Aug 08, 2015
MACHINE LEARNING TRENDS
Machine Learning Meetup, Sofia
OPENING MEETUP
What to expect?
open format
exchange knowledge/ideas
everyone can be on stage
be tolerant, respect the others
COGNITIVE COMPUTING
“A cognitive computer combines
artificial intelligence and machine-
learning algorithms, in an approach which
attempts to reproduce the behavior of the
human brain.”
Wikipedia
COGNITIVE COMPUTING
addresses complex situations that are
characterized by ambiguity and
uncertainty;
handles human kinds of problems
COGNITIVE COMPUTING
“The smart machine era will be the most
disruptive in the history of IT.”
Gartner
COGNITIVE COMPUTING
“By 2018 half of all consumers will interact
with services based on cognitive computing
on a regular basis.”
IDC
COGNITIVE COMPUTING
Why now?
Advances in enabling technology
Increasingly large complex datasets
Emerging Platforms – Cloud, Mobile, Big
Data, Analytics, Social
COGNITIVE COMPUTING
Enabling Technologies
Natural Language Processing
Semantic Analysis
Informational Retrieval
Automated Reasoning
Machine Learning / AI
TRENDS
Computers That Learn
Computers That Think
Computers That Interact with Humans
Computers That Interact with Computers
Research and Use Cases
Education and Training
TRENDS
Siri, Google Now, Cortana
Workplace Disruption
Industry Transformation
Window of Opportunity
ML Practically Means
Algorithms that can learn from and make
predictions on data
Building a model from example inputs in
order to make data-driven predictions or
decisions
ML Broad Categories
Supervised learning
Unsupervised learning
Reinforcement learning
ML Tasks by Desired Output
Classification (typically supervised)
Regression (typically supervised)
Clustering (typically unsupervised)
Density estimation
Dimensionality reduction
ML Approaches
Decision tree learning Artificial neural networks (ANN) Support vector machines (SVM) Clustering Bayesian networks Sparse dictionary learning Genetic algorithms
What We at Imagga Do
Image classification (supervised learning)
Use ANN
More precisely - Deep Learning
Even more precisely - CNN (not the TV
station)
Convolutional Neural Networks (CNN)
Get raster data as input
Typically deep networks - multiple
convolutional and hidden layers
Very useful for images - the convolution
parameters are produced as a result of the
learning
Why NOW
GPUs have thousands of cores
Big amount of data, lots of data sources
Affordable utility computing (e.g. AWS, Azure,
Google Cloud)
Demand for ML solutions
Challenges
Very data greedy
Requires clean data
Requires data variety
Still takes a lot of time (1-4 weeks until the
model converges)
Solutions
Data augmentation (increase robustness)
Auto-cleaning of data (remove outliers and re-
train)
Designing the model architecture for multiple
GPUs
Topic for Next Meetup?
Overview/Presentations of the Bulgarian
companies using ML
Commercial applications and use-cases
Open-source software packages for ML
other . . .
MACHINE LEARNING RESOURCES
IMAGGA blog - www.imagga.com/blog/
ML Flipboard - http://bit.ly/1GYL65j
IR Flipboard - http://bit.ly/1IkyOPA
Applied Deep Learning for Computer Vision
with Torch - http://torch.ch/docs/cvpr15.html
DIY Deep Learning: a Hands-On Tutorial with
Caffe - https://github.com/BVLC/caffe
QUESTIONS
Q & A