12. The branch of computer science concerned with making computers behave like
humans.
Field of study that gives computers the ability to learn without being explicitly
programmed.
The speed with which computers can carry out data or mathematical operations.
Using case, if-then-else, or other software structures to render human knowledge
in computer code.
In the context of machine learning, algorithms define specific machine learning
approaches (e.g., logistic regression, neural networks, classification trees).
Methods for minimizing or maximizing some objective function. In machine
learning, optimization is most often used to minimize some measure of error.
Artificial Intelligence
Machine Learning
Compute Power
Explicit Programming
Algorithm
Optimization
13. • Reviewing phone calls for quality
or compliance
• Reviewing transactions for fraud
• Confirming government-issued IDs
• Answering common customer
questions
• User login data where intrusion
detection is a concern
• Server log information that could
lead to proactive production
support
• Identifying customer segments
Human-like processing, where
our scale is limited by headcount
Situations where data is large or
complicated, making it difficult
for humans to find patterns
14. Talent / Capabilities
to deliver results
Machine Learning
Leader
Key RequirementsTo Succeed in ML
Organization and
Culture
to drive overall effort
Technology
to enable ML projects to
succeed
• Vision
• Leadership buy-in
• ML dedicated teams
• Experienced ML staff
• Knowledge sharing
• Data pipeline
• Computing platforms
15.
16.
17.
18.
19. “Original” machine learning algorithm in which values are predicted or estimated
Classification algorithm that was foundational to IBS ... And still is; high
interpretability satisfies fair lending requirements
Creating new variables from raw data to improve ML performance – can be manual
or automated
Algorithm (usually classification) that uses trees in a “serial” configuration; easily
captures interactions and nonlinearities
Algorithms (usually classification) that uses independent trees in a “parallel”
configuration; powerful and easily parallelizable
Algorithms that link multiple layers of inputs and outputs together with activation
nodes; most often used in classification
Collection of tools that, together, allow for extraction of meaning, topics, or sentiment
from spoken or written language
Linear Regression
Logistic Regression
Feature Engineering
Gradient BoostedTrees
Random Forrest (Bagged
Trees)
Neural Networks /
Deep Learning
NLP
Why does it work?
Implications for coding – easy to steal? Can you replicate easily?
Good for changing input, within same pattern. Compare regex to ML based approach
When data changes, ‘easy’ to retrain with new data.
Early approaches were based on rules based or expert systems. It is was extremely hard to quantify. What was called Artificial Intelligence then, is regarded as Artificial General intelligence (AGI) now. This can be contrasted with narrow intelligence.
Machine Intelligence is different from human intelligence. Building a system that mimics human general intelligence is not seeming very practical at the moment, and possible of questionable value. Machines can be good at complementary things or subset of human attributes. We will come back to this at the end of the presentation
In my own view, Deliberately put machine learning to occupy almost the whole spectrum of machine intelligence. There maybe very few activities contributing to intelligence outside of machine learning
And I am trying to be provocative here.
These three are not the same thing. Consider whether you need an analytics
One trick here is to cast everything as a classification problem
Non linear dimensionality reduction vs linear dimensionality reduction (ICA, PCA)
Presenter notes/script:
Each one of these can consume several classes and are the subject of many research projects.
Presenter notes/script:
Presenter notes/script:
Each one of these can consume several classes and are the subject of many research projects.
If we are a tech company, we care
Unfair advantage of data