This presentation briefs about machine learning technologies, its various learning methodologies, its types. Also it briefs about the Open Computer Vision, Graphics Processing Unit and CUDA Frameworks.
2. Overview
• An Artificial Intelligence (AI) technique which
provides the system to learn by itself.
• Machine acts without being explicitly
programmed
• Utilizes the historical data to make better
business decisions
• Evolved from pattern recognition and
computation theory of Artificial Intelligence
• Basically an algorithm not a magic
3. Application Areas
• Self driving cars
• Practical speech recognition
• Efficient web search
• Human genome understanding
• Object detection
• Face detection and recognition
• Vehicle monitoring for CCTV
• Email spam detection / Cyber fraud detection
• Online recommendation
4. When to use
• Cannot code the rules
– Scenario which cannot be solved using a
deterministic rule based solution
– Rule depends on too many factors
– Many rules overlap and needs to be fine tuned
– Difficult scenario for a human to code the rules
• Cannot scale
– Effective to handle large scale problems
5. Programming Language
• MATLAB
– Excellent tool for representation and working with
matrices
• R
– Platform used to understand and explore the data
using statistical methods and graphs
• Python
– Popular scientific language and rising star of ML.
• JAVA/C
7. Different Types
• Supervised Learning
– Analyses the training data and produces output
accordingly
– Algorithm iteratively makes predictions on the training
data
– Neural Networks, Multi Layer Perception, Decision Trees
• Un supervised Learning
– Learn to inherent structure from input data
– Clustering, Distances and Normalization, Self Organizing
maps.
• Semi – Supervised
– Mixture of supervised and un-supervised techniques
9. Process Involved
• Data Collection
• Data Preparation
• Model Selection
• Training
• Evaluation
• Prediction
10. Training Process
• Input training data source
• Name of the attribute that contains target to
be predicted
• Required data transformation instructions
• Parameters to control the learning algorithm
11. Model
• Refers to the model artifact created by the
training process
• Provides the ML algorithm with training data
to learn from.
• Model Zoo
– Created with multiple datasets like COCO, Kitti and
OpenImages
12. Models
• Binary Classification Model
– Predicts a binary outcome ( one of two possible
classes )
• Multi class Classification Model
– Generates predictions for multiple classes
• Regression Model
– Predicts a numeric value
– How many units will sell tomorrow
13. Dataset
• Training set
– Set of examples used for learning with known
target
• Validation set
– Set of examples used to fine tune the classifier
and estimate the error
• Test set
– Used to access the performance of the classifier
15. ML Examples
• Object Detection & Recognition
• Multi Vehicle / Car Detection
• Vehicle Speed detection
16. OpenCV
• Open Source Computer Vision Library
• Library functions mainly aimed at real-time
computer vision, image processing and
machine learning
• Has C++, JAVA, Python library interface
• Now features GPU Acceleration for real time
operations
17. GPU
• Graphic Processing Unit
• Used to render 3D graphics comprised on
polygons
• Technologies like OpenCV, OpenCL, CUDA used to
assist the GPU in non-graphics computations
• Improves the overall performance of the
computer
• Used to accelerate the deep learning, analytics
and engineering applications.
18. CUDA
• Parallel computing platform and programming
model developed by NVIDIA
• Able to speed up the computing applications
by harnessing the power of GPUs
• GPU accelerated computing
– Sequential part of workload runs on CPU
– Intensive portion of application runs on thousands
of GPU cores in parallel
19. Tensorflow
• Open source machine learning framework for
everyone
• Numerical computation using data flow graphics
• Deploy computation on one or more CPUs or
GPUs in desktop
• Developed by Google
• Also supports hardware acceleration with
Android Neural Networks APIs.