Era of Artificial Intelligence Lecture 5 and Lecture 6 Pietro Leo
@pieroleo
The Era of Artificial Intelligence
Lecture 5 and Lecture 6
Pietro Leo
IBM Italy Executive Architect and thought leader for Artificial Intelligence
Chief Scientist for IBM Italy Research & Business
IBM Academy of Technology Leadership
Member of ISO/SC42 Artificial Intelligence Standardization Committee
www.pieroleo.com
@pieroleo
Artificial
Intelligence
Machine
Learning
Deep
Learning
Data Mining
Data Science
Big Data
Defining AI & Cognitive, Machine Learning e Deep Learning
Cognitive
IT systems that aims
to build algorithms
to predict meaning in
features of human
languages (spoken,
written, visual) and
emulates related
forms
of human reasoning
and knowledge
representation
It is a class of
techniques
that aims to generate
knowledge by
training data to
recognize the
correlation between a
set of feature
patterns and
outcomes.
Systems that
leverage a
combination of AI
reasoning and
knowledge
representation
strategies and
other analytic and
classical
computing
techniques to
solve a complex
problem
It is a rapidly
maturing Machine
Learning space,
based on neural
network techniques,
that are taught to find
their own features
Internal Use Only
@pieroleo
BUSINESS
PROBLEMS
Cloud, hybrid cloud &
on premise
Infrastructure
Data
Applications &
Processes
Traditional Algorithms
COMPUTING
DATA
BUSINESS
LOGIC
Artificial
Intelligence is
emerging as a
new enterprise
platform layer to
move and
speedup business
IT Applications
and Processes to
a new level
Artificial Intelligence /
Cognitive
Internal Use Only
AI positioning in the Enterprise IT context
@pieroleo
Cloud Public/Dedicate and Private
Infrastructure
A highly scalable, security enabled infrastructure
that run on a cloud environments
Data
Tools to acquire, manage, analyze, govern and exchange all kinds of data generated within
the Enterprise as well as gathered form outside
Artificial Intelligence / Cognitive
A set of building blocks that complement traditional computer programming models with the
goal to close applications behavior to human language (spoken, written, visual) and
reasoning
Applications & Processes
All applications and processes that run the enterprise, both horizontal, common to a number
of industries, and Vertical, specific for a given industry
On-Premise Infrastructure
The computing infrastructure that
runs inside a company data center
Traditional Algorithms
A set of languages, middleware, and products to write enterprise software programs
Transitional DataUnstructured Data IoT Data
COMPUTING
DATA
BUSINESS
LOGIC
BUSINESS
PROBLEMS
AI positioning in the Enterprise IT context
@pieroleo
11
Deep learning basic utilization models
www.pieroleo.com
Given 1:Input and 2:Weights find the 3:Output à
Prediction/Discriminative
Given 1:Input and 3:Output find the 2:Weights à
Learning
Given 2: Weights and 3:Output find the 1:Inputs à
Generative
@pieroleo
13
Source: IBM Research automatic sport highlights generation https://www.ibm.com/blogs/research/2017/06/scaling-wimbledons-video-
production-highlight-reels-ai-technology/
@pieroleo
Key topics in research Learning & Reasoning to support business problems
Making Learning More Human-
Like
People learn by trail and error without a lot of labeled
data. We learn continuously throughout their lives,
remembering what we’ve learned and leveraging it for
new tasks.
Interpretability
Explaining AI decisions is crucial for customers,
government and regulators, enterprises.
Optimization
Beyond back-propagation
Neuro AI
Novel AI approaches based on brain function including
plasticity, attention, memory, reward processing,
motivation
Deep Document Understanding
People can access the accumulated knowledge of
humanity directly, by reading, viewing and listening.
And they can apply that knowledge directly to new
tasks.
Conversational Knowledge
Acquisition
Acquiring, Applying and Accumulating knowledge
during collaboration with humans.
Multi-step Reasoning
Humans can combine inputs and knowledge from
multiple sources to solve sub-problems and larger
complex tasks
Reliable, Approximate
Reasoning
Human reasoning can be exact and it can be flexible,
AI systems need to be able to span this range
@pieroleo
1
5
Video Face Extraction
12 Jun 2016
21:40 – 22:00
Video Time Tagging
Cleveland,
OH
Video Geotagging
Face Identity Attributes
Woman,
20-30
Face Expression
Pensive
Face Extrinsics
Full hair, blond,
no glasses, no
hat
Video Object Finding
Segmented
Objects
Bicycle:{
Colour:gray,
Brand:Raleigh,
Pose: inverted}
Object Recognition
Multimedia Retrieval
To: find examples of scenes in videos with sets of objects
fitting descriptions in a list L
• Retrieve candidates videos
• For each video, and object type
• Use appropriate extractor to find
spans with that object
• Segment those objects out
• Run attribute extraction on each obj o
giving a
• Remember span and o if a satisfies
any description in L
• Remember span if it contains objects
satisfying all descriptions in L
To: Answer a query x for user u,
• Identify the language of x, l,
• Use language to logic for l on x to make an equivalent
query y,
• Reason to answer y, yielding answers z
• Use logic to language to turn each zi into language l
equivalent ai
• Assemble ai into list a
• Find a convenient display d for u
• Display x and the list a on d
English to Logic
What’s the
population of
Auckland?
(nInhabitants
Auckland ?nu)
Language ID
47
60 90
Japanese {jp}
Logic to English
(nInhabitants
Auckland 1e6)
Auckland has a
million people.
The popn of
Auckland is 1m.
Problem Solving Methods
Machine Reasoning
Multi-step reasoning for Skill Composition
@pieroleo
Thanks!
Pietro Leo
IBM Italy Executive Architect and thought leader for Artificial Intelligence
Chief Scientist for IBM Italy Research & Business
IBM Academy of Technology Leadership
Member of ISO/SC42 Artificial Intelligence Standardization Committee
www.pieroleo.com