Francis Jeanson, PhD
IAMOPEN.AI
AI ML NLP
and how to apply them
Brief AI History
Cybernetics
1930’s-1940’s
Turing Test
1950
“AI” Coined
1956
SHDRLU
1970
Expert Systems
1980’s
Rise of Symbolic AI (GOFAI)
Control Systems Theory
Rosenblatt
“Perceptron”
1957
Parallel
Distributed
Processing
1986
Minsky & Papert
1969
Limits of NNs
Artificial Neural Nets
Brief AI History
PDP
1986
Deep Blue
1997
Ensembles
Netflix Recommender
2006
MNIST 0.21
2011
Artificial Neural Networks Deep NN Era
Statistical Machine Learning
Bayesian Markov Models
Pearson Correlation Eigenvectors
Value decomposition etc…
Symbolic AI ? ? ? Cognitive Computing
What is Intelligence?
Perception
Action
Problem Solving
Language
Thinking
Reasoning
Attention
Framing
Representations
Consciousness
What is Intelligence?
Memory
Learning
Reactive
Prediction
Control
Regulation
AI Space
AI
ML
NLP
• Philosophy
• Science
• Engineering
• Engineering
applications
• Focus on models
for learning
Computational
Linguistics
Today’s AI Landscape
Traditional ML
• Information theory
• Probability theory
• Statistical origin
• Bayesian, Markovian
• Learning is regression
Deep Neural Networks
• Multi-layered NN
• Feed-forward with Back Prop
through time
• Recurrent NN
• Convolutional NN
• Meta parameter optimization
Reinforcement Learning
(Agent inspired learning)
• Temporal Difference Learning
• Actor-Critic
• Q-Learning
Complex, Adaptive, Bio Systems
• Bottom-Up
• Sufficiency over Optimality
• Evolutionary and life-time
learning
• Dynamical systems
• Network theory and Complexity
What is ML?
Input
Function
(Model)
Output
Traditionally:
Human creates Function to get Output
Input
Function
(Model)
Output
“Learning
Machine”
What is ML?
A Function
(Model) X
B Function
(Model) X
“Learning
Machine”
C Function
(Model)
Not
Recognized
A B
A C
A
X
Two Important Classes of Learning
Supervised
e.g.
– Labelled Data
– Backpropagation
– Human correction
Input
Output
Input
Output
“Learning
Machine”
Known
Function
(Model)
Known
Unknown Fitness
Features
Trials
Hypothesis
“Learning
Machine”
Function
(Model)
Unsupervised
e.g.
– Clustering
– Hebbian Learning
– Reinforcement
– Genetic Algorithms
Test
Problems & Methods
When to use ML?
• Do you have complex data with hidden relationships or
structures to model: many dimensions, multiple sources,
multi-modal, etc?
• Do you have a noisy or unstructured data and a lot of data?
• Do you need fault tolerance?
• Do you need a model that can generalize to new samples?
• Do you need a model that can create new solutions?
Important Classes of Methods
• Regression
y = ax3+bx2+cx+d
• Linear
• Logistic
• Polynomial
• Discrete choice
• etc.
• Dimensionality reduction
• Principle Component Analysis
• Partial Least Square Regression
• Independent Component Analysis
• Discriminant analysis
• etc.
Important Classes of Methods
• Bayesian Probability • Decision Trees
• Condition Decision Tree
• Random Forest
• Classification and Regression Tree
• etc.
• Naïve Bayes Classification
• Bayes Belief Networks
• Markov Models
• Conditional Random Fields
• etc.
A
B
W
X
Y
Z
P ( W | A )
P ( X | A )
P ( Y | B )
P ( Z | B )
Important Classes of Methods
• Classification
“Circle”
“Square”
• Decision trees
• Neural Networks
• Regression
• Naïve Bayes
• etc.
• Clustering
• K-Means clustering
• Expectation Maximization
• Hierarchical Clustering
• Latent Dirichlet Allocation
• etc.
What about Natural Language?
• Natural Language Processing (NLP)
– Entity extraction: Person, Place, Date, etc.
– Sentiment: Happy, Sad, Angry, Excited, etc.
– Detecting intent, topic, summary, next word, etc.
• Natural Language Understanding (NLU)
– Generalize context from content
– Map action from context
• Natural Language Generation (NLG)
– Generate an answer from a condition, conversation,
or context
NLP Approaches
• Bag of words analysis and N-grams
Model the relationship between letters or words in
a sentence or in the same text to get sentiment,
topic, summary, etc.
• Syntactic structure and parsing
Explore word dependencies or parts of speech tags
to determine intentions, context, implication, etc.
• Semantic analysis
Model the meaning of language to summarize,
translate, converse, etc.
Using Machine Learning
How to apply ML?
What is your problem?
I need a
specific output!
I’m exploring solutions,
any output will do!
