Scaling face recognition with big data - Bogdan Bocse

@ITCAMPRO #ITCAMP17Community Conference for IT Professionals
Scaling Face Recognition
with Big Data
Bogdan BOCȘE
Solutions Architect & Co-founder VisageCloud
https://VisageCloud.com
https://www.linkedin.com/in/bogdanbocse/
https://twitter.com/bocse

Many thanks to our sponsors & partners!
GOLD
SILVER
PARTNERS
PLATINUM
POWERED BY

• How to learn ?
• What to learn?
• Defining learning objectives
• How to scale learning?
• Gotchas
• VisageCloud
–Architecture
–Use Cases
Agenda

• What questions to ask before writing the code?
• How to look at the data before feeding it to the
machine?
• What is the state of the art regarding ML?
• What frameworks to use?
• What are the common traps to avoid?
• How to design for scale?
Objectives

HOW TO LEARN?

Vision
• Convolutional Neural Networks
• Inception Paper
NLP
• Word2Vec
• GloVe: Global Vectors for Words Representation
Generic
• Classification
• Prediction
How to Learn?

Convolutional Neural Networks: Big Picture

• Pooling / Max Pooling
• Convolution
• Fully Connected Activation
– Activation Function, eg. ReLu
Convolutional Neural Networks : Components

• Learning is an optimization problem
–Find parameters of a system (neural network) that
minimize a fixed error function
–Not unlike planning orbital paths
• Defining the network architecture
• Defining the training algorithm
–Stochastic Gradient Descent
• With momentum
• With noisy
Taking a Step Back: The Math

• DeepLearning4j
– Independent company
– Java interface with C-bindings for performance
• TensorFlow
– Python & C++ API
– Developed by Google
– Compatible with TPU
• Torch
– Developed by Facebook
– Written in LuaJIT, with Python bindings
Frameworks

WHAT TO LEARN?

• Public data sets
–Labelled Faces in the Wild (LFW)
–Youtube faces
–Kaggle
• Private data sets
• Build your own
–Outsourcing: Mechanical Turk
–Crowsourcing: ReCaptcha model
Data Sets

Preparing Data
Clean
data
Cropping
Structure
Homogeneity
Normalization
Histograms
Filtering

• Machine learning is not magic
• If you can’t understand the data, a machine probably
won’t either
• Preprocessing makes the difference between results
• Applying filters, normalization, anomaly detection is
computationally inexpensive
Preparing Data

DEFINING LEARNING OBJECTIVES

• Supervised
–Classification
–Scoring and regression
–Identification
• Unsupervised
–Clustering
Defining learning objectives

• Projecting input onto a fixed set of classes
• “Don’t use a cannon to kill a fly”
–Support Vector Machines
• Linear
• Radial Based Functions
Classification

• Embedding
–Projecting input (image) onto an vector space with a
known property
• Triplet Loss Function
Identification

• Splitting a set of items into non-overlapping subsets,
based on item attributes
• Counting people in video streams
• Algorithms:
–Fixed threshold
–K-means
–Rank-order clustering
Clustering

HOW TO SCALE LEARNING?

• Scaling training
– Requires shared memory space
– Vertical scaling
• GPU
• Soon-to-come: TPU (tensor processing unit)
• Scaling evaluation
– Shared nothing architecture
– Neural network/classifier rarely change
– Load balancing pattern
– Partitioning data if needed
How to scale learning?

• There is no “reduce” for neural networks
• Averaging weights/parameters
– Usually not a good idea
• Genetic algorithms
– Requires a lot of processing power
– Running independent iterations on different machines
– Crossover between weights/parameters of independently
trained neural networks after each epoch
Ideas for horizontal scaling

GOTCHAS

• Our 2D and 3D intuition often fails in high dimensions
• Distances tend to become relatively “the same” as
number of dimensions increases
• Dimensionality reduction
– Embedding functions
– Principal component analysis
The Curse of Dimensionality

• “The bottom of a valley is not necessarily the lowest
point on Earth”
• Learning algorithms may get stuck in local optima
• Using momentum or some random noise reduces
this possibility
• Using genetic algorithms can be even more robust,
but it’s computationally expensive
Local Optima

Visualizing Local Optima
monkey saddle

“Based on state-of-the-art machine learning, our
weather forecast system can predict tomorrow’s
weather with 72% accuracy”
Evaluating of Learning
You get the same results by saying “it’s going to be the same as today”

• Don’t test on the data you train on
– Use different data set
– Split the data sets you have
• Beware of data biases
– Confirmation bias
– Survivorship bias
– Selection bias
• Compare against a benchmark, even a dummy one
– Coin flip
– Linear algorithms
– “Same-as-before”
Evaluation of Learning

Architecture and Use Cases

High Level Architecture
VisageCloud Production
HAProxy
(reverse proxy)
Image Storage
AWS S3
Service
(API Controller)
Cassandra
Containers
(Docker)
Neural Networks
(OpenCV, Dlib,
Torch, pixie magic)
CQL Binary
HTTP
API Consumer
(Customer Infrastructure)
HTTPS
HTTP
HTTPS

Detect
faces
Align faces
Pre-
processing
Feature
extraction
Feature
comparison
Processing Pipeline

• The collection
–Slice of data used together
–10K-100K records
• The Cache-Inside Pattern
–Loading / preloading collection in one application server
–Content based routing/balancing to maximize cache hits
–No logic in the database layer
–Requires periodic polling for updates
• Weaker consistency
Partitioning Data: Application Level Logic

Application Layer
Application Application Application
Cassandra (Database Layer)
Cassandra Node Cassandra Node Cassandra Node Cassandra Node
Content-based balancing/routing
Preload collectionPoll for updatesWrite updates

• Perform comparison logic in database
–User Defined Aggregate Functions
• Removes the need to move data around between
application and database
• Harder to deploy/test
• Stronger consistency

• It’s math, not magic
• If you don’t understand the data, neither will the
machine
• Preprocessing makes the difference
• Test against a benchmark, any benchmark
• Evaluate first, scale later
Key Take-away

Bogdan@VisageCloud.com
+(40) 724 714 234
https://www.linkedin.com/in/bogdanbocse/
https://twitter.com/bocse
Let’s keep in touch

Q & A

Scaling face recognition with big data - Bogdan Bocse

More Related Content

What's hot (20)

Viewers also liked (13)

Similar to Scaling face recognition with big data - Bogdan Bocse (20)

More from ITCamp (20)

Recently uploaded (20)

Scaling face recognition with big data - Bogdan Bocse