Professional Documents
Culture Documents
July4 SaketAnand FriendlyIntroToML
July4 SaketAnand FriendlyIntroToML
July4 SaketAnand FriendlyIntroToML
Machine Learning
with Applications to Visual Analytics
Saket Anand
Asst. Professor
Source: Rolnick et al., “Tackling Climate Change with Machine Learning”, ArXiv, 10th Jun. 2019
- 22 authors from 16 organizations
Machine Learning Overview
What should you learn?
• Modelling a learning problem
• Various algorithms (techniques) for solving ML problems
• Pitfalls while designing ML systems
• Modelling, Generalization, Regularization & Model Selection, (hyper)-Parameter
tuning, Overfitting, Underfitting
• Engineering Tricks Possibly the most important components
• Debugging ML systems
• Importance of Domain Knowledge
• Not treating ML techniques as a black box
• Simplify the learning problem by using domain knowledge
Machine Learning Paradigms
• Supervised Learning
• Labelled data – (Data, target value)
• Target value could be category/class labels, real value, real vector, etc.
• Classification, Regression
• Unsupervised Learning
• Only data, no labels
• Density Estimation, Clustering
• Semi-supervised Learning
• Some labelled data and lots of unlabelled data
• Multiple-Instance Learning
The Supervised Learning Problem
• Described through three components
• Data samples from some unknown distribution
• A supervisor (oracle) provides labels
• A learning machine capable of implementing a set of functions
• The learning problem is to choose from the given set of functions the
one which “best” approximates the supervisor’s response.
• The selection is based on training samples
• Evaluation Metrics
• Classification: Accuracy
• Regression: Mean Squared Error
• Retrieval: Precision/Recall, F-Score
• Ranking: mean Average Precision
• Clustering: Normalized Mutual Information
Performance Evaluation of Learning Tasks
• Entire population is unavailable
• obviously.ai
Python is the most popular
Support for other languages like R, Java,
C/C++, Ruby, Julia, etc.) And others…
Machine Learning Techniques
Regression: Linear Models
• Linear Regression
• Used for a continuous, normally
distributed response (Y ϵ ꓣm) variable
• Works with the following loss function
{
𝑛
1 𝑖𝑓 ∑ 𝑊 𝑖 𝑥𝑖 >0
𝑥 0=1 𝑂 ( 𝒙)= 𝑖 =0
𝑥1 𝑊1 𝑊0 −1 𝑜𝑡h𝑒𝑟𝑤𝑖𝑠𝑒
𝑥2
.
.
𝑊2
𝑊𝑛
∑ 𝑂
. Activation function: Usually a smooth
nonlinear, squishing function
𝑥𝑛
Perceptron
• A single perceptron can model a linearly separable boundary, but not
a non-linear boundary
Multi-Layered Perceptron
• Combine Perceptrons to model nonlinear decision boundaries
𝑥 𝑦
Neural Networks (NN) vs SVMs
Applications to Visual Analytics
Visual Recognition Tasks
Image Classification
Is it a natural or man
made scene
Is it a forest or a
beach?
Building:
42 m height
100 m away Car:
Police Car
Frontal View
Autonomous
and Assistive
Driving
Applications: Instance Recognition
Does this image
contains “India
Gate”?
Recognizing
landmarks in images
Recognizing products
in super market
Image Classification
• A core task in computer vision
Cat
The Problem: Semantic Gap
• Images are represented as
3D arrays of numbers, with
integers between [0, 255].
2008
WII’s 4-phase pan-India protocol
Source: WII
Population 1,411 3500 72
Camera Trap based Demography
• Extensive use of Camera Traps!
• Irrefutable evidence of tiger
presence
• Detection of poachers /
animals in Thermal Infra-
Red videos
• Very Challenging!!
77
Source: Mid Career Training Programme (MCT) Phase IV for Indian Forest Officers, 2014; Video Source: Youtube - World's Sneakiest Animals: Episode 2 Preview - BBC
Managing Macaques in Himachal
• Initiatives by Government • Existing strategy of capture is
• Translocation tedious
• commensals find nearby human settlements
• Identification of capture area, census of
• Mass awareness & legislation against macaque population, identification of
feeding of monkeys feeding sites etc.
• Too difficult to enforce
• Trap / bait/ capture team familiarization
78
Source: Mid Career Training Programme (MCT) Phase IV for Indian Forest Officers, 2014; Slide Credit: Dr. Sandeep Rattan
Our Efforts in Visual Wildlife Monitoring
• Tiger Census 2014
• ~9500 Camera Traps
• Tens of lakhs of images
• manually sorted by species
• Semi-automatic individual
tiger/leopard identification
Challenges ?
Pose Variations
● Snap & send crowdsourcing app
● ● Automatic flank matching followed
No annotation with a face
alignment-free model by user validation
Where we ●
● Learning representations robust Learning representations robust to
*Supported by
are today pose variations
to pose variations
MS AI for Earth
Acknowledgment
Prof. Y. V. Jhala Prof. Qamar Qureshi Prof. Milind Tambe Dr. Bistra Dilkina Liz Bondi
PhD Student
Ankita Shukla
PhD Student Dr. Ryan Farrell 81
Thank You!
Big Cat Mortality inLeopard
India
Tiger Mortality
Mortality
(2015-2017)
83
Slide source: Wildlife Protection Society of India (WPSI)
India – WPSI 201 20 Nov.
Tiger Data 2016 2017
4 15 2018
Documented
23 26 50 37 30
Poached Tigers
Other
Documented 58 65 82 78 68
Mortalities
Total 81 91 132 115 98
84
Current Approach: SPOT
Precision Recall
SPOT ESN ESE SPOT ESN ESE
0.4235 0.0024 0.0573 0.3697 0.0432 0.2836
86