Supervised Unsupervised
1. Collect training data: Input + Output upfront!
e.g. smart watch sensors classifier
2. Extract features; Reduce noise if possible; Normalize
3. Pick “the right” method
Regression, dimensionality reduction, classification, a set of
parameters,…
4. Pick the “right model”
Neural Net, SVM, HMM, POMDP, …
5. Train and Test
– Split your data into 80% for training and 20% for testing
– Test for Precision and Recall, and maybe Accuracy
1. What are your criteria for success?
2. Extract features; Reduce noise if possible; Normalize
1. Pick “the right” method
Regression, dimensionality reduction, clustering, a set of
parameters, …
2. Pick the “right model”
Clustering, Neural Net, Reinforcement Learning, Genetic
Algorithm, …
3. Test model on data and validate performance
Example: Time Series Classification
Image Processing
Activity Recognition using Cell Phone
Accelerometers, Kwapisz et al. 2010
Time series classification of Cell
Phone Accelerometer Data
Face Recognition in Real World Images
“Labeled Faces in the Wild”
Fontaine et al. 2017
Comparison with 2 and 5 training samples.
NN: Nearest Neighbor
SRC: Sparse Repres. Class.
CRC: Collab. Rep. Coding
MSPRCRC: Modified CRC
RSC: Robust Sparse Coding
Modern Types of Neural Networks
http://www.asimovinstitute.org/
Future of ML
Current Hot Topic
• Single Shot Learning
• Adversarial NN Defense
• Data and Model Privacy
• Expert Knowledge Integration
• Interpretable Inferencing
• Generative Conversational Agents
Near(ish) Future?
• Applied Spiking Neural Network
• Neuromorphic Computing
• Quantum Machine Learning
• Greater Generalization
• Fully Generative AI
Ramping up!
• Books
Cognitive Science and AI
– Mindware, by Andy Clark
Deep Learning
– http://www.deeplearningbook.org
Reinforcement Learning
– http://incompleteideas.net/sutton/book/book
draft2017june.pdf
Recommender Systems
– Recommender Systems by Agarwal
• Courses
– Coursera ML with Andrew Ng (Stanford,
Google Brain)
– Reinforcement learning with David Silver
(Google Deep Mind)
Deep Learning, ML, Reinforcement Learning
– http://videolectures.net/Top/Computer_
Science/
• Conferences
– NIPS (Neural Information Processing
Systems Conference Papers)
– ICML (International Conference on
Machine Learning)
• Programming
Frameworks (for Python)
– Deep Nets: Tensorflow, Theano, Caffe,
Keras
– Machine Learning: ScikitLearn
– NLP: NLTK, spaCy, Duckling
– Genetic Algorithms: NEAT
– Swarms: NetLogo
– Robotics: ROS
Josh Kerbel
Managing Partner
josh@extremeinnovations.com
Francis Jeanson
PhD
fjeanson@iamopen.ai

What is AI ML NLP and how to apply them

  • 1.
    Francis Jeanson, PhD IAMOPEN.AI AIML NLP and how to apply them
  • 2.
    Brief AI History Cybernetics 1930’s-1940’s TuringTest 1950 “AI” Coined 1956 SHDRLU 1970 Expert Systems 1980’s Rise of Symbolic AI (GOFAI) Control Systems Theory Rosenblatt “Perceptron” 1957 Parallel Distributed Processing 1986 Minsky & Papert 1969 Limits of NNs Artificial Neural Nets
  • 3.
    Brief AI History PDP 1986 DeepBlue 1997 Ensembles Netflix Recommender 2006 MNIST 0.21 2011 Artificial Neural Networks Deep NN Era Statistical Machine Learning Bayesian Markov Models Pearson Correlation Eigenvectors Value decomposition etc… Symbolic AI ? ? ? Cognitive Computing
  • 4.
    What is Intelligence? Perception Action ProblemSolving Language Thinking Reasoning Attention Framing Representations Consciousness
  • 5.
  • 6.
    AI Space AI ML NLP • Philosophy •Science • Engineering • Engineering applications • Focus on models for learning Computational Linguistics
  • 7.
    Today’s AI Landscape TraditionalML • Information theory • Probability theory • Statistical origin • Bayesian, Markovian • Learning is regression Deep Neural Networks • Multi-layered NN • Feed-forward with Back Prop through time • Recurrent NN • Convolutional NN • Meta parameter optimization Reinforcement Learning (Agent inspired learning) • Temporal Difference Learning • Actor-Critic • Q-Learning Complex, Adaptive, Bio Systems • Bottom-Up • Sufficiency over Optimality • Evolutionary and life-time learning • Dynamical systems • Network theory and Complexity
  • 8.
    What is ML? Input Function (Model) Output Traditionally: Humancreates Function to get Output Input Function (Model) Output “Learning Machine”
  • 9.
    What is ML? AFunction (Model) X B Function (Model) X “Learning Machine” C Function (Model) Not Recognized A B A C A X
  • 10.
    Two Important Classesof Learning Supervised e.g. – Labelled Data – Backpropagation – Human correction Input Output Input Output “Learning Machine” Known Function (Model) Known Unknown Fitness Features Trials Hypothesis “Learning Machine” Function (Model) Unsupervised e.g. – Clustering – Hebbian Learning – Reinforcement – Genetic Algorithms Test
  • 11.
