pythonic Frameworks, HPC Libraries, Microservices
Cloud Native AI
IoT
22. Applications of Deep Learning
• Computer Vision
• Natural Language Processing
• Speech recognition
• Bioinformatics and Chemistry
• Quantitative Finance
23. Computer Vision Applications of Deep
Learning
• Object Detection
• Face Recognition
• Event Recognition
• Human Pose Estimation
• Motion Tracking
24. (Goodfellow 2016)
Solving Object Recognition
2010 2011 2012 2013 2014 2015
Year
0.00
0.05
0.10
0.15
0.20
0.25
0.30
ILSVRCclassificationerrorrate
Figure 1.12: Since deep networks reached the scale necessary to compete in the ImageNet
Large Scale Visual Recognition Challenge, they have consistently won the competition
25. Some Popular CNNs
• LeNet (1990)
• AlexNet (2012)
• ZF net (2013)
• GoogLeNet (2014)
• VGGNet (2014)
• ResNet (2015)
26. Deep learning Requires Data
Little Data (More hand
engineering)
Lots of Data (Less hand
engineering and simpler
algorithms)
Speech
Recognition
Object
Recognition
Object
Detection
Source Andrew Ng
29. • Level 1
• Parallelism
across nodes
connected via a
network
interface.
• Level 2
• Parallelism
across GPUs
within the same
node connected
via an
interconnect
(e.g. NVLINK).
• Level 3
• Parallelism across the streaming
multiprocessors of the GPU
hardware.
Multi-level Parallelism
29
30. • Shorter training times
• Facilitates distributed machine learning
POWER, NVLink and V100 Advantage
POWER9
GPU
Memory
CPU Memory
150GB/s
150
GB/s
150 GB/s
GPU
Memory
POWER9 AC922 Two cores. Four NVIDIA V100 GPU