Topic 1. What is a Learning System?

Reference:
Tom Mitchell [1]: [1] Tom Mitchell, Machine Learning, McGraw Hill, 1997

A Computer learns a Task T

from experience E,

if its Performance P improves with E.

For Computers, Experience is a repository of historical data

Some observations on Mitchell’s
definition
• Focus on Data (Experience) for carrying out Tasks!

• Any learning approach must perform the Task better,

when given more Data

• Continuous assessment of Performance

 Applying Mitchell’s definition to Recruitment
task
T: Predict +ve (Hired) or –ve (Rejected) for candidates
E: Experience: data repository of past hirings.
P: Performance metrics 🡪 Tallies
True +ve (TP): Number predicted +ve and actually Hired
True –ve (TN): Number predicted –ve and actually Rejected
False +ve (FP): Number predicted +ve, but actually Rejected
False –ve (FN): Number predicted –ve, But actually Hired
 ML
APPROACH
1
approach1

ML
ML ML
APPROACH
TASKS
APPROACH
2 3
approach2 approach3

Adequate Data

Performance
Intermittent

Assessment
Algorithm
Learning
Topic 2. Approaches for Machine Learning
 Approach I. Supervised Learning – Discreet Output
Task: Predict results of Recruitment
Attributes
Highest Years of Coding Referenc Interview RESPONSE:
Qualifications Experience skills e perf. Accept/Reject

PG 1 Good Excellent Good */Accept

TRAINING UG 2.8 Med Med Bad */Reject
DATA
UG 5.5 Good Good Excellent */Accept TRAINING ERROR

………..
NEW DATA PG 6.3 Med Good Med ???/Reject True Error
TP TN
FP FN
 Approach I. Supervised Learning – Continuous output
Task: Recruitment – predict scores

Highest Years of Coding Reference Interview Score

Qualifications Experience skills marks

PG 1 Good Excellent 79 */80

TRAINING UG 2.8 Med Med 55 */50
TRAINING ERRORS
DATA
UG 5.5 Good Good 87 */75
Errors
……….. (Predicted score - Actual score)2
NEW DATA PG 6.3 Med Good 67 ???/70
 Types of supervised tasks
Supervised
Learning

Score Accept/Reject
Regression Classification
(Continuous (Discreet
Response) Response)
 The Process of Supervised ML
TASK: Predict Recruitment Results
VALIDATION/TEST
Learning Highes Years DATA
Level of Coding Referenc Interview Result
t of Skill e Perf
TRAINING Qualifi Exper
DATA
Algorithm cation ience
Highes Years Level of Reference Interview Result PG 5 Good Good Good ???/Accept
t of Coding Performanc
Qualifi Exper Skill e UG 3 Med Good Med ???/Reject
cation ience UG 6 Med Medium Good ???/Accept
UG 5 Good Excellent Good */Accept 1. Develop a UG 2 Good Excellent Bad ???Accept
PG 3 Med Good Bad */Reject
Model … … … …

TRAINING CYCLES:
VALID./TEST
…

UG 6 Poor Medium Excellent */Accept

UG 2 Good Excellent Good */Accept CYCLES:

… … … … …

Model
2. Optimize the
FIELD
High Year Level DATA Interview
Reference Result Model
est s of of Perf
Quali Exp Coding
ficati erie Skill 3. Apply Error
ALTERNATING TRAIN CYCLE
ASSESS / VALIDATE ON
PERFORMANCE
on nce
PG 5 Good Good Excellent ??? the Model CYCLE VALID./TEST DATA
Feedback-LEARN
UG 3 Med Good Good ???
UG 6 Med Medium Med ???
Algorithms for Supervised Learning
• Linear Regression
• Logistic Regression
• Bayesian Classification
• Decision Trees
• Support Vector Machines
• K-Nearest Neighbours
• Artificial Neural Networks
Column index Years of Level of Coding Skill Reference Interview Performance Result
Experience

Row Index

anil 5 Good Excellent Good */Accept

Sita 3 Med Good Bad */Reject

Gita 6 Poor Medium Excellent */Accept

Riva 2 Good Excellent Good */Accept

… … … … …
 Summary

• Supervised learning is Error reduction based learning.

• Goal of ML agent: Learn a mapping function from input

(feature values) to output (response) 🡪 Target
Hypothesis.

• Learning gained from examples is induced to new cases

 Approach II. Unsupervised –
Learning without a teacher
Extract Concepts in biodata of candidates

CONCEPT 1: Candidates who: CONCEPT 2:

• UG • UG/PG,
• Experience 1 to 2 years, • Experience any,
• Good/Excellent reference, • Bad reference,
• Good/Med coding skills, • Med/Low coding skills,
• Excellent or Good performance in • Med/Bad perf in interview
interview
Types of unsupervised tasks
Unsupervised
Learning

Principal Component
Analysis Association Rules
Clustering
( Extracts informative (Find associations
(Clubs similar data)
features and their between data items)
correlations)
 Algorithms for Unsupervised Learning
• Clustering:
• K-means Clustering
• Hierarchical Clustering
• Principal Component Analysis
• Association rules
• Self Organizing Maps (SOM)
• Singular Value Decomposition SVD
• Anomaly Detection
 Approach III. Semi-Supervised Learning
• Labelled pool D1 & Unlabelled pool D2
• Repeat till acceptable performance:
• Learn with labelled examples D1
• Classify D2
• Label confidently classified instances in D2
• Add them to D1
• Can surpass performance with either D1 or D2
• No burden of collecting lots of labelled data
 Approach IV. Reinforcement Learning

• No examples to begin with!

• Some inkling of good or bad steps
• Gradually learns from rewards (or penalties) from a series of
interactions with the environment
Environment Actions🡪 Change
Changed State (Roads &Traffic) Environment State
& Rewards

A
S C
E T
N U
S
ML+AI Agents A
O Navigator in Car T
R O
S R
S
Topic 3: Convergence of Technologies
Relationship between ML & other domains of
AI
AI

ML
Learn
Maths, Statistics,
Hear, Probability,
Plan & NLP Visualization
Talk Data Science
Act Knowledge
Robotics DL &
Interact

Vision

See n interpret
Great Applications!
• Robot-based manufacturing
• Human assistance systems
• Autonomous driving
• Understanding & speaking human language
• Translating one language to another
• Deep Sea/earth Space exploration
• ……..
 Learning Points
⮚Machine Learning has these components –
⮚ A Task to be learnt

⮚ Data on past experience

⮚ Performance to be assessed intermittently

⮚A learning approach to find mappings (hypothesis)

 Learning Points
⮚ The main approaches for ML are:
⮚ Supervised learning – Error reduction based
hypothesis learning
⮚ Unsupervised – Finds latent similarities,
correlations and associations in data
⮚ Semi-supervised learning – Boosts unlabelled
data starting with few labelled examples
⮚ Reinforcement Learning – interactive & reward
based learning
 Learning Points
⮚ The supervised ML Process includes:
⮚ Training with labelled data 🡪 Model

⮚ Validating and testing performance of model

⮚ Applying field data on trained model

 Learning Points
⮚Machine Learning is:
⮚ A sub-branch of AI
⮚ Includes Deep Learning
⮚ Interacts with interdisciplinary fields like Robotics,
NLP and Vision
⮚ Several sophisticated applications use convergence of
technologies
► Link for Self Assessment:

[Link]

The Secret to going ahead is getting started – Mark Twain