SLIDE 1
Learning in Intelligent Systems
October 14, 2016
Janyl Jumadinova
Overview of Learning
2/19
Learning in Intelligent Systems October 14, 2016 Janyl Jumadinova - - PowerPoint PPT Presentation
Learning in Intelligent Systems October 14, 2016 Janyl Jumadinova - - PowerPoint PPT Presentation
Learning in Intelligent Systems October 14, 2016 Janyl Jumadinova Overview of Learning 2/19 Learning in Humans The act / process of acquiring, modify or reinforcing knowledge or skills through synthesizing different types of new or existed
SLIDE 2
SLIDE 3
Learning in Humans
◮ The act / process of acquiring, modify or reinforcing knowledge
- r skills through synthesizing different types of new or existed
information.
3/19
SLIDE 4
Learning in Humans
◮ The act / process of acquiring, modify or reinforcing knowledge
- r skills through synthesizing different types of new or existed
information.
◮ Key to human survival. 3/19
SLIDE 5
Learning in Humans
◮ The act / process of acquiring, modify or reinforcing knowledge
- r skills through synthesizing different types of new or existed
information.
◮ Key to human survival. ◮ Progress over time tends to follow learning curves (relatively
permanent).
3/19
SLIDE 6
Learning in Computing Systems
◮ Computational methods using “experience” to improve
performance.
4/19
SLIDE 7
Learning in Computing Systems
◮ Computational methods using “experience” to improve
performance.
◮ Experience − data driven task 4/19
SLIDE 8
Learning in Computing Systems
◮ Computational methods using “experience” to improve
performance.
◮ Experience − data driven task ◮ Computer science – involves learning algorithms, analysis of
complexity, and theoretical guarantees.
4/19
SLIDE 9
Learning in Computing Systems
Artificial intelligence | Machine learning
5/19
SLIDE 10
Learning in Computing Systems
Artificial intelligence | Machine learning
◮ Computer program(s) with adaptive mechanisms that enable
computer / machine to learn from experience /example / analogy / rewards.
5/19
SLIDE 11
Learning in Computing Systems
Artificial intelligence | Machine learning
◮ Computer program(s) with adaptive mechanisms that enable
computer / machine to learn from experience /example / analogy / rewards.
◮ It improves the performance of an intelligent system over time
(e.g, reducing error rate, improving rewards).
5/19
SLIDE 12
Why Learning in Computing Systems?
◮ Understand and improve efficiency of human learning /
understanding.
6/19
SLIDE 13
Why Learning in Computing Systems?
◮ Understand and improve efficiency of human learning /
understanding.
◮ Discover new things or structure that is unknown to humans. 6/19
SLIDE 14
Why Learning in Computing Systems?
◮ Understand and improve efficiency of human learning /
understanding.
◮ Discover new things or structure that is unknown to humans. ◮ Fill in skeletal or incomplete knowledge / expert specifications
about a domain.
6/19
SLIDE 15
Applications of Learning
Mainly in decision making / pattern recognition / intelligent systems.
7/19
SLIDE 16
Applications of Learning
Mainly in decision making / pattern recognition / intelligent systems.
◮ Robot navigation. ◮ Automatic speech recognition (Siri in iPhone, Google
speech-to-text search).
◮ Search and recommendation (Google, Amazon, eBay). ◮ Financial prediction, fraud detection, medical diagnosis. ◮ Video games, data visualization. 7/19
SLIDE 17
Black-box Learning
8/19
SLIDE 18
Learning Architecture
9/19
SLIDE 19
Learning Paradigms
◮ Supervised learning
- input-output relationships
10/19
SLIDE 20
Learning Paradigms
◮ Supervised learning
- input-output relationships
◮ Unsupervised learning
- relationship among inputs
10/19
SLIDE 21
Learning Paradigms
◮ Supervised learning
- input-output relationships
◮ Unsupervised learning
- relationship among inputs
◮ Reinforcement learning
- input-action relates to rewards / punishment
10/19
SLIDE 22
Learning Paradigms
◮ Supervised learning
- input-output relationships
◮ Unsupervised learning
- relationship among inputs
◮ Reinforcement learning
- input-action relates to rewards / punishment
◮ Rule learning
- discovering common relationship to develop rules
10/19
SLIDE 23
Supervised Learning
Given examples of inputs and corresponding desired outputs.
11/19
SLIDE 24
Supervised Learning
Given examples of inputs and corresponding desired outputs. Tasks:
◮ Classification (categorizing output: correct class) ◮ Regression (continuous output
to predict output based for new inputs)
◮ Prediction (classify / regression on new input sequences) 11/19
SLIDE 25
Supervised Learning
12/19
SLIDE 26
Unsupervised Learning
Given only inputs and automatically discover representations, features, structure etc.
13/19
SLIDE 27
Unsupervised Learning
Given only inputs and automatically discover representations, features, structure etc. Tasks:
◮ Clustering (to group similar data into a finite number of clusters
/ groups)
◮ Vector Quantization (compress / decode dataset into a new
representation but maintaining internal information)
◮ Outlier Detection (select highly unusual cases) sequences) 13/19
SLIDE 28
Unsupervised Learning
14/19
SLIDE 29
Reinforcement Learning
◮ Learning approach of getting a computer system to act in the
world so as to maximize its rewards.
15/19
SLIDE 30
Reinforcement Learning
◮ Learning approach of getting a computer system to act in the
world so as to maximize its rewards.
◮ Consider teaching a domestic animal. We cannot tell it what to
do, but we can reward / punish if it does the right/ wrong thing.
15/19
SLIDE 31
Reinforcement Learning
◮ Learning approach of getting a computer system to act in the
world so as to maximize its rewards.
◮ Consider teaching a domestic animal. We cannot tell it what to
do, but we can reward / punish if it does the right/ wrong thing.
◮ Process to determine what it did that made it get the reward /
punishment – “credit assignment problem.”
15/19
SLIDE 32
Reinforcement Learning
16/19
SLIDE 33
Rule Learning
Given multiple measurements to discover very common settings in term of causal-effect.
17/19
SLIDE 34
Rule Learning
Given multiple measurements to discover very common settings in term of causal-effect. Tasks:
◮ Association rules (to group similar data into a finite number of
clusters / groups)
17/19
SLIDE 35
Rule Learning
Given multiple measurements to discover very common settings in term of causal-effect. Tasks:
◮ Association rules (to group similar data into a finite number of
clusters / groups)
◮ Classification rules (compress / decode dataset into a new
representation but maintaining internal information)
17/19
SLIDE 36
Rule Learning
18/19
SLIDE 37
Learning Paradigms and Some Techniques
19/19