Ottawa Machine Learning Meetup slides for event March 20, 2017, A whirlwind tour of AI and Machine Learning: When to use which technique, by Robin Grosset, Chief Technology Officer, MindBridge.Ai.
AI and Machine Learning: The many different approaches
1. We’ve built the world’s first commercial AI Auditor.
AI and Machine Learning:
The many different approaches
2. For Educational Purposes Only
Copyright 2017 MindBridge
certain content is copyright of the respective owners
3. About me
Robin Grosset
• Erstwhile Physicist from Scotland
• 20+ years in Business Analytics
• 16+ patents in the area of data processing, analytics and
security
• Founder, acquired twice: Cognos and IBM
• Appointed IBM Distinguished Engineer
• Watson Group in IBM, Chief Architect of Watson Analytics
• “InfoWorld's 2016 Technology of the Year” winner
Joined MindBridge as CTO in 2016, working on
revolutionizing financial anomaly detection with
Artificial Intelligence
8. Perceptron: 1958
1st Artificial Neural Network, Computer Vision
New York Times July 8th 1958
NEW NAVY DEVICE LEARNS BY DOING;
Psychologist Shows Embryo of Computer
Designed to Read and Grow Wiser
Frank Rosenblatt created controversy around the
abilities of perceptron: space exploration with
machines
20x20 photocell
IEEE
Award
f(x)
x1
x2
x3
Perceptron = Single Layer Neural Net
Book by Minsky and Papert
1970. “Perceptron” = no good
2nd “AI Winter”
9. Symbolic AI: 1983
• represents knowledge in human-readable form
• the dominant form of AI research until late 1980s
• Lisp Machines
• Most successful form is called an ‘Expert System’ ,
comprising inference engine and knowledge base
• Ontology Classification
• Fuzzy Logic
• Hypothetical Reasoning
“Lisp Machine”
Symbolics 3640
Workstation
c. 1984
Problems:
• Maintenance, they became brittle
• Could not learn or improve itself, needs coders
• No reusability of expertise
• Slow progress -> investment moved away
3rd “AI Winter”
16. Some of the Many ‘Tribes’ of AI
• Symbolists
• knowledge composition, rules, inverse deduction (inference), RDF, OWL.
• Connectionist
• multi-layer perceptrons with weights, backpropagation
• Analogizers
• Similarity, Kernel machines, Clustering
• Evolutionaries
• structure discovery, genetic algorithms
• Bayesians
• Uncertainty, Probabilistic inference, Monte-Carlo method (spam filter)
• Tree Huggers
• Decision Tree based models, Random Forest, + Gradient Boosting
And there are more…
17. The most common Machine Learning Algorithm
K-Means Clustering
Algorithms can learn to categorize by themselves
Algorithms tend to have an objective (cost function)
Iterates until done (cost is minimized)
K-Means Clustering
repeat {
for each point
find nearest centroid
assign to cluster
for each centroid
recalculate location
} until no point changes cluster