  • 12.
    When to useML? • Do you have complex data with hidden relationships or structures to model: many dimensions, multiple sources, multi-modal, etc? • Do you have a noisy or unstructured data and a lot of data? • Do you need fault tolerance? • Do you need a model that can generalize to new samples? • Do you need a model that can create new solutions?
  • 13.
    Important Classes ofMethods • Regression y = ax3+bx2+cx+d • Linear • Logistic • Polynomial • Discrete choice • etc. • Dimensionality reduction • Principle Component Analysis • Partial Least Square Regression • Independent Component Analysis • Discriminant analysis • etc.
  • 14.
    Important Classes ofMethods • Bayesian Probability • Decision Trees • Condition Decision Tree • Random Forest • Classification and Regression Tree • etc. • Naïve Bayes Classification • Bayes Belief Networks • Markov Models • Conditional Random Fields • etc. A B W X Y Z P ( W | A ) P ( X | A ) P ( Y | B ) P ( Z | B )
  • 15.
    Important Classes ofMethods • Classification “Circle” “Square” • Decision trees • Neural Networks • Regression • Naïve Bayes • etc. • Clustering • K-Means clustering • Expectation Maximization • Hierarchical Clustering • Latent Dirichlet Allocation • etc.
  • 16.
    What about NaturalLanguage? • Natural Language Processing (NLP) – Entity extraction: Person, Place, Date, etc. – Sentiment: Happy, Sad, Angry, Excited, etc. – Detecting intent, topic, summary, next word, etc. • Natural Language Understanding (NLU) – Generalize context from content – Map action from context • Natural Language Generation (NLG) – Generate an answer from a condition, conversation, or context
  • 17.
    NLP Approaches • Bagof words analysis and N-grams Model the relationship between letters or words in a sentence or in the same text to get sentiment, topic, summary, etc. • Syntactic structure and parsing Explore word dependencies or parts of speech tags to determine intentions, context, implication, etc. • Semantic analysis Model the meaning of language to summarize, translate, converse, etc.
  • 18.
  • 19.
    How to applyML? What is your problem? I need a specific output! I’m exploring solutions, any output will do! Supervised Unsupervised 1. Collect training data: Input + Output upfront! e.g. smart watch sensors classifier 2. Extract features; Reduce noise if possible; Normalize 3. Pick “the right” method Regression, dimensionality reduction, classification, a set of parameters,… 4. Pick the “right model” Neural Net, SVM, HMM, POMDP, … 5. Train and Test – Split your data into 80% for training and 20% for testing – Test for Precision and Recall, and maybe Accuracy 1. What are your criteria for success? 2. Extract features; Reduce noise if possible; Normalize 1. Pick “the right” method Regression, dimensionality reduction, clustering, a set of parameters, … 2. Pick the “right model” Clustering, Neural Net, Reinforcement Learning, Genetic Algorithm, … 3. Test model on data and validate performance
  • 20.
    Example: Time SeriesClassification Image Processing Activity Recognition using Cell Phone Accelerometers, Kwapisz et al. 2010 Time series classification of Cell Phone Accelerometer Data
  • 21.
    Face Recognition inReal World Images “Labeled Faces in the Wild” Fontaine et al. 2017 Comparison with 2 and 5 training samples. NN: Nearest Neighbor SRC: Sparse Repres. Class. CRC: Collab. Rep. Coding MSPRCRC: Modified CRC RSC: Robust Sparse Coding
  • 22.
    Modern Types ofNeural Networks http://www.asimovinstitute.org/
  • 23.
    Future of ML CurrentHot Topic • Single Shot Learning • Adversarial NN Defense • Data and Model Privacy • Expert Knowledge Integration • Interpretable Inferencing • Generative Conversational Agents Near(ish) Future? • Applied Spiking Neural Network • Neuromorphic Computing • Quantum Machine Learning • Greater Generalization • Fully Generative AI
  • 24.
    Ramping up! • Books CognitiveScience and AI – Mindware, by Andy Clark Deep Learning – http://www.deeplearningbook.org Reinforcement Learning – http://incompleteideas.net/sutton/book/book draft2017june.pdf Recommender Systems – Recommender Systems by Agarwal • Courses – Coursera ML with Andrew Ng (Stanford, Google Brain) – Reinforcement learning with David Silver (Google Deep Mind) Deep Learning, ML, Reinforcement Learning – http://videolectures.net/Top/Computer_ Science/ • Conferences – NIPS (Neural Information Processing Systems Conference Papers) – ICML (International Conference on Machine Learning) • Programming Frameworks (for Python) – Deep Nets: Tensorflow, Theano, Caffe, Keras – Machine Learning: ScikitLearn – NLP: NLTK, spaCy, Duckling – Genetic Algorithms: NEAT – Swarms: NetLogo – Robotics: ROS
  • 25.